Applied machine learning with a solid foundation in theory. Revised and expanded for TensorFlow 2, GANs, and reinforceme

*2,636*
*350*
*26MB*

*English*
*Pages 770
[771]*
*Year 2019*

- Author / Uploaded
- Sebastian Raschka
- Vahid Mirjalili

- Categories
- Computers
- Cybernetics: Artificial Intelligence

**Commentary**- True PDF

*Table of contents : 1. Giving Computers the Ability to Learn from Data2. Training Simple ML Algorithms for Classification3. ML Classifiers Using scikit-learn4. Building Good Training Datasets - Data Preprocessing5. Compressing Data via Dimensionality Reduction6. Best Practices for Model Evaluation and Hyperparameter Tuning7. Combining Different Models for Ensemble Learning8. Applying ML to Sentiment Analysis9. Embedding a ML Model into a Web Application10. Predicting Continuous Target Variables with Regression Analysis11. Working with Unlabeled Data - Clustering Analysis12. Implementing Multilayer Artificial Neural Networks13. Parallelizing Neural Network Training with TensorFlow14. TensorFlow Mechanics15. Classifying Images with Deep Convolutional Neural Networks16. Modeling Sequential Data Using Recurrent Neural Networks17. GANs for Synthesizing New Data18. RL for Decision Making in Complex Environments*

Python Machine Learning Third Edition

Machine Learning and Deep Learning with Python, scikit-learn, and TensorFlow 2

Sebastian Raschka Vahid Mirjalili

BIRMINGHAM - MUMBAI

Python Machine Learning Third Edition Copyright © 2019 Packt Publishing All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, without the prior written permission of the publisher, except in the case of brief quotations embedded in critical articles or reviews. Every effort has been made in the preparation of this book to ensure the accuracy of the information presented. However, the information contained in this book is sold without warranty, either express or implied. Neither the authors, nor Packt Publishing or its dealers and distributors, will be held liable for any damages caused or alleged to have been caused directly or indirectly by this book. Packt Publishing has endeavored to provide trademark information about all of the companies and products mentioned in this book by the appropriate use of capitals. However, Packt Publishing cannot guarantee the accuracy of this information. Acquisition Editor: Jonathan Malysiak Acquisition Editor – Peer Reviews: Suresh Jain Content Development Editors: Joanne Lovell, Chris Nelson Technical Editor: Saby D'silva Project Editor: Radhika Atitkar Proofreader: Safis Editing Indexer: Tejal Daruwale Soni Presentation Designer: Sandip Tadge First published: September 2015 Second edition: September 2017 Third edition: December 2019 Production reference: 1091219 Published by Packt Publishing Ltd. Livery Place 35 Livery Street Birmingham B3 2PB, UK. ISBN 978-1-78995-575-0 www.packt.com

packt.com

Subscribe to our online digital library for full access to over 7,000 books and videos, as well as industry leading tools to help you plan your personal development and advance your career. For more information, please visit our website.

Why subscribe?

• Spend less time learning and more time coding with practical eBooks and Videos from over 4,000 industry professionals • Learn better with Skill Plans built especially for you • Get a free eBook or video every month • Fully searchable for easy access to vital information • Copy and paste, print, and bookmark content

Did you know that Packt offers eBook versions of every book published, with PDF and ePub files available? You can upgrade to the eBook version at www.Packt.com and as a print book customer, you are entitled to a discount on the eBook copy. Get in touch with us at [email protected] for more details. At www.Packt.com, you can also read a collection of free technical articles, sign up for a range of free newsletters, and receive exclusive discounts and offers on Packt books and eBooks.

Contributors About the authors

Sebastian Raschka received his doctorate from Michigan State University, where he focused on developing methods at the intersection of computational biology and machine learning. In the summer of 2018, he joined the University of WisconsinMadison as Assistant Professor of Statistics. His research activities include the development of new deep learning architectures to solve problems in the field of biometrics. Sebastian has many years of experience with coding in Python and has given several seminars on the practical applications of data science, machine learning, and deep learning over the years, including a machine learning tutorial at SciPy, the leading conference for scientific computing in Python. Among Sebastian's achievements is his book Python Machine Learning, which is a bestselling title at Packt and on Amazon.com. The book received the ACM Best of Computing award in 2016 and was translated into many different languages, including German, Korean, Chinese, Japanese, Russian, Polish, and Italian. In his free time, Sebastian loves to contribute to open source projects, and methods that he implemented are now successfully used in machine learning competitions such as Kaggle. I would like to take this opportunity to thank the great Python community and the developers of open source packages who helped me create the perfect environment for scientific research and data science. Also, I want to thank my parents, who always encouraged and supported me in pursuing the path and career that I was so passionate about. Special thanks to the core developers of scikit-learn and TensorFlow. As a contributor and user, I had the pleasure of working with great people who are not only very knowledgeable when it comes to machine learning and deep learning but are also excellent programmers.

Vahid Mirjalili obtained his PhD in mechanical engineering working on novel methods for large-scale, computational simulations of molecular structures at Michigan State University. Being passionate about the field of machine learning, he joined the iPRoBe lab at Michigan State University, where he worked on applying machine learning in the computer vision and biometrics domains. After several productive years at the iPRoBe lab and many years in academia, Vahid recently joined 3M Company as a research scientist, where he can use his expertise and apply state-of-the-art machine learning and deep learning techniques to solve real-world problems in various applications to make life better.

I would like to thank my wife, Taban Eslami, who has been very supportive and encouraged me on my career path. Also, special thanks to my advisors, Nikolai Priezjev, Michael Feig, and Arun Ross, for supporting me during my PhD studies, as well as my professors, Vishnu Boddeti, Leslie Kuhn, and Xiaoming Liu, who have taught me so much and encouraged me to pursue my passion.

About the reviewers

Raghav Bali is a senior data scientist at one of the world's largest healthcare

organizations. His work involves the research and development of enterprise-level solutions based on machine learning, deep learning, and natural language processing for healthcare- and insurance-related use cases. In his previous role at Intel, he was involved in enabling proactive data-driven IT initiatives using natural language processing, deep learning, and traditional statistical methods. He has also worked in the finance domain with American Express, solving digital engagement and customer retention use cases. Raghav has also authored multiple books with leading publishers, the most recent being on the latest advancements in transfer learning research. Raghav has a master's degree (gold medalist) in information technology from the International Institute of Information Technology, Bangalore. Raghav loves reading and is a shutterbug, capturing moments when he isn't busy solving problems.

Motaz Saad holds a PhD in computer science from the University of Lorraine.

He loves data and he likes to play with it. He has over 10 years of professional experience in natural language processing, computational linguistics, data science, and machine learning. He currently works as an assistant professor at the faculty of Information Technology, IUG.

Table of Contents Prefacexiii Chapter 1: Giving Computers the Ability to Learn from Data 1 Building intelligent machines to transform data into knowledge The three different types of machine learning Making predictions about the future with supervised learning Classification for predicting class labels Regression for predicting continuous outcomes

Solving interactive problems with reinforcement learning Discovering hidden structures with unsupervised learning Finding subgroups with clustering Dimensionality reduction for data compression

1 2 3

3 4

6 7

7 8

Introduction to the basic terminology and notations 8 Notation and conventions used in this book 9 Machine learning terminology 11 A roadmap for building machine learning systems 11 Preprocessing – getting data into shape 12 Training and selecting a predictive model 13 Evaluating models and predicting unseen data instances 14 Using Python for machine learning 14 Installing Python and packages from the Python Package Index 14 Using the Anaconda Python distribution and package manager 15 Packages for scientific computing, data science, and machine learning 16 Summary16

[i]

Table of Contents

Chapter 2: Training Simple Machine Learning Algorithms for Classification

19

Chapter 3: A Tour of Machine Learning Classifiers Using scikit-learn

53

Artificial neurons – a brief glimpse into the early history of machine learning 20 The formal definition of an artificial neuron 21 The perceptron learning rule 23 Implementing a perceptron learning algorithm in Python 26 An object-oriented perceptron API 26 Training a perceptron model on the Iris dataset 30 Adaptive linear neurons and the convergence of learning 36 Minimizing cost functions with gradient descent 37 Implementing Adaline in Python 40 Improving gradient descent through feature scaling 44 Large-scale machine learning and stochastic gradient descent 46 Summary51

Choosing a classification algorithm First steps with scikit-learn – training a perceptron Modeling class probabilities via logistic regression Logistic regression and conditional probabilities Learning the weights of the logistic cost function Converting an Adaline implementation into an algorithm for logistic regression Training a logistic regression model with scikit-learn Tackling overfitting via regularization Maximum margin classification with support vector machines Maximum margin intuition Dealing with a nonlinearly separable case using slack variables Alternative implementations in scikit-learn Solving nonlinear problems using a kernel SVM Kernel methods for linearly inseparable data Using the kernel trick to find separating hyperplanes in a high-dimensional space Decision tree learning Maximizing IG – getting the most bang for your buck Building a decision tree Combining multiple decision trees via random forests K-nearest neighbors – a lazy learning algorithm Summary [ ii ]

53 54 60 60 65 67 72 75 79 79 81 83 84 84

86 90 91 96 100 103 108

Table of Contents

Chapter 4: Building Good Training Datasets – Data Preprocessing

109

Chapter 5: Compressing Data via Dimensionality Reduction

145

Dealing with missing data 109 Identifying missing values in tabular data 110 Eliminating training examples or features with missing values 111 Imputing missing values 112 Understanding the scikit-learn estimator API 113 Handling categorical data 115 Categorical data encoding with pandas 116 Mapping ordinal features 116 Encoding class labels 117 Performing one-hot encoding on nominal features 118 Partitioning a dataset into separate training and test datasets 121 Bringing features onto the same scale 124 Selecting meaningful features 127 L1 and L2 regularization as penalties against model complexity 128 A geometric interpretation of L2 regularization 128 Sparse solutions with L1 regularization 131 Sequential feature selection algorithms 135 Assessing feature importance with random forests 141 Summary143 Unsupervised dimensionality reduction via principal component analysis The main steps behind principal component analysis Extracting the principal components step by step Total and explained variance Feature transformation Principal component analysis in scikit-learn Supervised data compression via linear discriminant analysis Principal component analysis versus linear discriminant analysis The inner workings of linear discriminant analysis Computing the scatter matrices Selecting linear discriminants for the new feature subspace Projecting examples onto the new feature space LDA via scikit-learn Using kernel principal component analysis for nonlinear mappings Kernel functions and the kernel trick Implementing a kernel principal component analysis in Python Example 1 – separating half-moon shapes Example 2 – separating concentric circles

[ iii ]

145 146 148 151 152 155 159 159 160 161 164 167 168 169 170 175

177 180

Table of Contents

Projecting new data points 183 Kernel principal component analysis in scikit-learn 187 Summary188

Chapter 6: Learning Best Practices for Model Evaluation and Hyperparameter Tuning

191

Chapter 7: Combining Different Models for Ensemble Learning

223

Streamlining workflows with pipelines Loading the Breast Cancer Wisconsin dataset Combining transformers and estimators in a pipeline Using k-fold cross-validation to assess model performance The holdout method K-fold cross-validation Debugging algorithms with learning and validation curves Diagnosing bias and variance problems with learning curves Addressing over- and underfitting with validation curves Fine-tuning machine learning models via grid search Tuning hyperparameters via grid search Algorithm selection with nested cross-validation Looking at different performance evaluation metrics Reading a confusion matrix Optimizing the precision and recall of a classification model Plotting a receiver operating characteristic Scoring metrics for multiclass classification Dealing with class imbalance Summary Learning with ensembles Combining classifiers via majority vote Implementing a simple majority vote classifier Using the majority voting principle to make predictions Evaluating and tuning the ensemble classifier Bagging – building an ensemble of classifiers from bootstrap samples Bagging in a nutshell Applying bagging to classify examples in the Wine dataset Leveraging weak learners via adaptive boosting How boosting works Applying AdaBoost using scikit-learn Summary

Chapter 8: Applying Machine Learning to Sentiment Analysis Preparing the IMDb movie review data for text processing [ iv ]

191 192 193 195 196 197 201 201 205 207 207 209 211 211 213 216 219 220 222 223 227 228 234 237 243 244 245 249 250 254 257

259 259

Table of Contents

Obtaining the movie review dataset Preprocessing the movie dataset into a more convenient format Introducing the bag-of-words model Transforming words into feature vectors Assessing word relevancy via term frequency-inverse document frequency Cleaning text data Processing documents into tokens Training a logistic regression model for document classification Working with bigger data – online algorithms and out-of-core learning Topic modeling with Latent Dirichlet Allocation Decomposing text documents with LDA LDA with scikit-learn Summary

Chapter 9: Embedding a Machine Learning Model into a Web Application Serializing fitted scikit-learn estimators Setting up an SQLite database for data storage Developing a web application with Flask Our first Flask web application Form validation and rendering Setting up the directory structure Implementing a macro using the Jinja2 templating engine Adding style via CSS Creating the result page

260 260 262 263 265 267 269 272 274 278 279 279 283

285 285 289 291 292 294

295 296 296 298

Turning the movie review classifier into a web application 300 Files and folders – looking at the directory tree 301 Implementing the main application as app.py 302 Setting up the review form 305 Creating a results page template 306 Deploying the web application to a public server 309 Creating a PythonAnywhere account 309 Uploading the movie classifier application 310 Updating the movie classifier 311 Summary314

Chapter 10: Predicting Continuous Target Variables with Regression Analysis Introducing linear regression Simple linear regression Multiple linear regression

[v]

315 315 316 317

Table of Contents

Exploring the Housing dataset Loading the Housing dataset into a data frame Visualizing the important characteristics of a dataset Looking at relationships using a correlation matrix Implementing an ordinary least squares linear regression model Solving regression for regression parameters with gradient descent Estimating the coefficient of a regression model via scikit-learn Fitting a robust regression model using RANSAC Evaluating the performance of linear regression models Using regularized methods for regression Turning a linear regression model into a curve – polynomial regression Adding polynomial terms using scikit-learn Modeling nonlinear relationships in the Housing dataset Dealing with nonlinear relationships using random forests Decision tree regression Random forest regression Summary

318 318 320 322 325 325 330 332 334 337 339 340 342 345 346 348 350

Chapter 11: Working with Unlabeled Data – Clustering Analysis

353

Chapter 12: Implementing a Multilayer Artificial Neural Network from Scratch

383

Grouping objects by similarity using k-means K-means clustering using scikit-learn A smarter way of placing the initial cluster centroids using k-means++ Hard versus soft clustering Using the elbow method to find the optimal number of clusters Quantifying the quality of clustering via silhouette plots Organizing clusters as a hierarchical tree Grouping clusters in bottom-up fashion Performing hierarchical clustering on a distance matrix Attaching dendrograms to a heat map Applying agglomerative clustering via scikit-learn Locating regions of high density via DBSCAN Summary

Modeling complex functions with artificial neural networks Single-layer neural network recap Introducing the multilayer neural network architecture Activating a neural network via forward propagation Classifying handwritten digits Obtaining and preparing the MNIST dataset [ vi ]

353 354 358 359 361 363 367 368 369 373 375 376 382

383 385 387 391 393 394

Table of Contents

Implementing a multilayer perceptron Training an artificial neural network Computing the logistic cost function Developing your understanding of backpropagation Training neural networks via backpropagation About the convergence in neural networks A few last words about the neural network implementation Summary

Chapter 13: Parallelizing Neural Network Training with TensorFlow

400 412 412 415 417 421 422 423

425

TensorFlow and training performance 426 Performance challenges 426 What is TensorFlow? 427 How we will learn TensorFlow 429 First steps with TensorFlow 429 Installing TensorFlow 429 Creating tensors in TensorFlow 430 Manipulating the data type and shape of a tensor 431 Applying mathematical operations to tensors 432 Split, stack, and concatenate tensors 434 Building input pipelines using tf.data – the TensorFlow Dataset API 435 Creating a TensorFlow Dataset from existing tensors 436 Combining two tensors into a joint dataset 437 Shuffle, batch, and repeat 439 Creating a dataset from files on your local storage disk 441 Fetching available datasets from the tensorflow_datasets library 445 Building an NN model in TensorFlow 450 The TensorFlow Keras API (tf.keras) 451 Building a linear regression model 451 Model training via the .compile() and .fit() methods 456 Building a multilayer perceptron for classifying flowers in the Iris dataset 457 Evaluating the trained model on the test dataset 461 Saving and reloading the trained model 461 Choosing activation functions for multilayer neural networks 462 Logistic function recap 463 Estimating class probabilities in multiclass classification via the softmax function465 Broadening the output spectrum using a hyperbolic tangent 466 Rectified linear unit activation 468 Summary 470 [ vii ]

Table of Contents

Chapter 14: Going Deeper – The Mechanics of TensorFlow

471

Chapter 15: Classifying Images with Deep Convolutional Neural Networks

517

The key features of TensorFlow 472 TensorFlow's computation graphs: migrating to TensorFlow v2 473 Understanding computation graphs 473 Creating a graph in TensorFlow v1.x 474 Migrating a graph to TensorFlow v2 475 Loading input data into a model: TensorFlow v1.x style 476 Loading input data into a model: TensorFlow v2 style 476 Improving computational performance with function decorators 477 TensorFlow Variable objects for storing and updating model parameters 479 Computing gradients via automatic differentiation and GradientTape 483 Computing the gradients of the loss with respect to trainable variables 483 Computing gradients with respect to non-trainable tensors 485 Keeping resources for multiple gradient computations 485 Simplifying implementations of common architectures via the Keras API 486 Solving an XOR classification problem 489 Making model building more flexible with Keras' functional API 494 Implementing models based on Keras' Model class 496 Writing custom Keras layers 497 TensorFlow Estimators 501 Working with feature columns 501 Machine learning with pre-made Estimators 506 Using Estimators for MNIST handwritten digit classification 510 Creating a custom Estimator from an existing Keras model 512 Summary515

The building blocks of CNNs Understanding CNNs and feature hierarchies Performing discrete convolutions

Discrete convolutions in one dimension Padding inputs to control the size of the output feature maps Determining the size of the convolution output Performing a discrete convolution in 2D

Subsampling layers Putting everything together – implementing a CNN Working with multiple input or color channels Regularizing an NN with dropout Loss functions for classification [ viii ]

518 518 520

521 523 525 526

530 532 532 536 539

Table of Contents

Implementing a deep CNN using TensorFlow The multilayer CNN architecture Loading and preprocessing the data Implementing a CNN using the TensorFlow Keras API Configuring CNN layers in Keras Constructing a CNN in Keras

542 542 543 544

544 545

Gender classification from face images using a CNN 550 Loading the CelebA dataset 551 Image transformation and data augmentation 552 Training a CNN gender classifier 558 Summary564

Chapter 16: Modeling Sequential Data Using Recurrent Neural Networks

Introducing sequential data Modeling sequential data – order matters Representing sequences The different categories of sequence modeling RNNs for modeling sequences Understanding the RNN looping mechanism Computing activations in an RNN Hidden-recurrence versus output-recurrence The challenges of learning long-range interactions Long short-term memory cells Implementing RNNs for sequence modeling in TensorFlow Project one – predicting the sentiment of IMDb movie reviews Preparing the movie review data Embedding layers for sentence encoding Building an RNN model Building an RNN model for the sentiment analysis task

Project two – character-level language modeling in TensorFlow Preprocessing the dataset Building a character-level RNN model Evaluation phase – generating new text passages

Understanding language with the Transformer model Understanding the self-attention mechanism

A basic version of self-attention Parameterizing the self-attention mechanism with query, key, and value weights

567 568 568 569 570 571 571 574 577 580 582 584 585

585 590 592 594

600

601 607 609

613 614

614 616

Multi-head attention and the Transformer block 617 Summary618

Chapter 17: Generative Adversarial Networks for Synthesizing New Data Introducing generative adversarial networks [ ix ]

619 620

Table of Contents

Starting with autoencoders Generative models for synthesizing new data Generating new samples with GANs Understanding the loss functions of the generator and discriminator networks in a GAN model Implementing a GAN from scratch Training GAN models on Google Colab Implementing the generator and the discriminator networks Defining the training dataset Training the GAN model Improving the quality of synthesized images using a convolutional and Wasserstein GAN Transposed convolution Batch normalization Implementing the generator and discriminator Dissimilarity measures between two distributions Using EM distance in practice for GANs Gradient penalty Implementing WGAN-GP to train the DCGAN model Mode collapse Other GAN applications Summary

Chapter 18: Reinforcement Learning for Decision Making in Complex Environments

Introduction – learning from experience Understanding reinforcement learning Defining the agent-environment interface of a reinforcement learning system The theoretical foundations of RL Markov decision processes The mathematical formulation of Markov decision processes Visualization of a Markov process Episodic versus continuing tasks

620 623 624 626 628 628 631 636 638 646 647 648 651 657 661 662 663 667 669 670

671 672 672 674 676 676 677

679 680

RL terminology: return, policy, and value function

680

Dynamic programming using the Bellman equation Reinforcement learning algorithms Dynamic programming

685 686 686

The return 680 Policy682 Value function 683

Policy evaluation – predicting the value function with dynamic programming Improving the policy using the estimated value function [x]

687 688

Table of Contents Policy iteration Value iteration

688 689

Reinforcement learning with Monte Carlo

689

Temporal difference learning

691

State-value function estimation using MC Action-value function estimation using MC Finding an optimal policy using MC control Policy improvement – computing the greedy policy from the action-value function TD prediction On-policy TD control (SARSA) Off-policy TD control (Q-learning)

Implementing our first RL algorithm Introducing the OpenAI Gym toolkit

Working with the existing environments in OpenAI Gym A grid world example Implementing the grid world environment in OpenAI Gym

690 690 691 691 692 693 694

694 695

695 697 698

Solving the grid world problem with Q-learning

705

A glance at deep Q-learning

709

Chapter and book summary

717

Implementing the Q-learning algorithm

Training a DQN model according to the Q-learning algorithm Implementing a deep Q-learning algorithm

Other Books You May Enjoy Index

705 710 712

721 725

[ xi ]

Preface Through exposure to the news and social media, you are probably very familiar with the fact that machine learning has become one of the most exciting technologies of our time. Large companies, such as Google, Facebook, Apple, Amazon, and IBM, heavily invest in machine learning research and applications for good reason. While it may seem that machine learning has become the buzzword of our age, it is certainly not just hype. This exciting field opens up the way to new possibilities and has become indispensable to our daily lives. Think about talking to the voice assistant on our smartphones, recommending the right product for our customers, preventing credit card fraud, filtering out spam from our email inboxes, and detecting and diagnosing medical diseases; the list goes on and on.

Get started with machine learning

If you want to become a machine learning practitioner or a better problem solver, or maybe you are even considering a career in machine learning research, then this book is for you! For a novice, the theoretical concepts behind machine learning can be quite overwhelming, but the many practical books that have been published in recent years will help you to get started in machine learning by implementing powerful learning algorithms.

Practice and theory

Being exposed to practical code examples and working through example applications of machine learning are great ways to dive into this field. Also, concrete examples help to illustrate the broader concepts by putting the learned material directly into action. However, remember that with great power comes great responsibility!

[ xiii ]

Preface

In addition to offering hands-on experience with machine learning using the Python programming language and Python-based machine learning libraries, this book introduces the mathematical concepts behind machine learning algorithms, which are essential for using machine learning successfully. Thus, this book is different from a purely practical book; this is a book that discusses the necessary details regarding machine learning concepts and offers intuitive, yet informative, explanations on how machine learning algorithms work, how to use them, and, most importantly, how to avoid the most common pitfalls.

Why Python?

Before we dive deeper into the machine learning field, let's answer your most important question: "Why Python?" The answer is simple: it is powerful, yet very accessible. Python has become the most popular programming language for data science because it allows us to forget the tedious parts of programming and offers us an environment where we can quickly jot down our ideas and put concepts directly into action.

Explore the machine learning field

If you type "machine learning" as a search term into Google Scholar, it will return an overwhelmingly large number—3,250,000 publications. Of course, we cannot discuss all the nitty-gritty details of all the different algorithms and applications that have emerged in the last 60 years. However, in this book, we will embark on an exciting journey, covering all the essential topics and concepts to give you a head start in this field. If you find that your thirst for knowledge is not satisfied, you can use the many useful resources that this book references to follow up on the essential breakthroughs in this field. We, the authors, can truly say that the study of machine learning made us better scientists, thinkers, and problem solvers. In this book, we want to share this knowledge with you. Knowledge is gained by learning, the key to this is enthusiasm, and the real mastery of skills can only be achieved through practice. The road ahead may be bumpy on occasions, and some topics may be more challenging than others, but we hope that you will embrace this opportunity and focus on the reward. Remember that we are on this journey together, and throughout this book, we will add many powerful techniques to your arsenal that will help you to solve even the toughest problems the data-driven way.

[ xiv ]

Preface

Who this book is for

If you have already studied machine learning theory in detail, this book will show you how to put your knowledge into practice. If you have used machine learning techniques before and want to gain more insight into how machine learning actually works, this book is also for you. Don't worry if you are completely new to the machine learning field; you have even more reason to be excited! This is a promise that machine learning will change the way you think about the problems you want to solve and show you how to tackle them by unlocking the power of data. If you want to find out how to use Python to start answering critical questions about your data, pick up Python Machine Learning. Whether you want to start from scratch or extend your data science knowledge, this is an essential and unmissable resource.

What this book covers

Chapter 1, Giving Computers the Ability to Learn from Data, introduces the main subareas of machine learning used to tackle various problem tasks. In addition, it discusses the essential steps for creating a typical machine learning model-building pipeline that will guide us through the following chapters. Chapter 2, Training Simple Machine Learning Algorithms for Classification, goes back to the origin of machine learning and introduces binary perceptron classifiers and adaptive linear neurons. This chapter is a gentle introduction to the fundamentals of pattern classification and focuses on the interplay of optimization algorithms and machine learning. Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, describes the essential machine learning algorithms for classification and provides practical examples using one of the most popular and comprehensive open source machine learning libraries, scikit-learn. Chapter 4, Building Good Training Datasets – Data Preprocessing, discusses how to deal with the most common problems in unprocessed datasets, such as missing data. It also discusses several approaches to identify the most informative features in datasets and how to prepare variables of different types as proper inputs for machine learning algorithms. Chapter 5, Compressing Data via Dimensionality Reduction, describes the essential techniques to reduce the number of features in a dataset to smaller sets, while retaining most of their useful and discriminatory information. It also discusses the standard approach to dimensionality reduction via principal component analysis and compares it to supervised and nonlinear transformation techniques. [ xv ]

Preface

Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, discusses the dos and don'ts for estimating the performance of predictive models. Moreover, it discusses different metrics for measuring the performance of our models and techniques for fine-tuning machine learning algorithms. Chapter 7, Combining Different Models for Ensemble Learning, introduces the different concepts of combining multiple learning algorithms effectively. It explores how to build ensembles of experts to overcome the weaknesses of individual learners, resulting in more accurate and reliable predictions. Chapter 8, Applying Machine Learning to Sentiment Analysis, discusses the essential steps for transforming textual data into meaningful representations for machine learning algorithms to predict the opinions of people based on their writing. Chapter 9, Embedding a Machine Learning Model into a Web Application, continues with the predictive model from the previous chapter and walks through the essential steps of developing web applications with embedded machine learning models. Chapter 10, Predicting Continuous Target Variables with Regression Analysis, discusses the essential techniques for modeling linear relationships between target and response variables to make predictions on a continuous scale. After introducing different linear models, it also talks about polynomial regression and tree-based approaches. Chapter 11, Working with Unlabeled Data – Clustering Analysis, shifts the focus to a different subarea of machine learning, unsupervised learning. It covers algorithms from three fundamental families of clustering algorithms that find groups of objects that share a certain degree of similarity. Chapter 12, Implementing a Multilayer Artificial Neural Network from Scratch, extends the concept of gradient-based optimization, which we first introduced in Chapter 2, Training Simple Machine Learning Algorithms for Classification. In this chapter, we will build powerful, multilayer neural networks (NNs) based on the popular backpropagation algorithm in Python. Chapter 13, Parallelizing Neural Network Training with TensorFlow, builds upon the knowledge from the previous chapter to provide a practical guide for training NNs more efficiently. The focus of this chapter is on TensorFlow 2.0, an open source Python library that allows us to utilize multiple cores of modern graphics processing units (GPUs) and construct deep NNs from common building blocks via the userfriendly Keras API.

[ xvi ]

Preface

Chapter 14, Going Deeper – The Mechanics of TensorFlow, picks up where the previous chapter left off and introduces the more advanced concepts and functionality of TensorFlow 2.0. TensorFlow is an extraordinarily vast and sophisticated library, and this chapter walks through concepts such as compiling code into a static graph for faster execution and defining trainable model parameters. In addition, this chapter provides additional hands-on experience of training deep neural networks using TensorFlow's Keras API, as well as TensorFlow's pre-made Estimators. Chapter 15, Classifying Images with Deep Convolutional Neural Networks, introduces convolutional neural networks (CNNs). A CNN represents a particular type of deep NN architecture that is particularly well suited for image datasets. Due to their superior performance compared to traditional approaches, CNNs are now widely used in computer vision to achieve state-of-the-art results for various image recognition tasks. Throughout this chapter, you will learn how convolutional layers can be used as powerful feature extractors for image classification. Chapter 16, Modeling Sequential Data Using Recurrent Neural Networks, introduces another popular NN architecture for deep learning that is especially well suited to working with text and other types of sequential data and time series data. As a warm-up exercise, this chapter introduces recurrent NNs for predicting the sentiment of movie reviews. Then, the chapter covers teaching recurrent networks to digest information from books in order to generate entirely new text. Chapter 17, Generative Adversarial Networks for Synthesizing New Data, introduces a popular adversarial training regime for NNs that can be used to generate new, realistic-looking images. The chapter starts with a brief introduction to autoencoders, a particular type of NN architecture that can be used for data compression. The chapter then shows how to combine the decoder part of an autoencoder with a second NN that can distinguish between real and synthesized images. By letting two NNs compete with each other in an adversarial training approach, you will implement a generative adversarial network that generates new handwritten digits. Lastly, after introducing the basic concepts of generative adversarial networks, the chapter introduces improvements that can stabilize the adversarial training, such as using the Wasserstein distance metric. Chapter 18, Reinforcement Learning for Decision Making in Complex Environments, covers a subcategory of machine learning that is commonly used for training robots and other autonomous systems. This chapter starts by introducing the basics of reinforcement learning (RL) to make you familiar with agent/environment interactions, the reward process of RL systems, and the concept of learning from experience. The chapter covers the two main categories of RL, model-based and model-free RL. After learning about basic algorithmic approaches, such as Monte Carlo- and temporal distance-based learning, you will implement and train an agent that can navigate a grid world environment using the Q-learning algorithm. [ xvii ]

Preface

Finally, this chapter introduces the deep Q-learning algorithm, which is a variant of Q-learning that uses deep NNs.

What you need for this book

The execution of the code examples provided in this book requires an installation of Python 3.7.0 or newer on macOS, Linux, or Microsoft Windows. We will make frequent use of Python's essential libraries for scientific computing throughout this book, including SciPy, NumPy, scikit-learn, Matplotlib, and pandas. The first chapter will provide you with instructions and useful tips to set up your Python environment and these core libraries. We will add additional libraries to our repertoire, and installation instructions are provided in the respective chapters, for example, the NLTK library for natural language processing in Chapter 8, Applying Machine Learning to Sentiment Analysis, the Flask web framework in Chapter 9, Embedding a Machine Learning Model into a Web Application, and TensorFlow for efficient NN training on GPUs in Chapter 13 to Chapter 18.

To get the most out of this book

Now that you are the proud owner of a Packt book, we have a number of things to help you to get the most from your purchase.

Download the example code files

You can download the example code files from your account at http://www. packtpub.com for all the Packt Publishing books you have purchased. If you purchased this book elsewhere, you can visit http://www.packtpub.com/support and register to have the files emailed directly to you. You can download the code files by following these steps: 1. Log in or register at http://www.packt.com. 2. Select the Support tab. 3. Click on Code Downloads. 4. Enter the name of the book in the Search box and follow the on-screen instructions. Once the file is downloaded, please make sure that you unzip or extract the folder using the latest version of: • WinRAR/7-Zip for Windows [ xviii ]

Preface

• Zipeg/iZip/UnRarX for Mac • 7-Zip/PeaZip for Linux Alternatively, if you have obtained a copy of the book from elsewhere or do not wish to create an account at Packt, all code examples are also available for download through GitHub at https://github.com/rasbt/python-machine-learningbook-3rd-edition. All code in this book is also available in the form of Jupyter notebooks, and a short introduction can be found in the code directory of Chapter 1, Giving Computers the Ability to Learn from Data, at https://github.com/rasbt/python-machinelearning-book-3rd-edition/tree/master/ch01#pythonjupyter-notebook. For more information about the general Jupyter Notebook GUI, please see the official documentation at https://jupyter-notebook.readthedocs.io/en/stable/. While we recommend using Jupyter Notebook for executing code interactively, all code examples are available in both a Python script (for example, ch02/ch02. py) and a Jupyter Notebook format (for example, ch02/ch02.ipynb). Furthermore, we recommend that you view the README.md file that accompanies each individual chapter for additional information and updates (for example, https://github.

com/rasbt/python-machine-learning-book-3rd-edition/blob/master/ch01/ README.md).

We also have other code bundles from our rich catalog of books and videos available at https://github.com/PacktPublishing/. Check them out!

Download the color images

We also provide you with a PDF file that has color images of the screenshots/ diagrams used in this book. The color images will help you to better understand the changes in the output. You can download this file from https://static. packt-cdn.com/downloads/9781789955750_ColorImages.pdf. In addition, lower resolution color images are embedded in the code notebooks of this book that come bundled with the example code files.

Conventions used

In this book, you will find a number of text styles that distinguish between different kinds of information. Here are some examples of these styles and an explanation of their meaning. Code words in text are shown as follows: "And already installed packages can be updated via the --upgrade flag." [ xix ]

Preface

A block of code is set as follows: >>> >>> >>> >>> >>> >>> ... >>> ... >>> >>> >>> >>>

import matplotlib.pyplot as plt import numpy as np y = df.iloc[0:100, 4].values y = np.where(y == 'Iris-setosa', -1, 1) X = df.iloc[0:100, [0, 2]].values plt.scatter(X[:50, 0], X[:50, 1], color='red', marker='x', label='setosa') plt.scatter(X[50:100, 0], X[50:100, 1], color='blue', marker='o', label='versicolor') plt.xlabel('sepal length') plt.ylabel('petal length') plt.legend(loc='upper left') plt.show()

Any command-line input or output is written as follows: > dot -Tpng tree.dot -o tree.png

New terms and important words are shown in bold. Words that you see on the screen, for example, in menus or dialog boxes, appear in the text like this: "Clicking the Next button moves you to the next screen."

Warnings or important notes appear in a box like this.

Tips and tricks appear like this.

Get in touch

Feedback from our readers is always welcome. General feedback: If you have questions about any aspect of this book, mention the book title in the subject of your message and email us at [email protected] packtpub.com.

[ xx ]

Preface

Errata: Although we have taken every care to ensure the accuracy of our content, mistakes do happen. If you find a mistake in one of our books—maybe a mistake in the text or the code—we would be grateful if you could report this to us. By doing so, you can save other readers from frustration and help us to improve subsequent versions of this book. If you find any errata, please report them by visiting www. packtpub.com/support/errata, selecting your book, clicking on the Errata Submission Form link, and entering the details of your errata. Once your errata are verified, your submission will be accepted and the errata will be uploaded to our website or added to any list of existing errata under the Errata section of that title. To view the previously submitted errata, go to https://www.packtpub.com/books/ content/support and enter the name of the book in the search field. The required

information will appear under the Errata section.

Piracy: If you come across any illegal copies of our works in any form on the Internet, we would be grateful if you would provide us with the location address or website name. Please contact us at [email protected] with a link to the material. If you are interested in becoming an author: If there is a topic that you have expertise in and you are interested in either writing or contributing to a book, please visit authors.packtpub.com.

Reviews

Please leave a review. Once you have read and used this book, why not leave a review on the site that you purchased it from? Potential readers can then see and use your unbiased opinion to make purchase decisions, we at Packt can understand what you think about our products, and our authors can see your feedback on their book. Thank you! For more information about Packt, please visit packt.com.

[ xxi ]

Giving Computers the Ability to Learn from Data In my opinion, machine learning, the application and science of algorithms that make sense of data, is the most exciting field of all the computer sciences! We are living in an age where data comes in abundance; using self-learning algorithms from the field of machine learning, we can turn this data into knowledge. Thanks to the many powerful open source libraries that have been developed in recent years, there has probably never been a better time to break into the machine learning field and learn how to utilize powerful algorithms to spot patterns in data and make predictions about future events. In this chapter, you will learn about the main concepts and different types of machine learning. Together with a basic introduction to the relevant terminology, we will lay the groundwork for successfully using machine learning techniques for practical problem solving. In this chapter, we will cover the following topics: • The general concepts of machine learning • The three types of learning and basic terminology • The building blocks for successfully designing machine learning systems • Installing and setting up Python for data analysis and machine learning

Building intelligent machines to transform data into knowledge

In this age of modern technology, there is one resource that we have in abundance: a large amount of structured and unstructured data. In the second half of the 20th century, machine learning evolved as a subfield of artificial intelligence (AI) involving self-learning algorithms that derive knowledge from data in order to make predictions. [1]

Giving Computers the Ability to Learn from Data

Instead of requiring humans to manually derive rules and build models from analyzing large amounts of data, machine learning offers a more efficient alternative for capturing the knowledge in data to gradually improve the performance of predictive models and make data-driven decisions. Not only is machine learning becoming increasingly important in computer science research, but it is also playing an ever-greater role in our everyday lives. Thanks to machine learning, we enjoy robust email spam filters, convenient text and voice recognition software, reliable web search engines, and challenging chess-playing programs. Hopefully soon, we will add safe and efficient self-driving cars to this list. Also, notable progress has been made in medical applications; for example, researchers demonstrated that deep learning models can detect skin cancer with near-human accuracy (https://www.nature.com/articles/nature21056). Another milestone was recently achieved by researchers at DeepMind, who used deep learning to predict 3D protein structures, outperforming physics-based approaches for the first time (https://deepmind.com/blog/alphafold/).

The three different types of machine learning

In this section, we will take a look at the three types of machine learning: supervised learning, unsupervised learning, and reinforcement learning. We will learn about the fundamental differences between the three different learning types and, using conceptual examples, we will develop an understanding of the practical problem domains where they can be applied:

[2]

Chapter 1

Making predictions about the future with supervised learning

The main goal in supervised learning is to learn a model from labeled training data that allows us to make predictions about unseen or future data. Here, the term "supervised" refers to a set of training examples (data inputs) where the desired output signals (labels) are already known. The following figure summarizes a typical supervised learning workflow, where the labeled training data is passed to a machine learning algorithm for fitting a predictive model that can make predictions on new, unlabeled data inputs:

Considering the example of email spam filtering, we can train a model using a supervised machine learning algorithm on a corpus of labeled emails, which are correctly marked as spam or non-spam, to predict whether a new email belongs to either of the two categories. A supervised learning task with discrete class labels, such as in the previous email spam filtering example, is also called a classification task. Another subcategory of supervised learning is regression, where the outcome signal is a continuous value.

Classification for predicting class labels

Classification is a subcategory of supervised learning where the goal is to predict the categorical class labels of new instances, based on past observations. Those class labels are discrete, unordered values that can be understood as the group memberships of the instances. The previously mentioned example of email spam detection represents a typical example of a binary classification task, where the machine learning algorithm learns a set of rules in order to distinguish between two possible classes: spam and non-spam emails. [3]

Giving Computers the Ability to Learn from Data

The following figure illustrates the concept of a binary classification task given 30 training examples; 15 training examples are labeled as the negative class (minus signs) and 15 training examples are labeled as the positive class (plus signs). In this scenario, our dataset is two-dimensional, which means that each example has two values associated with it: x1 and x2. Now, we can use a supervised machine learning algorithm to learn a rule—the decision boundary represented as a dashed line—that can separate those two classes and classify new data into each of those two categories given its x1 and x2 values:

However, the set of class labels does not have to be of a binary nature. The predictive model learned by a supervised learning algorithm can assign any class label that was presented in the training dataset to a new, unlabeled instance. A typical example of a multiclass classification task is handwritten character recognition. We can collect a training dataset that consists of multiple handwritten examples of each letter in the alphabet. The letters ("A," "B," "C," and so on) will represent the different unordered categories or class labels that we want to predict. Now, if a user provides a new handwritten character via an input device, our predictive model will be able to predict the correct letter in the alphabet with certain accuracy. However, our machine learning system will be unable to correctly recognize any of the digits between 0 and 9, for example, if they were not part of the training dataset.

Regression for predicting continuous outcomes

We learned in the previous section that the task of classification is to assign categorical, unordered labels to instances. A second type of supervised learning is the prediction of continuous outcomes, which is also called regression analysis. In regression analysis, we are given a number of predictor (explanatory) variables and a continuous response variable (outcome), and we try to find a relationship between those variables that allows us to predict an outcome. [4]

Chapter 1

Note that in the field of machine learning, the predictor variables are commonly called "features," and the response variables are usually referred to as "target variables." We will adopt these conventions throughout this book. For example, let's assume that we are interested in predicting the math SAT scores of students. If there is a relationship between the time spent studying for the test and the final scores, we could use it as training data to learn a model that uses the study time to predict the test scores of future students who are planning to take this test. Regression toward the mean The term "regression" was devised by Francis Galton in his article Regression towards Mediocrity in Hereditary Stature in 1886. Galton described the biological phenomenon that the variance of height in a population does not increase over time. He observed that the height of parents is not passed on to their children, but instead, their children's height regresses toward the population mean.

The following figure illustrates the concept of linear regression. Given a feature variable, x, and a target variable, y, we fit a straight line to this data that minimizes the distance—most commonly the average squared distance—between the data points and the fitted line. We can now use the intercept and slope learned from this data to predict the target variable of new data:

[5]

Giving Computers the Ability to Learn from Data

Solving interactive problems with reinforcement learning

Another type of machine learning is reinforcement learning. In reinforcement learning, the goal is to develop a system (agent) that improves its performance based on interactions with the environment. Since the information about the current state of the environment typically also includes a so-called reward signal, we can think of reinforcement learning as a field related to supervised learning. However, in reinforcement learning, this feedback is not the correct ground truth label or value, but a measure of how well the action was measured by a reward function. Through its interaction with the environment, an agent can then use reinforcement learning to learn a series of actions that maximizes this reward via an exploratory trial-and-error approach or deliberative planning. A popular example of reinforcement learning is a chess engine. Here, the agent decides upon a series of moves depending on the state of the board (the environment), and the reward can be defined as win or lose at the end of the game:

There are many different subtypes of reinforcement learning. However, a general scheme is that the agent in reinforcement learning tries to maximize the reward through a series of interactions with the environment. Each state can be associated with a positive or negative reward, and a reward can be defined as accomplishing an overall goal, such as winning or losing a game of chess. For instance, in chess, the outcome of each move can be thought of as a different state of the environment. To explore the chess example further, let's think of visiting certain configurations on the chess board as being associated with states that will more likely lead to winning—for instance, removing an opponent's chess piece from the board or threatening the queen. Other positions, however, are associated with states that will more likely result in losing the game, such as losing a chess piece to the opponent in the following turn. Now, in the game of chess, the reward (either positive for winning or negative for losing the game) will not be given until the end of the game. In addition, the final reward will also depend on how the opponent plays. For example, the opponent may sacrifice the queen but eventually win the game. [6]

Chapter 1

Reinforcement learning is concerned with learning to choose a series of actions that maximizes the total reward, which could be earned either immediately after taking an action or via delayed feedback.

Discovering hidden structures with unsupervised learning

In supervised learning, we know the right answer beforehand when we train a model, and in reinforcement learning, we define a measure of reward for particular actions carried out by the agent. In unsupervised learning, however, we are dealing with unlabeled data or data of unknown structure. Using unsupervised learning techniques, we are able to explore the structure of our data to extract meaningful information without the guidance of a known outcome variable or reward function.

Finding subgroups with clustering

Clustering is an exploratory data analysis technique that allows us to organize a pile of information into meaningful subgroups (clusters) without having any prior knowledge of their group memberships. Each cluster that arises during the analysis defines a group of objects that share a certain degree of similarity but are more dissimilar to objects in other clusters, which is why clustering is also sometimes called unsupervised classification. Clustering is a great technique for structuring information and deriving meaningful relationships from data. For example, it allows marketers to discover customer groups based on their interests, in order to develop distinct marketing programs. The following figure illustrates how clustering can be applied to organizing unlabeled data into three distinct groups based on the similarity of their features, x1 and x2:

[7]

Giving Computers the Ability to Learn from Data

Dimensionality reduction for data compression

Another subfield of unsupervised learning is dimensionality reduction. Often, we are working with data of high dimensionality—each observation comes with a high number of measurements—that can present a challenge for limited storage space and the computational performance of machine learning algorithms. Unsupervised dimensionality reduction is a commonly used approach in feature preprocessing to remove noise from data, which can also degrade the predictive performance of certain algorithms, and compress the data onto a smaller dimensional subspace while retaining most of the relevant information. Sometimes, dimensionality reduction can also be useful for visualizing data; for example, a high-dimensional feature set can be projected onto one-, two-, or three-dimensional feature spaces in order to visualize it via 2D or 3D scatterplots or histograms. The following figure shows an example where nonlinear dimensionality reduction was applied to compress a 3D Swiss Roll onto a new 2D feature subspace:

Introduction to the basic terminology and notations

Now that we have discussed the three broad categories of machine learning— supervised, unsupervised, and reinforcement learning—let's have a look at the basic terminology that we will be using throughout this book. The following subsection covers the common terms we will be using when referring to different aspects of a dataset, as well as the mathematical notation to communicate more precisely and efficiently.

[8]

Chapter 1

As machine learning is a vast field and very interdisciplinary, you are guaranteed to encounter many different terms that refer to the same concepts sooner rather than later. The second subsection collects many of the most commonly used terms that are found in machine learning literature, which may be useful to you as a reference section when reading more diverse machine learning literature.

Notation and conventions used in this book

The following table depicts an excerpt of the Iris dataset, which is a classic example in the field of machine learning. The Iris dataset contains the measurements of 150 Iris flowers from three different species—Setosa, Versicolor, and Virginica. Here, each flower example represents one row in our dataset, and the flower measurements in centimeters are stored as columns, which we also call the features of the dataset:

To keep the notation and implementation simple yet efficient, we will make use of some of the basics of linear algebra. In the following chapters, we will use a matrix and vector notation to refer to our data. We will follow the common convention to represent each example as a separate row in a feature matrix, X, where each feature is stored as a separate column.

[9]

Giving Computers the Ability to Learn from Data

The Iris dataset, consisting of 150 examples and four features, can then be written as a 150 × 4 matrix, 𝑿𝑿 ∈ ℝ𝟏𝟏𝟏𝟏𝟏𝟏×𝟒𝟒 : (1)

𝑥𝑥1

(2)

𝑥𝑥1 ⋮ (150) [𝑥𝑥1

(1)

𝑥𝑥2

(2)

𝑥𝑥2 ⋮ (150) 𝑥𝑥2

(1)

𝑥𝑥3

(2)

𝑥𝑥3 ⋮ (150) 𝑥𝑥3

(1)

𝑥𝑥4

(2)

𝑥𝑥4 ⋮ (150) 𝑥𝑥4 ]

Notational conventions For the rest of this book, unless noted otherwise, we will use the superscript i to refer to the ith training example, and the subscript j to refer to the jth dimension of the training dataset. We will use lowercase, bold-face letters to refer to vectors (𝒙𝒙 ∈ ℝ𝒏𝒏×𝟏𝟏 ) and uppercase, bold-face letters to refer to matrices (𝑿𝑿 ∈ ℝ𝒏𝒏×𝒎𝒎 ). To refer to single elements in a vector or matrix, we (𝑛𝑛) will write the letters in italics ( 𝑥𝑥 (𝑛𝑛) or 𝑥𝑥𝑚𝑚 , respectively). For example, 𝑥𝑥 (150) refers to the first dimension of flower 1 example 150, the sepal length. Thus, each row in this feature matrix represents one flower instance and can be written as a four-dimensional row vector, 𝒙𝒙(𝑖𝑖) ∈ ℝ𝟏𝟏×𝟒𝟒 : 𝒙𝒙(𝒊𝒊) = [𝑥𝑥1(𝑖𝑖)

(𝑖𝑖)

𝑥𝑥2

(𝑖𝑖)

𝑥𝑥3

(𝑖𝑖) 𝑥𝑥4 ]

And each feature dimension is a 150-dimensional column vector, 𝒙𝒙(𝑖𝑖) ∈ ℝ𝟏𝟏𝟏𝟏𝟏𝟏×𝟏𝟏. For example: 𝒙𝒙𝒋𝒋 =

(1)

𝑥𝑥𝑗𝑗

(2)

𝑥𝑥𝑗𝑗 ... (150) [𝑥𝑥𝑗𝑗 ]

Similarly, we will store the target variables (here, class labels) as a 150-dimensional column vector: 𝑦𝑦 (1) 𝒚𝒚 = [ … ] (y ϵ {Setosa, Versicolor, Virginica}) 𝑦𝑦 (150)

[ 10 ]

Chapter 1

Machine learning terminology

Machine learning is a vast field and also very interdisciplinary as it brings together many scientists from other areas of research. As it happens, many terms and concepts have been rediscovered or redefined and may already be familiar to you but appear under different names. For your convenience, in the following list, you can find a selection of commonly used terms and their synonyms that you may find useful when reading this book and machine learning literature in general: • Training example: A row in a table representing the dataset and synonymous with an observation, record, instance, or sample (in most contexts, sample refers to a collection of training examples). • Training: Model fitting, for parametric models similar to parameter estimation. • Feature, abbrev. x: A column in a data table or data (design) matrix. Synonymous with predictor, variable, input, attribute, or covariate. • Target, abbrev. y: Synonymous with outcome, output, response variable, dependent variable, (class) label, and ground truth. • Loss function: Often used synonymously with a cost function. Sometimes the loss function is also called an error function. In some literature, the term "loss" refers to the loss measured for a single data point, and the cost is a measurement that computes the loss (average or summed) over the entire dataset.

A roadmap for building machine learning systems In previous sections, we discussed the basic concepts of machine learning and the three different types of learning. In this section, we will discuss the other important parts of a machine learning system accompanying the learning algorithm.

[ 11 ]

Giving Computers the Ability to Learn from Data

The following diagram shows a typical workflow for using machine learning in predictive modeling, which we will discuss in the following subsections:

Preprocessing – getting data into shape

Let's begin with discussing the roadmap for building machine learning systems. Raw data rarely comes in the form and shape that is necessary for the optimal performance of a learning algorithm. Thus, the preprocessing of the data is one of the most crucial steps in any machine learning application. If we take the Iris flower dataset from the previous section as an example, we can think of the raw data as a series of flower images from which we want to extract meaningful features. Useful features could be the color, hue, and intensity of the flowers, or the height, length, and width of the flowers. Many machine learning algorithms also require that the selected features are on the same scale for optimal performance, which is often achieved by transforming the features in the range [0, 1] or a standard normal distribution with zero mean and unit variance, as we will see in later chapters.

[ 12 ]

Chapter 1

Some of the selected features may be highly correlated and therefore redundant to a certain degree. In those cases, dimensionality reduction techniques are useful for compressing the features onto a lower dimensional subspace. Reducing the dimensionality of our feature space has the advantage that less storage space is required, and the learning algorithm can run much faster. In certain cases, dimensionality reduction can also improve the predictive performance of a model if the dataset contains a large number of irrelevant features (or noise); that is, if the dataset has a low signal-to-noise ratio. To determine whether our machine learning algorithm not only performs well on the training dataset but also generalizes well to new data, we also want to randomly divide the dataset into a separate training and test dataset. We use the training dataset to train and optimize our machine learning model, while we keep the test dataset until the very end to evaluate the final model.

Training and selecting a predictive model

As you will see in later chapters, many different machine learning algorithms have been developed to solve different problem tasks. An important point that can be summarized from David Wolpert's famous No free lunch theorems is that we can't get learning "for free" (The Lack of A Priori Distinctions Between Learning Algorithms, D.H. Wolpert, 1996; No free lunch theorems for optimization, D.H. Wolpert and W.G. Macready, 1997). We can relate this concept to the popular saying, "I suppose it is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail" (Abraham Maslow, 1966). For example, each classification algorithm has its inherent biases, and no single classification model enjoys superiority if we don't make any assumptions about the task. In practice, it is therefore essential to compare at least a handful of different algorithms in order to train and select the best performing model. But before we can compare different models, we first have to decide upon a metric to measure performance. One commonly used metric is classification accuracy, which is defined as the proportion of correctly classified instances. One legitimate question to ask is this: how do we know which model performs well on the final test dataset and real-world data if we don't use this test dataset for the model selection, but keep it for the final model evaluation? In order to address the issue embedded in this question, different techniques summarized as "cross-validation" can be used. In cross-validation, we further divide a dataset into training and validation subsets in order to estimate the generalization performance of the model. Finally, we also cannot expect that the default parameters of the different learning algorithms provided by software libraries are optimal for our specific problem task. Therefore, we will make frequent use of hyperparameter optimization techniques that help us to fine-tune the performance of our model in later chapters.

[ 13 ]

Giving Computers the Ability to Learn from Data

We can think of those hyperparameters as parameters that are not learned from the data but represent the knobs of a model that we can turn to improve its performance. This will become much clearer in later chapters when we see actual examples.

Evaluating models and predicting unseen data instances

After we have selected a model that has been fitted on the training dataset, we can use the test dataset to estimate how well it performs on this unseen data to estimate the so-called generalization error. If we are satisfied with its performance, we can now use this model to predict new, future data. It is important to note that the parameters for the previously mentioned procedures, such as feature scaling and dimensionality reduction, are solely obtained from the training dataset, and the same parameters are later reapplied to transform the test dataset, as well as any new data instances—the performance measured on the test data may be overly optimistic otherwise.

Using Python for machine learning

Python is one of the most popular programming languages for data science and thanks to its very active developer and open source community, a large number of useful libraries for scientific computing and machine learning have been developed. Although the performance of interpreted languages, such as Python, for computation-intensive tasks is inferior to lower-level programming languages, extension libraries such as NumPy and SciPy have been developed that build upon lower-layer Fortran and C implementations for fast vectorized operations on multidimensional arrays. For machine learning programming tasks, we will mostly refer to the scikit-learn library, which is currently one of the most popular and accessible open source machine learning libraries. In the later chapters, when we focus on a subfield of machine learning called deep learning, we will use the latest version of the TensorFlow library, which specializes in training so-called deep neural network models very efficiently by utilizing graphics cards.

Installing Python and packages from the Python Package Index

Python is available for all three major operating systems—Microsoft Windows, macOS, and Linux—and the installer, as well as the documentation, can be downloaded from the official Python website: https://www.python.org. [ 14 ]

Chapter 1

This book is written for Python version 3.7 or higher, and it is recommended that you use the most recent version of Python 3 that is currently available. Some of the code may also be compatible with Python 2.7, but as the official support for Python 2.7 ends in 2019, and the majority of open source libraries have already stopped supporting Python 2.7 (https://python3statement.org), we strongly advise that you use Python 3.7 or newer. The additional packages that we will be using throughout this book can be installed via the pip installer program, which has been part of the Python Standard Library since Python 3.3. More information about pip can be found at https://docs. python.org/3/installing/index.html. After we have successfully installed Python, we can execute pip from the terminal to install additional Python packages: pip install SomePackage

Already installed packages can be updated via the --upgrade flag: pip install SomePackage --upgrade

Using the Anaconda Python distribution and package manager

A highly recommended alternative Python distribution for scientific computing is Anaconda by Continuum Analytics. Anaconda is a free—including commercial use—enterprise-ready Python distribution that bundles all the essential Python packages for data science, math, and engineering into one user-friendly, crossplatform distribution. The Anaconda installer can be downloaded at https://docs. anaconda.com/anaconda/install/, and an Anaconda quick start guide is available at https://docs.anaconda.com/anaconda/user-guide/getting-started/. After successfully installing Anaconda, we can install new Python packages using the following command: conda install SomePackage

Existing packages can be updated using the following command: conda update SomePackage

[ 15 ]

Giving Computers the Ability to Learn from Data

Packages for scientific computing, data science, and machine learning

Throughout this book, we will mainly use NumPy's multidimensional arrays to store and manipulate data. Occasionally, we will make use of pandas, which is a library built on top of NumPy that provides additional higher-level data manipulation tools that make working with tabular data even more convenient. To augment your learning experience and visualize quantitative data, which is often extremely useful to make sense of it, we will use the very customizable Matplotlib library. The version numbers of the major Python packages that were used to write this book are mentioned in the following list. Please make sure that the version numbers of your installed packages are equal to, or greater than, these version numbers to ensure that the code examples run correctly: • NumPy 1.17.4 • SciPy 1.3.1 • scikit-learn 0.22.0 • Matplotlib 3.1.0 • pandas 0.25.3

Summary

In this chapter, we explored machine learning at a very high level and familiarized ourselves with the big picture and major concepts that we are going to explore in the following chapters in more detail. We learned that supervised learning is composed of two important subfields: classification and regression. While classification models allow us to categorize objects into known classes, we can use regression analysis to predict the continuous outcomes of target variables. Unsupervised learning not only offers useful techniques for discovering structures in unlabeled data, but it can also be useful for data compression in feature preprocessing steps. We briefly went over the typical roadmap for applying machine learning to problem tasks, which we will use as a foundation for deeper discussions and handson examples in the following chapters. Finally, we set up our Python environment and installed and updated the required packages to get ready to see machine learning in action.

[ 16 ]

Chapter 1

Later in this book, in addition to machine learning itself, we will introduce different techniques to preprocess a dataset, which will help you to get the best performance out of different machine learning algorithms. While we will cover classification algorithms quite extensively throughout the book, we will also explore different techniques for regression analysis and clustering. We have an exciting journey ahead, covering many powerful techniques in the vast field of machine learning. However, we will approach machine learning one step at a time, building upon our knowledge gradually throughout the chapters of this book. In the following chapter, we will start this journey by implementing one of the earliest machine learning algorithms for classification, which will prepare us for Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, where we will cover more advanced machine learning algorithms using the scikit-learn open source machine learning library.

[ 17 ]

Training Simple Machine Learning Algorithms for Classification In this chapter, we will make use of two of the first algorithmically described machine learning algorithms for classification: the perceptron and adaptive linear neurons. We will start by implementing a perceptron step by step in Python and training it to classify different flower species in the Iris dataset. This will help us to understand the concept of machine learning algorithms for classification and how they can be efficiently implemented in Python. Discussing the basics of optimization using adaptive linear neurons will then lay the groundwork for using more sophisticated classifiers via the scikit-learn machine learning library in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn. The topics that we will cover in this chapter are as follows: • Building an understanding of machine learning algorithms • Using pandas, NumPy, and Matplotlib to read in, process, and visualize data • Implementing linear classification algorithms in Python

[ 19 ]

Training Simple Machine Learning Algorithms for Classification

Artificial neurons – a brief glimpse into the early history of machine learning

Before we discuss the perceptron and related algorithms in more detail, let's take a brief tour of the beginnings of machine learning. Trying to understand how the biological brain works, in order to design artificial intelligence (AI), Warren McCulloch and Walter Pitts published the first concept of a simplified brain cell, the so-called McCulloch-Pitts (MCP) neuron, in 1943 (A Logical Calculus of the Ideas Immanent in Nervous Activity, W. S. McCulloch and W. Pitts, Bulletin of Mathematical Biophysics, 5(4): 115-133, 1943). Biological neurons are interconnected nerve cells in the brain that are involved in the processing and transmitting of chemical and electrical signals, which is illustrated in the following figure:

McCulloch and Pitts described such a nerve cell as a simple logic gate with binary outputs; multiple signals arrive at the dendrites, they are then integrated into the cell body, and, if the accumulated signal exceeds a certain threshold, an output signal is generated that will be passed on by the axon. Only a few years later, Frank Rosenblatt published the first concept of the perceptron learning rule based on the MCP neuron model (The Perceptron: A Perceiving and Recognizing Automaton, F. Rosenblatt, Cornell Aeronautical Laboratory, 1957). With his perceptron rule, Rosenblatt proposed an algorithm that would automatically learn the optimal weight coefficients that would then be multiplied with the input features in order to make the decision of whether a neuron fires (transmits a signal) or not. In the context of supervised learning and classification, such an algorithm could then be used to predict whether a new data point belongs to one class or the other.

[ 20 ]

Chapter 2

The formal definition of an artificial neuron

More formally, we can put the idea behind artificial neurons into the context of a binary classification task where we refer to our two classes as 1 (positive class) and –1 (negative class) for simplicity. We can then define a decision function ( 𝜙𝜙(𝑧𝑧) ) that takes a linear combination of certain input values, x, and a corresponding weight vector, w, where z is the so-called net input 𝑧𝑧 = 𝑤𝑤1 𝑥𝑥1 + 𝑤𝑤2 𝑥𝑥2 +. . . + 𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚 : 𝑤𝑤1 𝒘𝒘 = [ ⋮ ] , 𝑤𝑤𝑚𝑚

𝑥𝑥1 𝒙𝒙 = [ ⋮ ] 𝑥𝑥𝑚𝑚

1 𝜙𝜙(𝑧𝑧) = { −1

if 𝑧𝑧 ≥ 𝜃𝜃, otherwise.

Now, if the net input of a particular example, 𝒙𝒙(𝑖𝑖), is greater than a defined threshold, 𝜃𝜃 , we predict class 1, and class –1 otherwise. In the perceptron algorithm, the decision function, 𝜙𝜙(∙), is a variant of a unit step function:

For simplicity, we can bring the threshold, 𝜃𝜃 , to the left side of the equation and define a weight-zero as 𝑤𝑤0 = −𝜃𝜃 and 𝑥𝑥0 = 1 so that we write z in a more compact form:

And:

𝑧𝑧 = 𝑤𝑤0 𝑥𝑥0 + 𝑤𝑤1 𝑥𝑥1 + ⋯ + 𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚 = 𝒘𝒘𝑇𝑇 𝒙𝒙 𝜙𝜙(𝑧𝑧) = {

1 if 𝑧𝑧 ≥ 0, −1 otherwise.

In machine learning literature, the negative threshold, or weight, 𝑤𝑤0 = −𝜃𝜃 , is usually called the bias unit.

[ 21 ]

Training Simple Machine Learning Algorithms for Classification

Linear algebra basics: dot product and matrix transpose In the following sections, we will often make use of basic notations from linear algebra. For example, we will abbreviate the sum of the products of the values in x and w using a vector dot product, whereas superscript T stands for transpose, which is an operation that transforms a column vector into a row vector and vice versa: 𝑚𝑚

𝑧𝑧 = 𝑤𝑤0 𝑥𝑥0 + 𝑤𝑤1 𝑥𝑥1 + ⋯ + 𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚 = ∑ 𝑥𝑥𝑗𝑗 𝑤𝑤𝑗𝑗 = 𝒘𝒘𝑇𝑇 𝒙𝒙 𝑗𝑗=0

For example:

[1 2

4 3] × [ 5 ] = 1 × 4 + 2 × 5 + 3 × 6 = 32 6

Furthermore, the transpose operation can also be applied to matrices to reflect it over its diagonal, for example: 1 [3 5

2 𝑇𝑇 1 4] = [ 2 6

3 5 ] 4 6

Please note that the transpose operation is strictly only defined for matrices; however, in the context of machine learning, we refer to 𝑛𝑛 × 1 or 1 × 𝑚𝑚 matrices when we use the term "vector."

In this book, we will only use very basic concepts from linear algebra; however, if you need a quick refresher, please take a look at Zico Kolter's excellent Linear Algebra Review and Reference, which is freely available at http://www.cs.cmu. edu/~zkolter/course/linalg/linalg_notes.pdf.

The following figure illustrates how the net input 𝑧𝑧 = 𝒘𝒘𝑇𝑇 𝒙𝒙 is squashed into a binary output (–1 or 1) by the decision function of the perceptron (left subfigure) and how it can be used to discriminate between two linearly separable classes (right subfigure):

[ 22 ]

Chapter 2

The perceptron learning rule

The whole idea behind the MCP neuron and Rosenblatt's thresholded perceptron model is to use a reductionist approach to mimic how a single neuron in the brain works: it either fires or it doesn't. Thus, Rosenblatt's initial perceptron rule is fairly simple, and the perceptron algorithm can be summarized by the following steps: 1. Initialize the weights to 0 or small random numbers. 2. For each training example, 𝒙𝒙(𝑖𝑖) :

a. Compute the output value, 𝑦𝑦̂ . b. Update the weights.

Here, the output value is the class label predicted by the unit step function that we defined earlier, and the simultaneous update of each weight, 𝑤𝑤𝑗𝑗, in the weight vector, w, can be more formally written as:

𝑤𝑤𝑗𝑗 ∶= 𝑤𝑤𝑗𝑗 + Δ𝑤𝑤𝑗𝑗

[ 23 ]

Training Simple Machine Learning Algorithms for Classification

The update value for 𝑤𝑤𝑗𝑗 (or change in 𝑤𝑤𝑗𝑗 ), which we refer to as Δ𝑤𝑤𝑗𝑗 , is calculated by the perceptron learning rule as follows: (𝑖𝑖)

Δ𝑤𝑤𝑗𝑗 = 𝜂𝜂(𝑦𝑦 (𝑖𝑖) − 𝑦𝑦̂ (𝑖𝑖) )𝑥𝑥𝑗𝑗

Where 𝜂𝜂 is the learning rate (typically a constant between 0.0 and 1.0), 𝑦𝑦 (𝑖𝑖) is the true class label of the ith training example, and 𝑦𝑦̂ (𝑖𝑖) is the predicted class label. It is important to note that all weights in the weight vector are being updated simultaneously, which means that we don't recompute the predicted label, 𝑦𝑦̂ (𝑖𝑖), before all of the weights are updated via the respective update values, Δ𝑤𝑤𝑗𝑗 . Concretely, for a two-dimensional dataset, we would write the update as: Δ𝑤𝑤0 = 𝜂𝜂(𝑦𝑦 (𝑖𝑖) − 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 (𝑖𝑖) ) (𝑖𝑖)

Δ𝑤𝑤1 = 𝜂𝜂(𝑦𝑦 (𝑖𝑖) − 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 (𝑖𝑖) )𝑥𝑥1

(𝑖𝑖)

Δ𝑤𝑤2 = 𝜂𝜂(𝑦𝑦 (𝑖𝑖) − 𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 (𝑖𝑖) )𝑥𝑥2

Before we implement the perceptron rule in Python, let's go through a simple thought experiment to illustrate how beautifully simple this learning rule really is. In the two scenarios where the perceptron predicts the class label correctly, the weights remain unchanged, since the update values are 0: (1) (2)

𝑦𝑦 (𝑖𝑖) = −1,

𝑦𝑦̂ (𝑖𝑖) = −1,

𝑦𝑦 (𝑖𝑖) = 1,

𝑦𝑦̂ (𝑖𝑖) = 1,

𝑦𝑦 (𝑖𝑖) = 1,

𝑦𝑦̂ (𝑖𝑖) = −1,

(𝑖𝑖)

∆𝑤𝑤𝑗𝑗 = 𝜂𝜂(−1 − (−1))𝑥𝑥𝑗𝑗 = 0 (𝑖𝑖)

∆𝑤𝑤𝑗𝑗 = 𝜂𝜂(1 − 1)𝑥𝑥𝑗𝑗 = 0

However, in the case of a wrong prediction, the weights are being pushed toward the direction of the positive or negative target class: (3) (4)

𝑦𝑦 (𝑖𝑖) = −1,

𝑦𝑦̂ (𝑖𝑖) = 1,

(𝑖𝑖)

(𝑖𝑖)

∆𝑤𝑤𝑗𝑗 = 𝜂𝜂(1 − (−1))𝑥𝑥𝑗𝑗 = 𝜂𝜂(2)𝑥𝑥𝑗𝑗 (𝑖𝑖)

∆𝑤𝑤𝑗𝑗 = 𝜂𝜂(−1 − 1)𝑥𝑥𝑗𝑗

(𝑖𝑖)

= 𝜂𝜂(−2)𝑥𝑥𝑗𝑗 (𝑖𝑖)

To get a better understanding of the multiplicative factor, 𝑥𝑥𝑗𝑗 , let's go through another simple example, where:

𝑦𝑦̂ (𝑖𝑖) = −1,

𝑦𝑦 (𝑖𝑖) = +1, [ 24 ]

𝜂𝜂 = 1

Chapter 2 (𝑖𝑖)

Let's assume that 𝑥𝑥𝑗𝑗 = 0.5, and we misclassify this example as –1. In this case, we would increase the corresponding weight by 1 so that the net input, 𝑥𝑥𝑗𝑗(𝑖𝑖) × 𝑤𝑤𝑗𝑗, would be more positive the next time we encounter this example, and thus be more likely to be above the threshold of the unit step function to classify the example as +1: Δ𝑤𝑤𝑗𝑗 = (1 − (− 1))0.5 = (2)0.5 = 1 (𝑖𝑖)

The weight update is proportional to the value of 𝑥𝑥𝑗𝑗 . For instance, if we have (𝑖𝑖) another example, 𝑥𝑥𝑗𝑗 = 2, that is incorrectly classified as –1, we will push the decision boundary by an even larger extent to classify this example correctly the next time: Δ𝑤𝑤𝑗𝑗 = (1(𝑖𝑖) − (− 1)(𝑖𝑖) )2(𝑖𝑖) = (2)2(𝑖𝑖) = 4

It is important to note that the convergence of the perceptron is only guaranteed if the two classes are linearly separable and the learning rate is sufficiently small (interested readers can find the mathematical proof in my lecture notes: https:// sebastianraschka.com/pdf/lecture-notes/stat479ss19/L03_perceptron_ slides.pdf.). If the two classes can't be separated by a linear decision boundary,

we can set a maximum number of passes over the training dataset (epochs) and/ or a threshold for the number of tolerated misclassifications—the perceptron would never stop updating the weights otherwise:

Downloading the example code If you bought this book directly from Packt, you can download the example code files from your account at http://www. packtpub.com. If you purchased this book elsewhere, you can download all code examples and datasets directly from https://github.com/rasbt/python-machine-learningbook-3rd-edition.

[ 25 ]

Training Simple Machine Learning Algorithms for Classification

Now, before we jump into the implementation in the next section, what you just learned can be summarized in a simple diagram that illustrates the general concept of the perceptron:

The preceding diagram illustrates how the perceptron receives the inputs of an example, x, and combines them with the weights, w, to compute the net input. The net input is then passed on to the threshold function, which generates a binary output of –1 or +1—the predicted class label of the example. During the learning phase, this output is used to calculate the error of the prediction and update the weights.

Implementing a perceptron learning algorithm in Python

In the previous section, we learned how Rosenblatt's perceptron rule works; let's now implement it in Python and apply it to the Iris dataset that we introduced in Chapter 1, Giving Computers the Ability to Learn from Data.

An object-oriented perceptron API

We will take an object-oriented approach to defining the perceptron interface as a Python class, which will allow us to initialize new Perceptron objects that can learn from data via a fit method, and make predictions via a separate predict method. As a convention, we append an underscore (_) to attributes that are not created upon the initialization of the object, but we do this by calling the object's other methods, for example, self.w_.

[ 26 ]

Chapter 2

Additional resources for Python's scientific computing stack If you are not yet familiar with Python's scientific libraries or need a refresher, please see the following resources: • NumPy: https://sebastianraschka.com/pdf/ books/dlb/appendix_f_numpy-intro.pdf

• pandas: https://pandas.pydata.org/pandasdocs/stable/10min.html

• Matplotlib: https://matplotlib.org/tutorials/ introductory/usage.html

The following is the implementation of a perceptron in Python: import numpy as np class Perceptron(object): """Perceptron classifier. Parameters -----------eta : float Learning rate (between 0.0 and 1.0) n_iter : int Passes over the training dataset. random_state : int Random number generator seed for random weight initialization. Attributes ----------w_ : 1d-array Weights after fitting. errors_ : list Number of misclassifications (updates) in each epoch. """ def __init__(self, eta=0.01, n_iter=50, random_state=1): self.eta = eta self.n_iter = n_iter [ 27 ]

Training Simple Machine Learning Algorithms for Classification self.random_state = random_state def fit(self, X, y): """Fit training data. Parameters ---------X : {array-like}, shape = [n_examples, n_features] Training vectors, where n_examples is the number of examples and n_features is the number of features. y : array-like, shape = [n_examples] Target values. Returns ------self : object """ rgen = np.random.RandomState(self.random_state) self.w_ = rgen.normal(loc=0.0, scale=0.01, size=1 + X.shape[1]) self.errors_ = [] for _ in range(self.n_iter): errors = 0 for xi, target in zip(X, y): update = self.eta * (target - self.predict(xi)) self.w_[1:] += update * xi self.w_[0] += update errors += int(update != 0.0) self.errors_.append(errors) return self def net_input(self, X): """Calculate net input""" return np.dot(X, self.w_[1:]) + self.w_[0] def predict(self, X): """Return class label after unit step""" return np.where(self.net_input(X) >= 0.0, 1, -1)

Using this perceptron implementation, we can now initialize new Perceptron objects with a given learning rate, eta, and the number of epochs, n_iter (passes over the training dataset). [ 28 ]

Chapter 2

Via the fit method, we initialize the weights in self.w_ to a vector, ℝ𝑚𝑚+1 , where m stands for the number of dimensions (features) in the dataset, and we add 1 for the first element in this vector that represents the bias unit. Remember that the first element in this vector, self.w_[0], represents the so-called bias unit that we discussed earlier. Also notice that this vector contains small random numbers drawn from a normal distribution with standard deviation 0.01 via rgen.normal(loc=0.0, scale=0.01, size=1 + X.shape[1]), where rgen is a NumPy random number generator that we seeded with a user-specified random seed so that we can reproduce previous results if desired. It is important to keep in mind that we don't initialize the weights to zero because the learning rate, 𝜂𝜂 (eta), only has an effect on the classification outcome if the weights are initialized to non-zero values. If all the weights are initialized to zero, the learning rate parameter, eta, affects only the scale of the weight vector, not the direction. If you are familiar with trigonometry, consider a vector, 𝑣𝑣1 = [1 2 3] , where the angle between 𝑣𝑣1 and a vector, 𝑣𝑣2 = 0.5 × 𝑣𝑣1, would be exactly zero, as demonstrated by the following code snippet: >>> v1 = np.array([1, 2, 3]) >>> v2 = 0.5 * v1 >>> np.arccos(v1.dot(v2) / (np.linalg.norm(v1) * ... np.linalg.norm(v2))) 0.0

Here, np.arccos is the trigonometric inverse cosine, and np.linalg.norm is a function that computes the length of a vector (our decision to draw the random numbers from a random normal distribution—for example, instead of from a uniform distribution—and to use a standard deviation of 0.01 was arbitrary; remember, we are just interested in small random values to avoid the properties of all-zero vectors, as discussed earlier). NumPy array indexing NumPy indexing for one-dimensional arrays works similarly to Python lists using the square-bracket ([]) notation. For twodimensional arrays, the first indexer refers to the row number and the second indexer to the column number. For example, we would use X[2, 3] to select the third row and fourth column of a two-dimensional array, X.

After the weights have been initialized, the fit method loops over all individual examples in the training dataset and updates the weights according to the perceptron learning rule that we discussed in the previous section. [ 29 ]

Training Simple Machine Learning Algorithms for Classification

The class labels are predicted by the predict method, which is called in the fit method during training to get the class label for the weight update; but predict can also be used to predict the class labels of new data after we have fitted our model. Furthermore, we also collect the number of misclassifications during each epoch in the self.errors_ list so that we can later analyze how well our perceptron performed during the training. The np.dot function that is used in the net_input method simply calculates the vector dot product, 𝒘𝒘𝑇𝑇 𝒙𝒙 . Instead of using NumPy to calculate the vector dot product between two arrays, a and b, via a.dot(b) or np.dot(a, b), we could also perform the calculation in pure Python via sum([i * j for i, j in zip(a, b)]). However, the advantage of using NumPy over classic Python for loop structures is that its arithmetic operations are vectorized. Vectorization means that an elemental arithmetic operation is automatically applied to all elements in an array. By formulating our arithmetic operations as a sequence of instructions on an array, rather than performing a set of operations for each element at a time, we can make better use of our modern central processing unit (CPU) architectures with single instruction, multiple data (SIMD) support. Furthermore, NumPy uses highly optimized linear algebra libraries, such as Basic Linear Algebra Subprograms (BLAS) and Linear Algebra Package (LAPACK), that have been written in C or Fortran. Lastly, NumPy also allows us to write our code in a more compact and intuitive way using the basics of linear algebra, such as vector and matrix dot products.

Training a perceptron model on the Iris dataset

To test our perceptron implementation, we will restrict the following analyses and examples in the remainder of this chapter to two feature variables (dimensions). Although the perceptron rule is not restricted to two dimensions, considering only two features, sepal length and petal length, will allow us to visualize the decision regions of the trained model in a scatter plot for learning purposes. Note that we will also only consider two flower classes, Setosa and Versicolor, from the Iris dataset for practical reasons—remember, the perceptron is a binary classifier. However, the perceptron algorithm can be extended to multi-class classification— for example, the one-vs.-all (OvA) technique.

[ 30 ]

Chapter 2

The OvA method for multi-class classification OvA, which is sometimes also called one-vs.-rest (OvR), is a technique that allows us to extend any binary classifier to multiclass problems. Using OvA, we can train one classifier per class, where the particular class is treated as the positive class and the examples from all other classes are considered negative classes. If we were to classify a new, unlabeled data instance, we would use our n classifiers, where n is the number of class labels, and assign the class label with the highest confidence to the particular instance we want to classify. In the case of the perceptron, we would use OvA to choose the class label that is associated with the largest absolute net input value.

First, we will use the pandas library to load the Iris dataset directly from the UCI Machine Learning Repository into a DataFrame object and print the last five lines via the tail method to check that the data was loaded correctly: >>> import os >>> import pandas as pd >>> s = os.path.join('https://archive.ics.uci.edu', 'ml', ... 'machine-learning-databases', ... 'iris','iris.data') >>> print('URL:', s) URL: https://archive.ics.uci.edu/ml/machine-learning-databases/iris/ iris.data >>> df = pd.read_csv(s, ... header=None, ... encoding='utf-8') >>> df.tail()

[ 31 ]

Training Simple Machine Learning Algorithms for Classification

Loading the Iris dataset You can find a copy of the Iris dataset (and all other datasets used in this book) in the code bundle of this book, which you can use if you are working offline or the UCI server at https:// archive.ics.uci.edu/ml/machine-learning-databases/ iris/iris.data is temporarily unavailable. For instance, to

load the Iris dataset from a local directory, you can replace this line, df = pd.read_csv( 'https://archive.ics.uci.edu/ml/' 'machine-learning-databases/iris/iris.data', header=None, encoding='utf-8')

with the following one: df = pd.read_csv( 'your/local/path/to/iris.data', header=None, encoding='utf-8')

Next, we extract the first 100 class labels that correspond to the 50 Iris-setosa and 50 Iris-versicolor flowers and convert the class labels into the two integer class labels, 1 (versicolor) and -1 (setosa), that we assign to a vector, y, where the values method of a pandas DataFrame yields the corresponding NumPy representation. Similarly, we extract the first feature column (sepal length) and the third feature column (petal length) of those 100 training examples and assign them to a feature matrix, X, which we can visualize via a two-dimensional scatterplot: >>> import matplotlib.pyplot as plt >>> import numpy as np >>> # select setosa and versicolor >>> y = df.iloc[0:100, 4].values >>> y = np.where(y == 'Iris-setosa', -1, 1) >>> # extract sepal length and petal length >>> X = df.iloc[0:100, [0, 2]].values >>> # plot data >>> plt.scatter(X[:50, 0], X[:50, 1], ... color='red', marker='o', label='setosa') >>> plt.scatter(X[50:100, 0], X[50:100, 1], ... color='blue', marker='x', label='versicolor')

[ 32 ]

Chapter 2 >>> >>> >>> >>>

plt.xlabel('sepal length [cm]') plt.ylabel('petal length [cm]') plt.legend(loc='upper left') plt.show()

After executing the preceding code example, we should now see the following scatterplot:

The preceding scatterplot shows the distribution of flower examples in the Iris dataset along the two feature axes: petal length and sepal length (measured in centimeters). In this two-dimensional feature subspace, we can see that a linear decision boundary should be sufficient to separate Setosa from Versicolor flowers. Thus, a linear classifier such as the perceptron should be able to classify the flowers in this dataset perfectly. Now, it's time to train our perceptron algorithm on the Iris data subset that we just extracted. Also, we will plot the misclassification error for each epoch to check whether the algorithm converged and found a decision boundary that separates the two Iris flower classes: >>> >>> >>> ... >>> >>> >>>

ppn = Perceptron(eta=0.1, n_iter=10) ppn.fit(X, y) plt.plot(range(1, len(ppn.errors_) + 1), ppn.errors_, marker='o') plt.xlabel('Epochs') plt.ylabel('Number of updates') plt.show()

[ 33 ]

Training Simple Machine Learning Algorithms for Classification

After executing the preceding code, we should see the plot of the misclassification errors versus the number of epochs, as shown in the following graph:

As we can see in the preceding plot, our perceptron converged after the sixth epoch and should now be able to classify the training examples perfectly. Let's implement a small convenience function to visualize the decision boundaries for two-dimensional datasets: from matplotlib.colors import ListedColormap def plot_decision_regions(X, y, classifier, resolution=0.02): # setup marker generator and color map markers = ('s', 'x', 'o', '^', 'v') colors = ('red', 'blue', 'lightgreen', 'gray', 'cyan') cmap = ListedColormap(colors[:len(np.unique(y))]) # plot the decision surface x1_min, x1_max = X[:, 0].min() - 1, X[:, 0].max() + 1 x2_min, x2_max = X[:, 1].min() - 1, X[:, 1].max() + 1 xx1, xx2 = np.meshgrid(np.arange(x1_min, x1_max, resolution), np.arange(x2_min, x2_max, resolution)) Z = classifier.predict(np.array([xx1.ravel(), xx2.ravel()]).T) Z = Z.reshape(xx1.shape) plt.contourf(xx1, xx2, Z, alpha=0.3, cmap=cmap) plt.xlim(xx1.min(), xx1.max()) plt.ylim(xx2.min(), xx2.max()) # plot class examples for idx, cl in enumerate(np.unique(y)): plt.scatter(x=X[y == cl, 0], [ 34 ]

Chapter 2 y=X[y == cl, 1], alpha=0.8, c=colors[idx], marker=markers[idx], label=cl, edgecolor='black')

First, we define a number of colors and markers and create a colormap from the list of colors via ListedColormap. Then, we determine the minimum and maximum values for the two features and use those feature vectors to create a pair of grid arrays, xx1 and xx2, via the NumPy meshgrid function. Since we trained our perceptron classifier on two feature dimensions, we need to flatten the grid arrays and create a matrix that has the same number of columns as the Iris training subset so that we can use the predict method to predict the class labels, Z, of the corresponding grid points. After reshaping the predicted class labels, Z, into a grid with the same dimensions as xx1 and xx2, we can now draw a contour plot via Matplotlib's contourf function, which maps the different decision regions to different colors for each predicted class in the grid array: >>> >>> >>> >>> >>>

plot_decision_regions(X, y, classifier=ppn) plt.xlabel('sepal length [cm]') plt.ylabel('petal length [cm]') plt.legend(loc='upper left') plt.show()

After executing the preceding code example, we should now see a plot of the decision regions, as shown in the following figure:

[ 35 ]

Training Simple Machine Learning Algorithms for Classification

As we can see in the plot, the perceptron learned a decision boundary that is able to classify all flower examples in the Iris training subset perfectly. Perceptron convergence Although the perceptron classified the two Iris flower classes perfectly, convergence is one of the biggest problems of the perceptron. Rosenblatt proved mathematically that the perceptron learning rule converges if the two classes can be separated by a linear hyperplane. However, if the classes cannot be separated perfectly by such a linear decision boundary, the weights will never stop updating unless we set a maximum number of epochs. Interested readers can find a summary of the proof in my lecture notes at https://sebastianraschka. com/pdf/lecture-notes/stat479ss19/L03_perceptron_ slides.pdf.

Adaptive linear neurons and the convergence of learning

In this section, we will take a look at another type of single-layer neural network (NN): ADAptive LInear NEuron (Adaline). Adaline was published by Bernard Widrow and his doctoral student Tedd Hoff only a few years after Rosenblatt's perceptron algorithm, and it can be considered an improvement on the latter (An Adaptive "Adaline" Neuron Using Chemical "Memistors", Technical Report Number 1553-2, B. Widrow and others, Stanford Electron Labs, Stanford, CA, October 1960). The Adaline algorithm is particularly interesting because it illustrates the key concepts of defining and minimizing continuous cost functions. This lays the groundwork for understanding more advanced machine learning algorithms for classification, such as logistic regression, support vector machines, and regression models, which we will discuss in future chapters. The key difference between the Adaline rule (also known as the Widrow-Hoff rule) and Rosenblatt's perceptron is that the weights are updated based on a linear activation function rather than a unit step function like in the perceptron. In Adaline, this linear activation function, 𝜙𝜙(𝑧𝑧), is simply the identity function of the net input, so that:

𝜙𝜙(𝒘𝒘𝑇𝑇 𝒙𝒙) = 𝒘𝒘𝑇𝑇 𝒙𝒙 [ 36 ]

Chapter 2

While the linear activation function is used for learning the weights, we still use a threshold function to make the final prediction, which is similar to the unit step function that we covered earlier. The main differences between the perceptron and Adaline algorithm are highlighted in the following figure:

As the illustration indicates, the Adaline algorithm compares the true class labels with the linear activation function's continuous valued output to compute the model error and update the weights. In contrast, the perceptron compares the true class labels to the predicted class labels.

Minimizing cost functions with gradient descent

One of the key ingredients of supervised machine learning algorithms is a defined objective function that is to be optimized during the learning process. This objective function is often a cost function that we want to minimize. In the case of Adaline, we can define the cost function, J, to learn the weights as the sum of squared errors (SSE) between the calculated outcome and the true class label:

𝐽𝐽(𝒘𝒘) =

2 1 ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 2 𝑖𝑖

[ 37 ]

Training Simple Machine Learning Algorithms for Classification

The term 12 is just added for our convenience and will make it easier to derive the gradient of the cost or loss function with respect to the weight parameters, as we will see in the following paragraphs. The main advantage of this continuous linear activation function, in contrast to the unit step function, is that the cost function becomes differentiable. Another nice property of this cost function is that it is convex; thus, we can use a very simple yet powerful optimization algorithm called gradient descent to find the weights that minimize our cost function to classify the examples in the Iris dataset. As illustrated in the following figure, we can describe the main idea behind gradient descent as climbing down a hill until a local or global cost minimum is reached. In each iteration, we take a step in the opposite direction of the gradient, where the step size is determined by the value of the learning rate, as well as the slope of the gradient:

Using gradient descent, we can now update the weights by taking a step in the opposite direction of the gradient, ∇J(𝐰𝐰), of our cost function, J(𝐰𝐰):

𝒘𝒘 ∶= 𝒘𝒘 + ∆𝒘𝒘

The weight change, ∆𝒘𝒘 , is defined as the negative gradient multiplied by the learning rate, 𝜂𝜂:

∆𝒘𝒘 = −𝜂𝜂∇J(𝐰𝐰)

To compute the gradient of the cost function, we need to compute the partial derivative of the cost function with respect to each weight, 𝑤𝑤𝑗𝑗: 𝜕𝜕𝜕𝜕 (𝑖𝑖) = − ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑥𝑥𝑗𝑗 𝜕𝜕𝑤𝑤𝑗𝑗 𝑖𝑖

[ 38 ]

Chapter 2

So we can write the update of weight 𝑤𝑤𝑗𝑗 as: ∆𝑤𝑤𝑗𝑗 = − 𝜂𝜂

𝜕𝜕𝜕𝜕 (𝑖𝑖) = 𝜂𝜂 ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑥𝑥𝑗𝑗 𝜕𝜕𝑤𝑤𝑗𝑗 𝑖𝑖

Since we update all weights simultaneously, our Adaline learning rule becomes:

𝒘𝒘 ∶= 𝒘𝒘 + ∆𝒘𝒘

The squared error derivative

If you are familiar with calculus, the partial derivative of the SSE cost function with respect to the jth weight can be obtained as follows: 2 𝜕𝜕𝜕𝜕 𝜕𝜕 1 = ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝜕𝜕𝑤𝑤𝑗𝑗 𝜕𝜕𝑤𝑤𝑗𝑗 2 𝑖𝑖

=

=

2 1 𝜕𝜕 ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 2 𝜕𝜕𝑤𝑤𝑗𝑗 𝑖𝑖

𝜕𝜕 1 ∑ 2 (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝜕𝜕𝑤𝑤𝑗𝑗 2 𝑖𝑖

= ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑖𝑖

𝜕𝜕 (𝑖𝑖) (𝑖𝑖) (𝑦𝑦 (𝑖𝑖) − ∑(𝑤𝑤𝑗𝑗 𝑥𝑥𝑗𝑗 )) 𝜕𝜕𝑤𝑤𝑗𝑗 𝑖𝑖

(𝑖𝑖)

= ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) (−𝑥𝑥𝑗𝑗 ) 𝑖𝑖

(𝑖𝑖)

= − ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑥𝑥𝑗𝑗 𝑖𝑖

Although the Adaline learning rule looks identical to the perceptron rule, we should note that 𝜙𝜙(𝑧𝑧 (𝑖𝑖) ) with 𝑧𝑧 (𝑖𝑖) = 𝒘𝒘𝑇𝑇 𝒙𝒙(𝑖𝑖) is a real number and not an integer class label. Furthermore, the weight update is calculated based on all examples in the training dataset (instead of updating the weights incrementally after each training example), which is why this approach is also referred to as batch gradient descent.

[ 39 ]

Training Simple Machine Learning Algorithms for Classification

Implementing Adaline in Python

Since the perceptron rule and Adaline are very similar, we will take the perceptron implementation that we defined earlier and change the fit method so that the weights are updated by minimizing the cost function via gradient descent: class AdalineGD(object): """ADAptive LInear NEuron classifier. Parameters -----------eta : float Learning rate (between 0.0 and 1.0) n_iter : int Passes over the training dataset. random_state : int Random number generator seed for random weight initialization. Attributes ----------w_ : 1d-array Weights after fitting. cost_ : list Sum-of-squares cost function value in each epoch. """ def __init__(self, eta=0.01, n_iter=50, random_state=1): self.eta = eta self.n_iter = n_iter self.random_state = random_state def fit(self, X, y): """ Fit training data. Parameters ---------X : {array-like}, shape = [n_examples, n_features] Training vectors, where n_examples is the number of examples and n_features is the number of features. y : array-like, shape = [n_examples] Target values. Returns

[ 40 ]

Chapter 2 ------self : object """ rgen = np.random.RandomState(self.random_state) self.w_ = rgen.normal(loc=0.0, scale=0.01, size=1 + X.shape[1]) self.cost_ = [] for i in range(self.n_iter): net_input = self.net_input(X) output = self.activation(net_input) errors = (y - output) self.w_[1:] += self.eta * X.T.dot(errors) self.w_[0] += self.eta * errors.sum() cost = (errors**2).sum() / 2.0 self.cost_.append(cost) return self def net_input(self, X): """Calculate net input""" return np.dot(X, self.w_[1:]) + self.w_[0] def activation(self, X): """Compute linear activation""" return X def predict(self, X): """Return class label after unit step""" return np.where(self.activation(self.net_input(X)) >= 0.0, 1, -1)

Instead of updating the weights after evaluating each individual training example, as in the perceptron, we calculate the gradient based on the whole training dataset via self.eta * errors.sum() for the bias unit (zero-weight), and via self.eta * X.T.dot(errors) for the weights 1 to m, where X.T.dot(errors) is a matrix-vector multiplication between our feature matrix and the error vector. Please note that the activation method has no effect in the code since it is simply an identity function. Here, we added the activation function (computed via the activation method) to illustrate the general concept with regard to how information flows through a single-layer NN: features from the input data, net input, activation, and output. [ 41 ]

Training Simple Machine Learning Algorithms for Classification

In the next chapter, we will learn about a logistic regression classifier that uses a nonidentity, nonlinear activation function. We will see that a logistic regression model is closely related to Adaline, with the only difference being its activation and cost function. Now, similar to the previous perceptron implementation, we collect the cost values in a self.cost_ list to check whether the algorithm converged after training. Matrix multiplication Performing a matrix multiplication is similar to calculating a vector dot-product where each row in the matrix is treated as a single row vector. This vectorized approach represents a more compact notation and results in a more efficient computation using NumPy. For example: [

1 2 4 5

7 1×7+ 2×8+ 3×9 3 50 ] × [8] = [ ]=[ ] 4×7+ 5×8+ 6×9 6 122 9

Please note that in the preceding equation, we are multiplying a matrix with a vector, which is mathematically not defined. However, remember that we use the convention that this preceding vector is regarded as a 3 × 1 matrix.

In practice, it often requires some experimentation to find a good learning rate, 𝜂𝜂, for optimal convergence. So, let's choose two different learning rates, 𝜂𝜂 = 0.1 and 𝜂𝜂 = 0.0001, to start with and plot the cost functions versus the number of epochs to see how well the Adaline implementation learns from the training data. Perceptron hyperparameters The learning rate, 𝜂𝜂, (eta), as well as the number of epochs (n_iter), are the so-called hyperparameters (or tuning parameters) of the perceptron and Adaline learning algorithms. In Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, we will take a look at different techniques to automatically find the values of different hyperparameters that yield optimal performance of the classification model.

Let's now plot the cost against the number of epochs for the two different learning rates: >>> fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(10, 4)) >>> ada1 = AdalineGD(n_iter=10, eta=0.01).fit(X, y) [ 42 ]

Chapter 2

>>> ... >>> >>> >>>

ax[0].plot(range(1, len(ada1.cost_) + 1), np.log10(ada1.cost_), marker='o') ax[0].set_xlabel('Epochs') ax[0].set_ylabel('log(Sum-squared-error)') ax[0].set_title('Adaline - Learning rate 0.01')

>>> >>> ... >>> >>> >>> >>>

ada2 = AdalineGD(n_iter=10, eta=0.0001).fit(X, y) ax[1].plot(range(1, len(ada2.cost_) + 1), ada2.cost_, marker='o') ax[1].set_xlabel('Epochs') ax[1].set_ylabel('Sum-squared-error') ax[1].set_title('Adaline - Learning rate 0.0001') plt.show()

As we can see in the resulting cost-function plots, we encountered two different types of problem. The left chart shows what could happen if we choose a learning rate that is too large. Instead of minimizing the cost function, the error becomes larger in every epoch, because we overshoot the global minimum. On the other hand, we can see that the cost decreases on the right plot, but the chosen learning rate, 𝜂𝜂 = 0.0001, is so small that the algorithm would require a very large number of epochs to converge to the global cost minimum:

The following figure illustrates what might happen if we change the value of a particular weight parameter to minimize the cost function, J. The left subfigure illustrates the case of a well-chosen learning rate, where the cost decreases gradually, moving in the direction of the global minimum.

[ 43 ]

Training Simple Machine Learning Algorithms for Classification

The subfigure on the right, however, illustrates what happens if we choose a learning rate that is too large—we overshoot the global minimum:

Improving gradient descent through feature scaling

Many machine learning algorithms that we will encounter throughout this book require some sort of feature scaling for optimal performance, which we will discuss in more detail in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, and Chapter 4, Building Good Training Datasets – Data Preprocessing. Gradient descent is one of the many algorithms that benefit from feature scaling. In this section, we will use a feature scaling method called standardization, which gives our data the properties of a standard normal distribution: zero-mean and unit variance. This normalization procedure helps gradient descent learning to converge more quickly; however, it does not make the original dataset normally distributed. Standardization shifts the mean of each feature so that it is centered at zero and each feature has a standard deviation of 1 (unit variance). For instance, to standardize the jth feature, we can simply subtract the sample mean, 𝜇𝜇𝑗𝑗 , from every training example and divide it by its standard deviation, 𝜎𝜎𝑗𝑗:

𝒙𝒙′𝑗𝑗 =

𝒙𝒙𝑗𝑗 − 𝜇𝜇𝑗𝑗 𝜎𝜎𝑗𝑗

Here, 𝒙𝒙𝑗𝑗 is a vector consisting of the jth feature values of all training examples, n, and this standardization technique is applied to each feature, j, in our dataset.

One of the reasons why standardization helps with gradient descent learning is that the optimizer has to go through fewer steps to find a good or optimal solution (the global cost minimum), as illustrated in the following figure, where the subfigures represent the cost surface as a function of two model weights in a two-dimensional classification problem: [ 44 ]

Chapter 2

Standardization can easily be achieved by using the built-in NumPy methods mean and std: >>> X_std = np.copy(X) >>> X_std[:,0] = (X[:,0] - X[:,0].mean()) / X[:,0].std() >>> X_std[:,1] = (X[:,1] - X[:,1].mean()) / X[:,1].std()

After standardization, we will train Adaline again and we will see that it now converges after a small number of epochs using a learning rate of 𝜂𝜂 = 0.01: >>> ada_gd = AdalineGD(n_iter=15, eta=0.01) >>> ada_gd.fit(X_std, y) >>> >>> >>> >>> >>> >>> >>>

plot_decision_regions(X_std, y, classifier=ada_gd) plt.title('Adaline - Gradient Descent') plt.xlabel('sepal length [standardized]') plt.ylabel('petal length [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

>>> ... >>> >>> >>> >>>

plt.plot(range(1, len(ada_gd.cost_) + 1), ada_gd.cost_, marker='o') plt.xlabel('Epochs') plt.ylabel('Sum-squared-error') plt.tight_layout() plt.show()

[ 45 ]

Training Simple Machine Learning Algorithms for Classification

After executing this code, we should see a figure of the decision regions, as well as a plot of the declining cost, as shown in the following figure:

As we can see in the plots, Adaline has now converged after training on the standardized features using a learning rate of 𝜂𝜂 = 0.01. However, note that the SSE remains non-zero even though all flower examples were classified correctly.

Large-scale machine learning and stochastic gradient descent

In the previous section, we learned how to minimize a cost function by taking a step in the opposite direction of a cost gradient that is calculated from the whole training dataset; this is why this approach is sometimes also referred to as batch gradient descent. Now imagine that we have a very large dataset with millions of data points, which is not uncommon in many machine learning applications. Running batch gradient descent can be computationally quite costly in such scenarios, since we need to reevaluate the whole training dataset each time that we take one step toward the global minimum. A popular alternative to the batch gradient descent algorithm is stochastic gradient descent (SGD), which is sometimes also called iterative or online gradient descent. Instead of updating the weights based on the sum of the accumulated errors over all training examples, 𝒙𝒙(𝑖𝑖):

∆𝒘𝒘 = 𝜂𝜂 ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝒙𝒙(𝑖𝑖) 𝑖𝑖

we update the weights incrementally for each training example:

𝜂𝜂 (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝒙𝒙(𝑖𝑖) [ 46 ]

Chapter 2

Although SGD can be considered as an approximation of gradient descent, it typically reaches convergence much faster because of the more frequent weight updates. Since each gradient is calculated based on a single training example, the error surface is noisier than in gradient descent, which can also have the advantage that SGD can escape shallow local minima more readily if we are working with nonlinear cost functions, as we will see later in Chapter 12, Implementing a Multilayer Artificial Neural Network from Scratch. To obtain satisfying results via SGD, it is important to present training data in a random order; also, we want to shuffle the training dataset for every epoch to prevent cycles. Adjusting the learning rate during training In SGD implementations, the fixed learning rate, 𝜂𝜂, is often replaced by an adaptive learning rate that decreases over time, for example:

𝑐𝑐1 [𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 𝑜𝑜𝑜𝑜 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖] + 𝑐𝑐2

where 𝑐𝑐1 and 𝑐𝑐2 are constants. Note that SGD does not reach the global minimum but an area very close to it. And using an adaptive learning rate, we can achieve further annealing to the cost minimum.

Another advantage of SGD is that we can use it for online learning. In online learning, our model is trained on the fly as new training data arrives. This is especially useful if we are accumulating large amounts of data, for example, customer data in web applications. Using online learning, the system can immediately adapt to changes, and the training data can be discarded after updating the model if storage space is an issue. Mini-batch gradient descent A compromise between batch gradient descent and SGD is so-called mini-batch learning. Mini-batch learning can be understood as applying batch gradient descent to smaller subsets of the training data, for example, 32 training examples at a time. The advantage over batch gradient descent is that convergence is reached faster via mini-batches because of the more frequent weight updates. Furthermore, mini-batch learning allows us to replace the for loop over the training examples in SGD with vectorized operations leveraging concepts from linear algebra (for example, implementing a weighted sum via a dot product), which can further improve the computational efficiency of our learning algorithm.

[ 47 ]

Training Simple Machine Learning Algorithms for Classification

Since we already implemented the Adaline learning rule using gradient descent, we only need to make a few adjustments to modify the learning algorithm to update the weights via SGD. Inside the fit method, we will now update the weights after each training example. Furthermore, we will implement an additional partial_fit method, which does not reinitialize the weights, for online learning. In order to check whether our algorithm converged after training, we will calculate the cost as the average cost of the training examples in each epoch. Furthermore, we will add an option to shuffle the training data before each epoch to avoid repetitive cycles when we are optimizing the cost function; via the random_state parameter, we allow the specification of a random seed for reproducibility: class AdalineSGD(object): """ADAptive LInear NEuron classifier. Parameters -----------eta : float Learning rate (between 0.0 and 1.0) n_iter : int Passes over the training dataset. shuffle : bool (default: True) Shuffles training data every epoch if True to prevent cycles. random_state : int Random number generator seed for random weight initialization.

Attributes ----------w_ : 1d-array Weights after fitting. cost_ : list Sum-of-squares cost function value averaged over all training examples in each epoch.

""" def __init__(self, eta=0.01, n_iter=10, shuffle=True, random_state=None): self.eta = eta self.n_iter = n_iter self.w_initialized = False self.shuffle = shuffle

[ 48 ]

Chapter 2 self.random_state = random_state def fit(self, X, y): """ Fit training data. Parameters ---------X : {array-like}, shape = [n_examples, n_features] Training vectors, where n_examples is the number of examples and n_features is the number of features. y : array-like, shape = [n_examples] Target values. Returns ------self : object """ self._initialize_weights(X.shape[1]) self.cost_ = [] for i in range(self.n_iter): if self.shuffle: X, y = self._shuffle(X, y) cost = [] for xi, target in zip(X, y): cost.append(self._update_weights(xi, target)) avg_cost = sum(cost) / len(y) self.cost_.append(avg_cost) return self def partial_fit(self, X, y): """Fit training data without reinitializing the weights""" if not self.w_initialized: self._initialize_weights(X.shape[1]) if y.ravel().shape[0] > 1: for xi, target in zip(X, y): self._update_weights(xi, target) else: self._update_weights(X, y) return self def _shuffle(self, X, y): """Shuffle training data""" r = self.rgen.permutation(len(y)) [ 49 ]

Training Simple Machine Learning Algorithms for Classification return X[r], y[r] def _initialize_weights(self, m): """Initialize weights to small random numbers""" self.rgen = np.random.RandomState(self.random_state) self.w_ = self.rgen.normal(loc=0.0, scale=0.01, size=1 + m) self.w_initialized = True def _update_weights(self, xi, target): """Apply Adaline learning rule to update the weights""" output = self.activation(self.net_input(xi)) error = (target - output) self.w_[1:] += self.eta * xi.dot(error) self.w_[0] += self.eta * error cost = 0.5 * error**2 return cost def net_input(self, X): """Calculate net input""" return np.dot(X, self.w_[1:]) + self.w_[0] def activation(self, X): """Compute linear activation""" return X def predict(self, X): """Return class label after unit step""" return np.where(self.activation(self.net_input(X)) >= 0.0, 1, -1)

The _shuffle method that we are now using in the AdalineSGD classifier works as follows: via the permutation function in np.random, we generate a random sequence of unique numbers in the range 0 to 100. Those numbers can then be used as indices to shuffle our feature matrix and class label vector. We can then use the fit method to train the AdalineSGD classifier and use our plot_decision_regions to plot our training results: >>> ada_sgd = AdalineSGD(n_iter=15, eta=0.01, random_state=1) >>> ada_sgd.fit(X_std, y) >>> plot_decision_regions(X_std, y, classifier=ada_sgd) >>> plt.title('Adaline - Stochastic Gradient Descent') >>> plt.xlabel('sepal length [standardized]') [ 50 ]

Chapter 2 >>> >>> >>> >>> >>> ... >>> >>> >>> >>>

plt.ylabel('petal length [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show() plt.plot(range(1, len(ada_sgd.cost_) + 1), ada_sgd.cost_, marker='o') plt.xlabel('Epochs') plt.ylabel('Average Cost') plt.tight_layout() plt.show()

The two plots that we obtain from executing the preceding code example are shown in the following figure:

As you can see, the average cost goes down pretty quickly, and the final decision boundary after 15 epochs looks similar to the batch gradient descent Adaline. If we want to update our model, for example, in an online learning scenario with streaming data, we could simply call the partial_fit method on individual training examples—for instance ada_sgd.partial_fit(X_std[0, :], y[0]).

Summary

In this chapter, we gained a good understanding of the basic concepts of linear classifiers for supervised learning. After we implemented a perceptron, we saw how we can train adaptive linear neurons efficiently via a vectorized implementation of gradient descent and online learning via SGD. Now that we have seen how to implement simple classifiers in Python, we are ready to move on to the next chapter, where we will use the Python scikit-learn machine learning library to get access to more advanced and powerful machine learning classifiers, which are commonly used in academia as well as in industry. [ 51 ]

Training Simple Machine Learning Algorithms for Classification

The object-oriented approach that we used to implement the perceptron and Adaline algorithms will help with understanding the scikit-learn API, which is implemented based on the same core concepts that we used in this chapter: the fit and predict methods. Based on these core concepts, we will learn about logistic regression for modeling class probabilities and support vector machines for working with nonlinear decision boundaries. In addition, we will introduce a different class of supervised learning algorithms, tree-based algorithms, which are commonly combined into robust ensemble classifiers.

[ 52 ]

A Tour of Machine Learning Classifiers Using scikit-learn In this chapter, we will take a tour of a selection of popular and powerful machine learning algorithms that are commonly used in academia as well as in industry. While learning about the differences between several supervised learning algorithms for classification, we will also develop an appreciation of their individual strengths and weaknesses. In addition, we will take our first steps with the scikit-learn library, which offers a user-friendly and consistent interface for using those algorithms efficiently and productively. The topics that will be covered throughout this chapter are as follows: • An introduction to robust and popular algorithms for classification, such as logistic regression, support vector machines, and decision trees • Examples and explanations using the scikit-learn machine learning library, which provides a wide variety of machine learning algorithms via a userfriendly Python API • Discussions about the strengths and weaknesses of classifiers with linear and nonlinear decision boundaries

Choosing a classification algorithm

Choosing an appropriate classification algorithm for a particular problem task requires practice and experience; each algorithm has its own quirks and is based on certain assumptions. To restate the no free lunch theorem by David H. Wolpert, no single classifier works best across all possible scenarios (The Lack of A Priori Distinctions Between Learning Algorithms, Wolpert, David H, Neural Computation 8.7 (1996): 13411390). In practice, it is always recommended that you compare the performance of at least a handful of different learning algorithms to select the best model for the particular problem; these may differ in the number of features or examples, the amount of noise in a dataset, and whether the classes are linearly separable or not. [ 53 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Eventually, the performance of a classifier—computational performance as well as predictive power—depends heavily on the underlying data that is available for learning. The five main steps that are involved in training a supervised machine learning algorithm can be summarized as follows: 1. Selecting features and collecting labeled training examples. 2. Choosing a performance metric. 3. Choosing a classifier and optimization algorithm. 4. Evaluating the performance of the model. 5. Tuning the algorithm. Since the approach of this book is to build machine learning knowledge step by step, we will mainly focus on the main concepts of the different algorithms in this chapter and revisit topics such as feature selection and preprocessing, performance metrics, and hyperparameter tuning for more detailed discussions later in the book.

First steps with scikit-learn – training a perceptron

In Chapter 2, Training Simple Machine Learning Algorithms for Classification, you learned about two related learning algorithms for classification, the perceptron rule and Adaline, which we implemented in Python and NumPy by ourselves. Now we will take a look at the scikit-learn API, which, as mentioned, combines a userfriendly and consistent interface with a highly optimized implementation of several classification algorithms. The scikit-learn library offers not only a large variety of learning algorithms, but also many convenient functions to preprocess data and to fine-tune and evaluate our models. We will discuss this in more detail, together with the underlying concepts, in Chapter 4, Building Good Training Datasets – Data Preprocessing, and Chapter 5, Compressing Data via Dimensionality Reduction. To get started with the scikit-learn library, we will train a perceptron model similar to the one that we implemented in Chapter 2. For simplicity, we will use the already familiar Iris dataset throughout the following sections. Conveniently, the Iris dataset is already available via scikit-learn, since it is a simple yet popular dataset that is frequently used for testing and experimenting with algorithms. Similar to the previous chapter, we will only use two features from the Iris dataset for visualization purposes. We will assign the petal length and petal width of the 150 flower examples to the feature matrix, X, and the corresponding class labels of the flower species to the vector array, y: [ 54 ]

Chapter 3 >>> from sklearn import datasets >>> import numpy as np >>> iris = datasets.load_iris() >>> X = iris.data[:, [2, 3]] >>> y = iris.target >>> print('Class labels:', np.unique(y)) Class labels: [0 1 2]

The np.unique(y) function returned the three unique class labels stored in iris.target, and as we can see, the Iris flower class names, Iris-setosa, Iris-versicolor, and Iris-virginica, are already stored as integers (here: 0, 1, 2). Although many scikit-learn functions and class methods also work with class labels in string format, using integer labels is a recommended approach to avoid technical glitches and improve computational performance due to a smaller memory footprint; furthermore, encoding class labels as integers is a common convention among most machine learning libraries. To evaluate how well a trained model performs on unseen data, we will further split the dataset into separate training and test datasets. In Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, we will discuss the best practices around model evaluation in more detail: >>> from sklearn.model_selection import train_test_split >>> X_train, X_test, y_train, y_test = train_test_split( ... X, y, test_size=0.3, random_state=1, stratify=y)

Using the train_test_split function from scikit-learn's model_selection module, we randomly split the X and y arrays into 30 percent test data (45 examples) and 70 percent training data (105 examples). Note that the train_test_split function already shuffles the training datasets internally before splitting; otherwise, all examples from class 0 and class 1 would have ended up in the training datasets, and the test dataset would consist of 45 examples from class 2. Via the random_state parameter, we provided a fixed random seed (random_state=1) for the internal pseudo-random number generator that is used for shuffling the datasets prior to splitting. Using such a fixed random_ state ensures that our results are reproducible. Lastly, we took advantage of the built-in support for stratification via stratify=y. In this context, stratification means that the train_test_split method returns training and test subsets that have the same proportions of class labels as the input dataset. We can use NumPy's bincount function, which counts the number of occurrences of each value in an array, to verify that this is indeed the case: >>> print('Labels counts in y:', np.bincount(y)) [ 55 ]

A Tour of Machine Learning Classifiers Using scikit-learn Labels counts in y: [50 50 50] >>> print('Labels counts in y_train:', np.bincount(y_train)) Labels counts in y_train: [35 35 35] >>> print('Labels counts in y_test:', np.bincount(y_test)) Labels counts in y_test: [15 15 15]

Many machine learning and optimization algorithms also require feature scaling for optimal performance, as we saw in the gradient descent example in Chapter 2, Training Simple Machine Learning Algorithms for Classification. Here, we will standardize the features using the StandardScaler class from scikit-learn's preprocessing module: >>> >>> >>> >>> >>>

from sklearn.preprocessing import StandardScaler sc = StandardScaler() sc.fit(X_train) X_train_std = sc.transform(X_train) X_test_std = sc.transform(X_test)

Using the preceding code, we loaded the StandardScaler class from the preprocessing module and initialized a new StandardScaler object that we assigned to the sc variable. Using the fit method, StandardScaler estimated the parameters, 𝜇𝜇 (sample mean) and 𝜎𝜎 (standard deviation), for each feature dimension from the training data. By calling the transform method, we then standardized the training data using those estimated parameters, 𝜇𝜇 and 𝜎𝜎 . Note that we used the same scaling parameters to standardize the test dataset so that both the values in the training and test dataset are comparable to each other. Having standardized the training data, we can now train a perceptron model. Most algorithms in scikit-learn already support multiclass classification by default via the one-vs.-rest (OvR) method, which allows us to feed the three flower classes to the perceptron all at once. The code is as follows: >>> from sklearn.linear_model import Perceptron >>> ppn = Perceptron(eta0=0.1, random_state=1) >>> ppn.fit(X_train_std, y_train)

The scikit-learn interface will remind you of our perceptron implementation in Chapter 2, Training Simple Machine Learning Algorithms for Classification. After loading the Perceptron class from the linear_model module, we initialized a new Perceptron object and trained the model via the fit method. Here, the model parameter, eta0, is equivalent to the learning rate, eta, that we used in our own perceptron implementation, and the n_iter parameter defines the number of epochs (passes over the training dataset).

[ 56 ]

Chapter 3

As you will remember from Chapter 2, finding an appropriate learning rate requires some experimentation. If the learning rate is too large, the algorithm will overshoot the global cost minimum. If the learning rate is too small, the algorithm will require more epochs until convergence, which can make the learning slow—especially for large datasets. Also, we used the random_state parameter to ensure the reproducibility of the initial shuffling of the training dataset after each epoch. Having trained a model in scikit-learn, we can make predictions via the predict method, just like in our own perceptron implementation in Chapter 2. The code is as follows: >>> y_pred = ppn.predict(X_test_std) >>> print('Misclassified examples: %d' % (y_test != y_pred).sum()) Misclassified examples: 1

Executing the code, we can see that the perceptron misclassifies 1 out of the 45 flower examples. Thus, the misclassification error on the test dataset is approximately 0.022 or 2.2 percent ( 1/45 ≈ 0.022 ). Classification error versus accuracy Instead of the misclassification error, many machine learning practitioners report the classification accuracy of a model, which is simply calculated as follows: 1–error = 0.978 or 97.8 percent Whether we use the classification error or accuracy is merely a matter of preference.

Note that scikit-learn also implements a large variety of different performance metrics that are available via the metrics module. For example, we can calculate the classification accuracy of the perceptron on the test dataset as follows: >>> from sklearn.metrics import accuracy_score >>> print('Accuracy: %.3f' % accuracy_score(y_test, y_pred)) Accuracy: 0.978

Here, y_test are the true class labels and y_pred are the class labels that we predicted previously. Alternatively, each classifier in scikit-learn has a score method, which computes a classifier's prediction accuracy by combining the predict call with accuracy_score, as shown here: >>> print('Accuracy: %.3f' % ppn.score(X_test_std, y_test)) Accuracy: 0.978 [ 57 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Overfitting Note that we will evaluate the performance of our models based on the test dataset in this chapter. In Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, you will learn about useful techniques, including graphical analysis, such as learning curves, to detect and prevent overfitting. Overfitting, which we will return to later in this chapter, means that the model captures the patterns in the training data well but fails to generalize well to unseen data.

Finally, we can use our plot_decision_regions function from Chapter 2, Training Simple Machine Learning Algorithms for Classification, to plot the decision regions of our newly trained perceptron model and visualize how well it separates the different flower examples. However, let's add a small modification to highlight the data instances from the test dataset via small circles: from matplotlib.colors import ListedColormap import matplotlib.pyplot as plt

def plot_decision_regions(X, y, classifier, test_idx=None, resolution=0.02): # setup marker generator and color map markers = ('s', 'x', 'o', '^', 'v') colors = ('red', 'blue', 'lightgreen', 'gray', 'cyan') cmap = ListedColormap(colors[:len(np.unique(y))]) # plot the decision surface x1_min, x1_max = X[:, 0].min() - 1, X[:, 0].max() + 1 x2_min, x2_max = X[:, 1].min() - 1, X[:, 1].max() + 1 xx1, xx2 = np.meshgrid(np.arange(x1_min, x1_max, resolution), np.arange(x2_min, x2_max, resolution)) Z = classifier.predict(np.array([xx1.ravel(), xx2.ravel()]).T) Z = Z.reshape(xx1.shape) plt.contourf(xx1, xx2, Z, alpha=0.3, cmap=cmap) plt.xlim(xx1.min(), xx1.max()) plt.ylim(xx2.min(), xx2.max()) for idx, cl in enumerate(np.unique(y)): plt.scatter(x=X[y == cl, 0], y=X[y == cl, 1], alpha=0.8, c=colors[idx], marker=markers[idx], label=cl, edgecolor='black') [ 58 ]

Chapter 3

# highlight test examples if test_idx: # plot all examples X_test, y_test = X[test_idx, :], y[test_idx] plt.scatter(X_test[:, 0], X_test[:, 1], c='', edgecolor='black', alpha=1.0, linewidth=1, marker='o', s=100, label='test set')

With the slight modification that we made to the plot_decision_regions function, we can now specify the indices of the examples that we want to mark on the resulting plots. The code is as follows: >>> >>> >>> ... ... ... >>> >>> >>> >>> >>>

X_combined_std = np.vstack((X_train_std, X_test_std)) y_combined = np.hstack((y_train, y_test)) plot_decision_regions(X=X_combined_std, y=y_combined, classifier=ppn, test_idx=range(105, 150)) plt.xlabel('petal length [standardized]') plt.ylabel('petal width [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

As we can see in the resulting plot, the three flower classes cannot be perfectly separated by a linear decision boundary:

[ 59 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Remember from our discussion in Chapter 2, Training Simple Machine Learning Algorithms for Classification, that the perceptron algorithm never converges on datasets that aren't perfectly linearly separable, which is why the use of the perceptron algorithm is typically not recommended in practice. In the following sections, we will look at more powerful linear classifiers that converge to a cost minimum even if the classes are not perfectly linearly separable. Additional perceptron settings The Perceptron, as well as other scikit-learn functions and classes, often has additional parameters that we omit for clarity. You can read more about those parameters using the help function in Python (for instance, help(Perceptron)) or by going through the excellent scikit-learn online documentation at http://scikit-learn.org/stable/.

Modeling class probabilities via logistic regression

Although the perceptron rule offers a nice and easy-going introduction to machine learning algorithms for classification, its biggest disadvantage is that it never converges if the classes are not perfectly linearly separable. The classification task in the previous section would be an example of such a scenario. The reason for this is that the weights are continuously being updated, since there is always at least one misclassified training example present in each epoch. Of course, you can change the learning rate and increase the number of epochs, but be warned that the perceptron will never converge on this dataset. To make better use of our time, we will now take a look at another simple, yet more powerful, algorithm for linear and binary classification problems: logistic regression. Note that, in spite of its name, logistic regression is a model for classification, not regression.

Logistic regression and conditional probabilities

Logistic regression is a classification model that is very easy to implement and performs very well on linearly separable classes. It is one of the most widely used algorithms for classification in industry. Similar to the perceptron and Adaline, the logistic regression model in this chapter is also a linear model for binary classification. [ 60 ]

Chapter 3

Logistic regression for multiple classes Note that logistic regression can be readily generalized to multiclass settings, which is known as multinomial logistic regression or softmax regression. A more detailed coverage of multinomial logistic regression is outside the scope of this book, but the interested reader can find more information in my lecture notes at https://sebastianraschka.com/pdf/ lecture-notes/stat479ss19/L05_gradient-descent_ slides.pdf or http://rasbt.github.io/mlxtend/user_ guide/classifier/SoftmaxRegression/.

Another way to use logistic regression in multiclass settings is via the OvR technique, which we discussed previously.

To explain the idea behind logistic regression as a probabilistic model for binary classification, let's first introduce the odds: the odds in favor of a particular event. 𝑝𝑝 The odds can be written as where p stands for the probability of the positive (1 − 𝑝𝑝)

event. The term "positive event" does not necessarily mean "good," but refers to the event that we want to predict, for example, the probability that a patient has a certain disease; we can think of the positive event as class label y = 1. We can then further define the logit function, which is simply the logarithm of the odds (log-odds): 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙(𝑝𝑝) = log

𝑝𝑝 (1 − 𝑝𝑝)

Note that log refers to the natural logarithm, as it is the common convention in computer science. The logit function takes input values in the range 0 to 1 and transforms them to values over the entire real-number range, which we can use to express a linear relationship between feature values and the log-odds: 𝑚𝑚

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙(𝑝𝑝(𝑦𝑦 = 1|𝒙𝒙)) = 𝑤𝑤0 𝑥𝑥0 + 𝑤𝑤1 𝑥𝑥1 + ⋯ +𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚 = ∑ 𝑤𝑤𝑖𝑖 𝑥𝑥𝑖𝑖 = 𝒘𝒘𝑇𝑇 𝒙𝒙 𝑖𝑖=0

Here, 𝑝𝑝(𝑦𝑦 = 1|𝒙𝒙) is the conditional probability that a particular example belongs to class 1 given its features, x. Now, we are actually interested in predicting the probability that a certain example belongs to a particular class, which is the inverse form of the logit function.

[ 61 ]

A Tour of Machine Learning Classifiers Using scikit-learn

It is also called the logistic sigmoid function, which is sometimes simply abbreviated to sigmoid function due to its characteristic S-shape:

𝜙𝜙(𝑧𝑧) =

1 1 + 𝑒𝑒 −𝑧𝑧

Here, z is the net input, the linear combination of weights, and the inputs (that is, the features associated with the training examples):

𝑧𝑧 = 𝒘𝒘𝑇𝑇 𝒙𝒙 = 𝑤𝑤0 𝑥𝑥0 + 𝑤𝑤1 𝑥𝑥1 + ⋯ + 𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚

Note that similar to the convention we used in Chapter 2, Training Simple Machine Learning Algorithms for Classification, 𝑤𝑤0 refers to the bias unit and is an additional input value that we provide to 𝑥𝑥0 , which is set equal to 1.

Now, let's simply plot the sigmoid function for some values in the range –7 to 7 to see how it looks: >>> >>> >>> ... >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>>

import matplotlib.pyplot as plt import numpy as np def sigmoid(z): return 1.0 / (1.0 + np.exp(-z)) z = np.arange(-7, 7, 0.1) phi_z = sigmoid(z) plt.plot(z, phi_z) plt.axvline(0.0, color='k') plt.ylim(-0.1, 1.1) plt.xlabel('z') plt.ylabel('$\phi (z)$') # y axis ticks and gridline plt.yticks([0.0, 0.5, 1.0]) ax = plt.gca() ax.yaxis.grid(True) plt.tight_layout() plt.show()

[ 62 ]

Chapter 3

As a result of executing the previous code example, we should now see the S-shaped (sigmoidal) curve:

We can see that 𝜙𝜙(𝑧𝑧) approaches 1 if z goes toward infinity ( 𝑧𝑧 → ∞ ) since 𝑒𝑒 −𝑧𝑧 becomes very small for large values of z. Similarly, 𝜙𝜙(𝑧𝑧) goes toward 0 for 𝑧𝑧 → −∞ as a result of an increasingly large denominator. Thus, we can conclude that this sigmoid function takes real-number values as input and transforms them into values in the range [0, 1] with an intercept at 𝜙𝜙(𝑧𝑧) = 0.5 .

To build some understanding of the logistic regression model, we can relate it to Chapter 2. In Adaline, we used the identity function, 𝜙𝜙(𝑧𝑧) = 𝑧𝑧 , as the activation function. In logistic regression, this activation function simply becomes the sigmoid function that we defined earlier.

[ 63 ]

A Tour of Machine Learning Classifiers Using scikit-learn

The difference between Adaline and logistic regression is illustrated in the following figure:

The output of the sigmoid function is then interpreted as the probability of a particular example belonging to class 1, 𝜙𝜙(𝑧𝑧) = 𝑃𝑃(𝑦𝑦 = 1|𝒙𝒙; 𝒘𝒘) , given its features, x, parameterized by the weights, w. For example, if we compute 𝜙𝜙(𝑧𝑧) = 0.8 for a particular flower example, it means that the chance that this example is an Irisversicolor flower is 80 percent. Therefore, the probability that this flower is an Iris-setosa flower can be calculated as 𝑃𝑃(𝑦𝑦 = 0|𝒙𝒙; 𝒘𝒘) = 1 − 𝑃𝑃(𝑦𝑦 = 1|𝒙𝒙; 𝒘𝒘) = 0.2 or 20 percent. The predicted probability can then simply be converted into a binary outcome via a threshold function:

𝑦𝑦̂ = {

1 if 𝜙𝜙(𝑧𝑧) ≥ 0.5 0 otherwise

If we look at the preceding plot of the sigmoid function, this is equivalent to the following:

𝑦𝑦̂ = {

1 if 𝑧𝑧 ≥ 0.0 0 otherwise [ 64 ]

Chapter 3

In fact, there are many applications where we are not only interested in the predicted class labels, but where the estimation of the class-membership probability is particularly useful (the output of the sigmoid function prior to applying the threshold function). Logistic regression is used in weather forecasting, for example, not only to predict whether it will rain on a particular day but also to report the chance of rain. Similarly, logistic regression can be used to predict the chance that a patient has a particular disease given certain symptoms, which is why logistic regression enjoys great popularity in the field of medicine.

Learning the weights of the logistic cost function

You have learned how we can use the logistic regression model to predict probabilities and class labels; now, let's briefly talk about how we fit the parameters of the model, for instance the weights, w. In the previous chapter, we defined the sum-squared-error cost function as follows:

𝐽𝐽(𝒘𝒘) = ∑ 𝑖𝑖

1 2 (𝜙𝜙(𝑧𝑧 (𝑖𝑖) ) − 𝑦𝑦 (𝑖𝑖) ) 2

We minimized this function in order to learn the weights, w, for our Adaline classification model. To explain how we can derive the cost function for logistic regression, let's first define the likelihood, L, that we want to maximize when we build a logistic regression model, assuming that the individual examples in our dataset are independent of one another. The formula is as follows: 𝑛𝑛

𝐿𝐿(𝒘𝒘) = 𝑃𝑃(𝒚𝒚 | 𝒙𝒙; 𝒘𝒘) = ∏ 𝑃𝑃(𝑦𝑦 𝑖𝑖=1

(𝑖𝑖)

𝑛𝑛

(𝑖𝑖)

| 𝑥𝑥 ; 𝒘𝒘) = ∏ (𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑖𝑖=1

𝑦𝑦 (𝑖𝑖)

1−𝑦𝑦 (𝑖𝑖)

(1 − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) ))

In practice, it is easier to maximize the (natural) log of this equation, which is called the log-likelihood function: 𝑛𝑛

𝑙𝑙(𝒘𝒘) = log 𝐿𝐿(𝒘𝒘) = ∑ [𝑦𝑦 (𝑖𝑖) log (𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) + (1 − 𝑦𝑦 (𝑖𝑖) ) log (1 − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) ))] 𝑖𝑖=1

Firstly, applying the log function reduces the potential for numerical underflow, which can occur if the likelihoods are very small. Secondly, we can convert the product of factors into a summation of factors, which makes it easier to obtain the derivative of this function via the addition trick, as you may remember from calculus. [ 65 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Now, we could use an optimization algorithm such as gradient ascent to maximize this log-likelihood function. Alternatively, let's rewrite the log-likelihood as a cost function, J, that can be minimized using gradient descent as in Chapter 2, Training Simple Machine Learning Algorithms for Classification: 𝑛𝑛

J(𝐰𝐰) = ∑ [−𝑦𝑦 (𝑖𝑖) log (𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) − (1 − 𝑦𝑦 (𝑖𝑖) ) log (1 − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) ))] 𝑖𝑖=1

To get a better grasp of this cost function, let's take a look at the cost that we calculate for one single training example: 𝐽𝐽(𝜙𝜙(𝑧𝑧), 𝑦𝑦; 𝒘𝒘) = −𝑦𝑦 log(𝜙𝜙(𝑧𝑧)) − (1 − 𝑦𝑦)log(1 − 𝜙𝜙(𝑧𝑧))

Looking at the equation, we can see that the first term becomes zero if y = 0, and the second term becomes zero if y = 1:

− log(𝜙𝜙(𝑧𝑧)) if 𝑦𝑦 = 1 𝐽𝐽(𝜙𝜙(𝑧𝑧), 𝑦𝑦; 𝒘𝒘) = { − log(1 − 𝜙𝜙(𝑧𝑧)) if 𝑦𝑦 = 0

Let's write a short code snippet to create a plot that illustrates the cost of classifying a single training example for different values of 𝜙𝜙(𝑧𝑧) : >>> ... >>> ... >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>>

def cost_1(z): return - np.log(sigmoid(z)) def cost_0(z): return - np.log(1 - sigmoid(z)) z = np.arange(-10, 10, 0.1) phi_z = sigmoid(z) c1 = [cost_1(x) for x in z] plt.plot(phi_z, c1, label='J(w) if y=1') c0 = [cost_0(x) for x in z] plt.plot(phi_z, c0, linestyle='--', label='J(w) if y=0') plt.ylim(0.0, 5.1) plt.xlim([0, 1]) plt.xlabel('$\phi$(z)') plt.ylabel('J(w)') plt.legend(loc='best') plt.tight_layout() plt.show()

[ 66 ]

Chapter 3

The resulting plot shows the sigmoid activation on the x-axis in the range 0 to 1 (the inputs to the sigmoid function were z values in the range –10 to 10) and the associated logistic cost on the y-axis:

We can see that the cost approaches 0 (continuous line) if we correctly predict that an example belongs to class 1. Similarly, we can see on the y-axis that the cost also approaches 0 if we correctly predict y = 0 (dashed line). However, if the prediction is wrong, the cost goes toward infinity. The main point is that we penalize wrong predictions with an increasingly larger cost.

Converting an Adaline implementation into an algorithm for logistic regression If we were to implement logistic regression ourselves, we could simply substitute the cost function, J, in our Adaline implementation from Chapter 2, Training Simple Machine Learning Algorithms for Classification, with the new cost function:

𝐽𝐽(𝒘𝒘) = − ∑ 𝑦𝑦 (𝑖𝑖) log (𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) + (1 − 𝑦𝑦 (𝑖𝑖) ) log (1 − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑖𝑖

[ 67 ]

A Tour of Machine Learning Classifiers Using scikit-learn

We use this to compute the cost of classifying all training examples per epoch. Also, we need to swap the linear activation function with the sigmoid activation and change the threshold function to return class labels 0 and 1, instead of –1 and 1. If we make those changes to the Adaline code, we will end up with a working logistic regression implementation, as shown here: class LogisticRegressionGD(object): """Logistic Regression Classifier using gradient descent. Parameters -----------eta : float Learning rate (between 0.0 and 1.0) n_iter : int Passes over the training dataset. random_state : int Random number generator seed for random weight initialization.

Attributes ----------w_ : 1d-array Weights after fitting. cost_ : list Logistic cost function value in each epoch. """ def __init__(self, eta=0.05, n_iter=100, random_state=1): self.eta = eta self.n_iter = n_iter self.random_state = random_state def fit(self, X, y): """ Fit training data. Parameters ---------X : {array-like}, shape = [n_examples, n_features] Training vectors, where n_examples is the number of examples and n_features is the number of features. y : array-like, shape = [n_examples] Target values.

[ 68 ]

Chapter 3 Returns ------self : object """ rgen = np.random.RandomState(self.random_state) self.w_ = rgen.normal(loc=0.0, scale=0.01, size=1 + X.shape[1]) self.cost_ = [] for i in range(self.n_iter): net_input = self.net_input(X) output = self.activation(net_input) errors = (y - output) self.w_[1:] += self.eta * X.T.dot(errors) self.w_[0] += self.eta * errors.sum() # note that we compute the logistic 'cost' now # instead of the sum of squared errors cost cost = (-y.dot(np.log(output)) ((1 - y).dot(np.log(1 - output)))) self.cost_.append(cost) return self def net_input(self, X): """Calculate net input""" return np.dot(X, self.w_[1:]) + self.w_[0] def activation(self, z): """Compute logistic sigmoid activation""" return 1. / (1. + np.exp(-np.clip(z, -250, 250))) def predict(self, X): """Return class label after unit step""" return np.where(self.net_input(X) >= 0.0, 1, 0) # equivalent to: # return np.where(self.activation(self.net_input(X)) # >= 0.5, 1, 0)

When we fit a logistic regression model, we have to keep in mind that it only works for binary classification tasks.

[ 69 ]

A Tour of Machine Learning Classifiers Using scikit-learn

So, let's consider only Iris-setosa and Iris-versicolor flowers (classes 0 and 1) and check that our implementation of logistic regression works: >>> >>> >>> ... ... >>> ... >>> ... ... >>> >>> >>> >>> >>>

X_train_01_subset = X_train[(y_train == 0) | (y_train == 1)] y_train_01_subset = y_train[(y_train == 0) | (y_train == 1)] lrgd = LogisticRegressionGD(eta=0.05, n_iter=1000, random_state=1) lrgd.fit(X_train_01_subset, y_train_01_subset) plot_decision_regions(X=X_train_01_subset, y=y_train_01_subset, classifier=lrgd) plt.xlabel('petal length [standardized]') plt.ylabel('petal width [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

The resulting decision region plot looks as follows:

[ 70 ]

Chapter 3

The gradient descent learning algorithm for logistic regression Using calculus, we can show that the weight update in logistic regression via gradient descent is equal to the equation that we used in Adaline in Chapter 2, Training Simple Machine Learning Algorithms for Classification. However, please note that the following derivation of the gradient descent learning rule is intended for readers who are interested in the mathematical concepts behind the gradient descent learning rule for logistic regression. It is not essential for following the rest of this chapter. Let's start by calculating the partial derivative of the loglikelihood function with respect to the jth weight: 1 1 𝜕𝜕 𝜕𝜕 𝑙𝑙(𝒘𝒘) = (𝑦𝑦 − (1 − 𝑦𝑦) ) 𝜙𝜙(𝑧𝑧) 𝜙𝜙(𝑧𝑧) 1 − 𝜙𝜙(𝑧𝑧) 𝜕𝜕𝑤𝑤𝑗𝑗 𝜕𝜕𝑤𝑤𝑗𝑗

Before we continue, let's also calculate the partial derivative of the sigmoid function: 𝜕𝜕 1 1 1 1 𝜕𝜕 𝜙𝜙(𝑧𝑧) = = 𝑒𝑒 −𝑧𝑧 = (1 − ) = 𝜙𝜙(𝑧𝑧)(1 − 𝜙𝜙(𝑧𝑧)) (1 + 𝑒𝑒 −𝑧𝑧 )2 𝜕𝜕𝜕𝜕 1 + 𝑒𝑒 −𝑧𝑧 1 + 𝑒𝑒 −𝑧𝑧 1 + 𝑒𝑒 −𝑧𝑧 𝜕𝜕𝜕𝜕

Now, we can resubstitute

𝜕𝜕 𝜙𝜙(𝑧𝑧) = 𝜙𝜙(𝑧𝑧)(1 − 𝜙𝜙(𝑧𝑧)) 𝜕𝜕𝜕𝜕

in our first

equation to obtain the following: (𝑦𝑦

1 1 𝜕𝜕 1 1 𝜕𝜕 − (1 − 𝑦𝑦) ) 𝜙𝜙(𝑧𝑧) = (𝑦𝑦 − (1 − 𝑦𝑦) ) 𝜙𝜙(𝑧𝑧)(1 − 𝜙𝜙(𝑧𝑧)) 𝑧𝑧 𝜙𝜙(𝑧𝑧) 1 − 𝜙𝜙(𝑧𝑧) 𝜕𝜕𝑤𝑤𝑗𝑗 𝜙𝜙(𝑧𝑧) 1 − 𝜙𝜙(𝑧𝑧) 𝜕𝜕𝑤𝑤𝑗𝑗 = (𝑦𝑦(1 − 𝜙𝜙(𝑧𝑧)) − (1 − 𝑦𝑦)𝜙𝜙(𝑧𝑧)) 𝑥𝑥𝑗𝑗 = (𝑦𝑦 − 𝜙𝜙(𝑧𝑧))𝑥𝑥𝑗𝑗

Remember that the goal is to find the weights that maximize the log-likelihood so that we perform the update for each weight as follows: 𝑛𝑛

𝑤𝑤𝑗𝑗 ≔ 𝑤𝑤𝑗𝑗 + 𝜂𝜂 ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑥𝑥𝑗𝑗 (𝑖𝑖) 𝑖𝑖=1

[ 71 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Since we update all weights simultaneously, we can write the general update rule as follows:

We define Δ𝒘𝒘 as follows:

𝒘𝒘 ≔ 𝒘𝒘 + Δ𝒘𝒘 Δ𝒘𝒘 = 𝜂𝜂∇𝑙𝑙(𝒘𝒘)

Since maximizing the log-likelihood is equal to minimizing the cost function, J, that we defined earlier, we can write the gradient descent update rule as follows: Δ𝑤𝑤𝑗𝑗 = −𝜂𝜂

𝑛𝑛

𝜕𝜕𝜕𝜕 = 𝜂𝜂 ∑ (𝑦𝑦 (𝑖𝑖) − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) 𝑥𝑥𝑗𝑗 (𝑖𝑖) 𝜕𝜕𝑤𝑤𝑗𝑗 𝑖𝑖=1

𝒘𝒘 ≔ 𝒘𝒘 + Δ𝒘𝒘, Δ𝒘𝒘 = −𝜂𝜂∇𝐽𝐽(𝒘𝒘)

This is equal to the gradient descent rule for Adaline in Chapter 2, Training Simple Machine Learning Algorithms for Classification.

Training a logistic regression model with scikit-learn

We just went through useful coding and math exercises in the previous subsection, which helped to illustrate the conceptual differences between Adaline and logistic regression. Now, let's learn how to use scikit-learn's more optimized implementation of logistic regression, which also supports multiclass settings off the shelf. Note that in recent versions of scikit-learn, the technique used for multiclass classification, multinomial, or OvR, is chosen automatically. In the following code example, we will use the sklearn.linear_model.LogisticRegression class as well as the familiar fit method to train the model on all three classes in the standardized flower training dataset. Also, we set multi_class='ovr' for illustration purposes. As an exercise for the reader, you may want to compare the results with multi_ class='multinomial'. Note that the multinomial setting is usually recommended in practice for mutually exclusive classes, such as those found in the Iris dataset. Here, "mutually exclusive" means that each training example can only belong to a single class (in contrast to multilabel classification, where a training example can be a member of multiple classes).

[ 72 ]

Chapter 3

Now, let's have a look at the code example: >>> >>> ... >>> >>> ... ... ... >>> >>> >>> >>> >>>

from sklearn.linear_model import LogisticRegression lr = LogisticRegression(C=100.0, random_state=1, solver='lbfgs', multi_class='ovr') lr.fit(X_train_std, y_train) plot_decision_regions(X_combined_std, y_combined, classifier=lr, test_idx=range(105, 150)) plt.xlabel('petal length [standardized]') plt.ylabel('petal width [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

After fitting the model on the training data, we plotted the decision regions, training examples, and test examples, as shown in the following plot:

[ 73 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Note that there exist many different optimization algorithms for solving optimization problems. For minimizing convex loss functions, such as the logistic regression loss, it is recommended to use more advanced approaches than regular stochastic gradient descent (SGD). In fact, scikit-learn implements a whole range of such optimization algorithms, which can be specified via the solver parameter, namely, 'newton-cg', 'lbfgs', 'liblinear', 'sag', and 'saga'. While the logistic regression loss is convex, most optimization algorithms should converge to the global loss minimum with ease. However, there are certain advantages of using one algorithm over the other. For instance, in the current version (v 0.21), scikit-learn uses 'liblinear' as a default, which cannot handle the multinomial loss and is limited to the OvR scheme for multi-class classification. However, in future versions of scikit-learn (that is, v 0.22), the default solver will be changed to 'lbfgs', which stands for the limited-memory Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm (https://en.wikipedia.org/wiki/Limited-memory_BFGS) and is more flexible in this regard. To adopt this new default choice, we will specify solver='lbfgs' explicitly when using logistic regression throughout this book.

Looking at the preceding code that we used to train the LogisticRegression model, you might now be wondering, "What is this mysterious parameter C?" We will discuss this parameter in the next subsection, where we will introduce the concepts of overfitting and regularization. However, before we move on to those topics, let's finish our discussion of class-membership probabilities. The probability that training examples belong to a certain class can be computed using the predict_proba method. For example, we can predict the probabilities of the first three examples in the test dataset as follows: >>> lr.predict_proba(X_test_std[:3, :])

This code snippet returns the following array: array([[3.81527885e-09, 1.44792866e-01, 8.55207131e-01], [8.34020679e-01, 1.65979321e-01, 3.25737138e-13], [8.48831425e-01, 1.51168575e-01, 2.62277619e-14]])

[ 74 ]

Chapter 3

The first row corresponds to the class-membership probabilities of the first flower, the second row corresponds to the class-membership probabilities of the second flower, and so forth. Notice that the columns all sum up to one, as expected. (You can confirm this by executing lr.predict_proba(X_test_std[:3, :]).sum(axis=1).) The highest value in the first row is approximately 0.85, which means that the first example belongs to class three (Iris-virginica) with a predicted probability of 85 percent. So, as you may have already noticed, we can get the predicted class labels by identifying the largest column in each row, for example, using NumPy's argmax function: >>> lr.predict_proba(X_test_std[:3, :]).argmax(axis=1)

The returned class indices are shown here (they correspond to Iris-virginica, Iris-setosa, and Iris-setosa): array([2, 0, 0])

In the preceding code example, we computed the conditional probabilities and converted these into class labels manually by using NumPy's argmax function. In practice, the more convenient way of obtaining class labels when using scikit-learn is to call the predict method directly: >>> lr.predict(X_test_std[:3, :]) array([2, 0, 0])

Lastly, a word of caution if you want to predict the class label of a single flower example: scikit-learn expects a two-dimensional array as data input; thus, we have to convert a single row slice into such a format first. One way to convert a single row entry into a two-dimensional data array is to use NumPy's reshape method to add a new dimension, as demonstrated here: >>> lr.predict(X_test_std[0, :].reshape(1, -1)) array([2])

Tackling overfitting via regularization

Overfitting is a common problem in machine learning, where a model performs well on training data but does not generalize well to unseen data (test data). If a model suffers from overfitting, we also say that the model has a high variance, which can be caused by having too many parameters, leading to a model that is too complex given the underlying data. Similarly, our model can also suffer from underfitting (high bias), which means that our model is not complex enough to capture the pattern in the training data well and therefore also suffers from low performance on unseen data. [ 75 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Although we have only encountered linear models for classification so far, the problems of overfitting and underfitting can be best illustrated by comparing a linear decision boundary to more complex, nonlinear decision boundaries, as shown in the following figure:

The bias-variance tradeoff Often, researchers use the terms "bias" and "variance" or "biasvariance tradeoff" to describe the performance of a model—that is, you may stumble upon talks, books, or articles where people say that a model has a "high variance" or "high bias." So, what does that mean? In general, we might say that "high variance" is proportional to overfitting and "high bias" is proportional to underfitting. In the context of machine learning models, variance measures the consistency (or variability) of the model prediction for classifying a particular example if we retrain the model multiple times, for example, on different subsets of the training dataset. We can say that the model is sensitive to the randomness in the training data. In contrast, bias measures how far off the predictions are from the correct values in general if we rebuild the model multiple times on different training datasets; bias is the measure of the systematic error that is not due to randomness. If you are interested in the technical specification and derivation of the "bias" and "variance" terms, I've written about it in my lecture notes here: https://sebastianraschka.com/pdf/ lecture-notes/stat479fs18/08_eval-intro_notes.pdf.

One way of finding a good bias-variance tradeoff is to tune the complexity of the model via regularization. Regularization is a very useful method for handling collinearity (high correlation among features), filtering out noise from data, and eventually preventing overfitting. [ 76 ]

Chapter 3

The concept behind regularization is to introduce additional information (bias) to penalize extreme parameter (weight) values. The most common form of regularization is so-called L2 regularization (sometimes also called L2 shrinkage or weight decay), which can be written as follows: 𝑚𝑚

𝜆𝜆 𝜆𝜆 ‖𝒘𝒘‖2 = ∑ 𝑤𝑤𝑗𝑗2 2 2 𝑗𝑗=1

Here, 𝜆𝜆 is the so-called regularization parameter.

Regularization and feature normalization Regularization is another reason why feature scaling such as standardization is important. For regularization to work properly, we need to ensure that all our features are on comparable scales.

The cost function for logistic regression can be regularized by adding a simple regularization term, which will shrink the weights during model training: 𝑛𝑛

𝐽𝐽(𝒘𝒘) = ∑ [−𝑦𝑦 (𝑖𝑖) log (𝜙𝜙(𝑧𝑧 (𝑖𝑖) )) − (1 − 𝑦𝑦 (𝑖𝑖) ) log (1 − 𝜙𝜙(𝑧𝑧 (𝑖𝑖) ))] + 𝑖𝑖=1

𝜆𝜆 ‖𝒘𝒘‖2 2

Via the regularization parameter, 𝜆𝜆 , we can then control how well we fit the training data, while keeping the weights small. By increasing the value of 𝜆𝜆, we increase the regularization strength.

The parameter, C, that is implemented for the LogisticRegression class in scikitlearn comes from a convention in support vector machines, which will be the topic of the next section. The term C is directly related to the regularization parameter, 𝜆𝜆, which is its inverse. Consequently, decreasing the value of the inverse regularization parameter, C, means that we are increasing the regularization strength, which we can visualize by plotting the L2 regularization path for the two weight coefficients: >>> weights, params = [], [] >>> for c in np.arange(-5, 5): ... lr = LogisticRegression(C=10.**c, random_state=1, ... solver='lbfgs', multi_class='ovr') ... lr.fit(X_train_std, y_train) ... weights.append(lr.coef_[1]) ... params.append(10.**c) >>> weights = np.array(weights) [ 77 ]

A Tour of Machine Learning Classifiers Using scikit-learn >>> ... >>> ... >>> >>> >>> >>> >>>

plt.plot(params, weights[:, 0], label='petal length') plt.plot(params, weights[:, 1], linestyle='--', label='petal width') plt.ylabel('weight coefficient') plt.xlabel('C') plt.legend(loc='upper left') plt.xscale('log') plt.show()

By executing the preceding code, we fitted 10 logistic regression models with different values for the inverse-regularization parameter, C. For the purposes of illustration, we only collected the weight coefficients of class 1 (here, the second class in the dataset: Iris-versicolor) versus all classifiers—remember that we are using the OvR technique for multiclass classification. As we can see in the resulting plot, the weight coefficients shrink if we decrease parameter C, that is, if we increase the regularization strength:

An additional resource on logistic regression Since an in-depth coverage of the individual classification algorithms exceeds the scope of this book, Logistic Regression: From Introductory to Advanced Concepts and Applications, Dr. Scott Menard, Sage Publications, 2009, is recommended to readers who want to learn more about logistic regression.

[ 78 ]

Chapter 3

Maximum margin classification with support vector machines

Another powerful and widely used learning algorithm is the support vector machine (SVM), which can be considered an extension of the perceptron. Using the perceptron algorithm, we minimized misclassification errors. However, in SVMs our optimization objective is to maximize the margin. The margin is defined as the distance between the separating hyperplane (decision boundary) and the training examples that are closest to this hyperplane, which are the so-called support vectors. This is illustrated in the following figure:

Maximum margin intuition

The rationale behind having decision boundaries with large margins is that they tend to have a lower generalization error, whereas models with small margins are more prone to overfitting. To get an idea of the margin maximization, let's take a closer look at those positive and negative hyperplanes that are parallel to the decision boundary, which can be expressed as follows:

𝑤𝑤𝟎𝟎 + 𝒘𝒘𝑇𝑇 𝒙𝒙𝒑𝒑𝒑𝒑𝒑𝒑 = 1

(1)

𝑤𝑤𝟎𝟎 + 𝒘𝒘𝑇𝑇 𝒙𝒙𝒏𝒏𝒏𝒏𝒏𝒏 = −1

(2)

[ 79 ]

A Tour of Machine Learning Classifiers Using scikit-learn

If we subtract those two linear equations (1) and (2) from each other, we get:

⇒ 𝒘𝒘𝑇𝑇 (𝒙𝒙𝒑𝒑𝒑𝒑𝒑𝒑 − 𝒙𝒙𝒏𝒏𝒏𝒏𝒏𝒏 ) = 2

We can normalize this equation by the length of the vector w, which is defined as follows: 𝑚𝑚

‖𝒘𝒘‖ = √∑ 𝑤𝑤𝑗𝑗2 𝑗𝑗=1

So, we arrive at the following equation:

𝒘𝒘𝑻𝑻 (𝒙𝒙𝒑𝒑𝒑𝒑𝒑𝒑 − 𝒙𝒙𝒏𝒏𝒏𝒏𝒏𝒏 ) 2 = ‖𝒘𝒘‖ ‖𝒘𝒘‖

The left side of the preceding equation can then be interpreted as the distance between the positive and negative hyperplane, which is the so-called margin that we want to maximize. Now, the objective function of the SVM becomes the maximization of this margin by maximizing

2 under the constraint that the examples are classified correctly, ‖𝒘𝒘‖

which can be written as:

𝑤𝑤𝟎𝟎 + 𝒘𝒘𝑇𝑇 𝒙𝒙(𝑖𝑖) ≥ 1 𝑖𝑖𝑖𝑖 𝑦𝑦 (𝑖𝑖) = 1

𝑤𝑤𝟎𝟎 + 𝒘𝒘𝑇𝑇 𝒙𝒙(𝑖𝑖) ≤ −1 𝑖𝑖𝑖𝑖 𝑦𝑦 (𝑖𝑖) = −1 for 𝑖𝑖 = 1. . . 𝑁𝑁

Here, N is the number of examples in our dataset.

[ 80 ]

Chapter 3

These two equations basically say that all negative-class examples should fall on one side of the negative hyperplane, whereas all the positive-class examples should fall behind the positive hyperplane, which can also be written more compactly as follows:

𝑦𝑦 (𝑖𝑖) (𝑤𝑤𝟎𝟎 + 𝒘𝒘𝑇𝑇 𝒙𝒙(𝑖𝑖) ) ≥ 1

∀𝑖𝑖

1 2

In practice, though, it is easier to minimize the reciprocal term, ‖𝒘𝒘‖2 , which can be

solved by quadratic programming. However, a detailed discussion about quadratic programming is beyond the scope of this book. You can learn more about SVMs in The Nature of Statistical Learning Theory, Springer Science+Business Media, Vladimir Vapnik, 2000, or read Chris J.C. Burges' excellent explanation in A Tutorial on Support Vector Machines for Pattern Recognition (Data Mining and Knowledge Discovery, 2(2): 121-167, 1998).

Dealing with a nonlinearly separable case using slack variables

Although we don't want to dive much deeper into the more involved mathematical concepts behind the maximum-margin classification, let's briefly mention the slack variable, 𝜉𝜉 , which was introduced by Vladimir Vapnik in 1995 and led to the socalled soft-margin classification. The motivation for introducing the slack variable was that the linear constraints need to be relaxed for nonlinearly separable data to allow the convergence of the optimization in the presence of misclassifications, under appropriate cost penalization. The positive-valued slack variable is simply added to the linear constraints:

𝑤𝑤𝟎𝟎 + 𝒘𝒘𝑇𝑇 𝒙𝒙(𝑖𝑖) ≥ 1 − 𝜉𝜉 (𝑖𝑖) 𝑖𝑖𝑖𝑖 𝑦𝑦 (𝑖𝑖) = 1 𝑤𝑤𝟎𝟎 + 𝒘𝒘𝑇𝑇 𝒙𝒙(𝑖𝑖) ≤ −1 + 𝜉𝜉 (𝑖𝑖) 𝑖𝑖𝑖𝑖 𝑦𝑦 (𝑖𝑖) = −1 for 𝑖𝑖 = 1. . . 𝑁𝑁

[ 81 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Here, N is the number of examples in our dataset. So, the new objective to be minimized (subject to the constraints) becomes: 1 ‖𝒘𝒘‖2 + 𝐶𝐶 (∑ 𝜉𝜉 (𝑖𝑖) ) 2 𝑖𝑖

Via the variable, C, we can then control the penalty for misclassification. Large values of C correspond to large error penalties, whereas we are less strict about misclassification errors if we choose smaller values for C. We can then use the C parameter to control the width of the margin and therefore tune the bias-variance tradeoff, as illustrated in the following figure:

This concept is related to regularization, which we discussed in the previous section in the context of regularized regression, where decreasing the value of C increases the bias and lowers the variance of the model. Now that we have learned the basic concepts behind a linear SVM, let's train an SVM model to classify the different flowers in our Iris dataset: >>> >>> >>> >>> ... ... ... >>> >>>

from sklearn.svm import SVC svm = SVC(kernel='linear', C=1.0, random_state=1) svm.fit(X_train_std, y_train) plot_decision_regions(X_combined_std, y_combined, classifier=svm, test_idx=range(105, 150)) plt.xlabel('petal length [standardized]') plt.ylabel('petal width [standardized]')

[ 82 ]

Chapter 3 >>> plt.legend(loc='upper left') >>> plt.tight_layout() >>> plt.show()

The three decision regions of the SVM, visualized after training the classifier on the Iris dataset by executing the preceding code example, are shown in the following plot:

Logistic regression versus SVMs In practical classification tasks, linear logistic regression and linear SVMs often yield very similar results. Logistic regression tries to maximize the conditional likelihoods of the training data, which makes it more prone to outliers than SVMs, which mostly care about the points that are closest to the decision boundary (support vectors). On the other hand, logistic regression has the advantage that it is a simpler model and can be implemented more easily. Furthermore, logistic regression models can be easily updated, which is attractive when working with streaming data.

Alternative implementations in scikit-learn

The scikit-learn library's LogisticRegression class, which we used in the previous sections, makes use of the LIBLINEAR library, which is a highly optimized C/C++ library developed at the National Taiwan University (http://www.csie.ntu.edu. tw/~cjlin/liblinear/). [ 83 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Similarly, the SVC class that we used to train an SVM makes use of LIBSVM, which is an equivalent C/C++ library specialized for SVMs (http://www.csie.ntu.edu. tw/~cjlin/libsvm/). The advantage of using LIBLINEAR and LIBSVM over native Python implementations is that they allow the extremely quick training of large amounts of linear classifiers. However, sometimes our datasets are too large to fit into computer memory. Thus, scikit-learn also offers alternative implementations via the SGDClassifier class, which also supports online learning via the partial_fit method. The concept behind the SGDClassifier class is similar to the stochastic gradient algorithm that we implemented in Chapter 2, Training Simple Machine Learning Algorithms for Classification, for Adaline. We could initialize the SGD version of the perceptron, logistic regression, and an SVM with default parameters, as follows: >>> >>> >>> >>>

from sklearn.linear_model import SGDClassifier ppn = SGDClassifier(loss='perceptron') lr = SGDClassifier(loss='log') svm = SGDClassifier(loss='hinge')

Solving nonlinear problems using a kernel SVM

Another reason why SVMs enjoy high popularity among machine learning practitioners is that they can be easily kernelized to solve nonlinear classification problems. Before we discuss the main concept behind the so-called kernel SVM, the most common variant of SVMs, let's first create a synthetic dataset to see what such a nonlinear classification problem may look like.

Kernel methods for linearly inseparable data Using the following code, we will create a simple dataset that has the form of an XOR gate using the logical_or function from NumPy, where 100 examples will be assigned the class label 1, and 100 examples will be assigned the class label -1: >>> >>> >>> >>> >>> ... >>> >>>

import matplotlib.pyplot as plt import numpy as np np.random.seed(1) X_xor = np.random.randn(200, 2) y_xor = np.logical_xor(X_xor[:, 0] > 0, X_xor[:, 1] > 0) y_xor = np.where(y_xor, 1, -1) plt.scatter(X_xor[y_xor == 1, 0], [ 84 ]

Chapter 3 ... ... ... >>> ... ... ... ... >>> >>> >>> >>> >>>

X_xor[y_xor == 1, 1], c='b', marker='x', label='1') plt.scatter(X_xor[y_xor == -1, 0], X_xor[y_xor == -1, 1], c='r', marker='s', label='-1') plt.xlim([-3, 3]) plt.ylim([-3, 3]) plt.legend(loc='best') plt.tight_layout() plt.show()

After executing the code, we will have an XOR dataset with random noise, as shown in the following plot:

Obviously, we would not be able to separate the examples from the positive and negative class very well using a linear hyperplane as a decision boundary via the linear logistic regression or linear SVM model that we discussed in earlier sections. The basic idea behind kernel methods to deal with such linearly inseparable data is to create nonlinear combinations of the original features to project them onto a higher-dimensional space via a mapping function, 𝜙𝜙 , where the data becomes linearly separable. As shown in the following plot, we can transform a twodimensional dataset into a new three-dimensional feature space, where the classes become separable via the following projection: 𝜙𝜙(𝑥𝑥1 , 𝑥𝑥2 ) = (𝑧𝑧1 , 𝑧𝑧2 , 𝑧𝑧3 ) = (𝑥𝑥1 , 𝑥𝑥2 , 𝑥𝑥12 + 𝑥𝑥22 ) [ 85 ]

A Tour of Machine Learning Classifiers Using scikit-learn

This allows us to separate the two classes shown in the plot via a linear hyperplane that becomes a nonlinear decision boundary if we project it back onto the original feature space:

Using the kernel trick to find separating hyperplanes in a high-dimensional space

To solve a nonlinear problem using an SVM, we would transform the training data into a higher-dimensional feature space via a mapping function, 𝜙𝜙 , and train a linear SVM model to classify the data in this new feature space. Then, we could use the same mapping function, 𝜙𝜙 , to transform new, unseen data to classify it using the linear SVM model. However, one problem with this mapping approach is that the construction of the new features is computationally very expensive, especially if we are dealing with high-dimensional data. This is where the so-called kernel trick comes into play.

[ 86 ]

Chapter 3

Although we did not go into much detail about how to solve the quadratic programming task to train an SVM, in practice, we just need to replace the dot 𝑇𝑇 product 𝒙𝒙(𝑖𝑖) 𝑇𝑇 𝒙𝒙(𝑗𝑗) by 𝜙𝜙(𝒙𝒙(𝑖𝑖)) 𝜙𝜙(𝒙𝒙(𝑗𝑗) ) . In order to save the expensive step of calculating this dot product between two points explicitly, we define a so-called kernel function: 𝑇𝑇

𝜅𝜅(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = 𝜙𝜙(𝒙𝒙(𝑖𝑖) ) 𝜙𝜙(𝒙𝒙(𝑗𝑗) )

One of the most widely used kernels is the radial basis function (RBF) kernel, which can simply be called the Gaussian kernel: 2

‖𝒙𝒙(𝑖𝑖) −𝒙𝒙(𝑗𝑗) ‖ 𝜅𝜅(𝒙𝒙 , 𝒙𝒙 ) = exp (− ) 2𝜎𝜎 2 (𝑖𝑖)

(𝑗𝑗)

This is often simplified to:

2

1

𝜅𝜅(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = exp (− 𝛾𝛾‖𝒙𝒙(𝑖𝑖) −𝒙𝒙(𝑗𝑗) ‖ )

Here, 𝛾𝛾 = 2𝜎𝜎 2 is a free parameter to be optimized.

Roughly speaking, the term "kernel" can be interpreted as a similarity function between a pair of examples. The minus sign inverts the distance measure into a similarity score, and, due to the exponential term, the resulting similarity score will fall into a range between 1 (for exactly similar examples) and 0 (for very dissimilar examples). Now that we have covered the big picture behind the kernel trick, let's see if we can train a kernel SVM that is able to draw a nonlinear decision boundary that separates the XOR data well. Here, we simply use the SVC class from scikit-learn that we imported earlier and replace the kernel='linear' parameter with kernel='rbf': >>> >>> >>> >>> >>> >>>

svm = SVC(kernel='rbf', random_state=1, gamma=0.10, C=10.0) svm.fit(X_xor, y_xor) plot_decision_regions(X_xor, y_xor, classifier=svm) plt.legend(loc='upper left') plt.tight_layout() plt.show()

[ 87 ]

A Tour of Machine Learning Classifiers Using scikit-learn

As we can see in the resulting plot, the kernel SVM separates the XOR data relatively well:

The 𝛾𝛾 parameter, which we set to gamma=0.1, can be understood as a cut-off parameter for the Gaussian sphere. If we increase the value for 𝛾𝛾 , we increase the influence or reach of the training examples, which leads to a tighter and bumpier decision boundary. To get a better understanding of 𝛾𝛾, let's apply an RBF kernel SVM to our Iris flower dataset: >>> >>> >>> ... ... >>> >>> >>> >>> >>>

svm = SVC(kernel='rbf', random_state=1, gamma=0.2, C=1.0) svm.fit(X_train_std, y_train) plot_decision_regions(X_combined_std, y_combined, classifier=svm, test_idx=range(105,150)) plt.xlabel('petal length [standardized]') plt.ylabel('petal width [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

[ 88 ]

Chapter 3

Since we chose a relatively small value for 𝛾𝛾, the resulting decision boundary of the RBF kernel SVM model will be relatively soft, as shown in the following plot:

Now, let's increase the value of 𝛾𝛾 and observe the effect on the decision boundary: >>> >>> >>> ... ... >>> >>> >>> >>> >>>

svm = SVC(kernel='rbf', random_state=1, gamma=100.0, C=1.0) svm.fit(X_train_std, y_train) plot_decision_regions(X_combined_std, y_combined, classifier=svm, test_idx=range(105,150)) plt.xlabel('petal length [standardized]') plt.ylabel('petal width [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

[ 89 ]

A Tour of Machine Learning Classifiers Using scikit-learn

In the resulting plot, we can now see that the decision boundary around the classes 0 and 1 is much tighter using a relatively large value of 𝛾𝛾:

Although the model fits the training dataset very well, such a classifier will likely have a high generalization error on unseen data. This illustrates that the 𝛾𝛾 parameter also plays an important role in controlling overfitting or variance when the algorithm is too sensitive to fluctuations in the training dataset.

Decision tree learning

Decision tree classifiers are attractive models if we care about interpretability. As the name "decision tree" suggests, we can think of this model as breaking down our data by making a decision based on asking a series of questions. Let's consider the following example in which we use a decision tree to decide upon an activity on a particular day:

[ 90 ]

Chapter 3

Based on the features in our training dataset, the decision tree model learns a series of questions to infer the class labels of the examples. Although the preceding figure illustrates the concept of a decision tree based on categorical variables, the same concept applies if our features are real numbers, like in the Iris dataset. For example, we could simply define a cut-off value along the sepal width feature axis and ask a binary question: "Is the sepal width ≥ 2.8 cm?" Using the decision algorithm, we start at the tree root and split the data on the feature that results in the largest information gain (IG), which will be explained in more detail in the following section. In an iterative process, we can then repeat this splitting procedure at each child node until the leaves are pure. This means that the training examples at each node all belong to the same class. In practice, this can result in a very deep tree with many nodes, which can easily lead to overfitting. Thus, we typically want to prune the tree by setting a limit for the maximal depth of the tree.

Maximizing IG – getting the most bang for your buck

In order to split the nodes at the most informative features, we need to define an objective function that we want to optimize via the tree learning algorithm. Here, our objective function is to maximize the IG at each split, which we define as follows: 𝑚𝑚

𝐼𝐼𝐼𝐼(𝐷𝐷𝑝𝑝 , 𝑓𝑓) = 𝐼𝐼(𝐷𝐷𝑝𝑝 ) − ∑ 𝑗𝑗=1

[ 91 ]

𝑁𝑁𝑗𝑗 𝐼𝐼(𝐷𝐷𝑗𝑗 ) 𝑁𝑁𝑝𝑝

A Tour of Machine Learning Classifiers Using scikit-learn

Here, f is the feature to perform the split; 𝐷𝐷𝑝𝑝 and 𝐷𝐷𝑗𝑗 are the dataset of the parent and jth child node; I is our impurity measure; 𝑁𝑁𝑝𝑝 is the total number of training examples at the parent node; and 𝑁𝑁𝑗𝑗 is the number of examples in the jth child node. As we can see, the information gain is simply the difference between the impurity of the parent node and the sum of the child node impurities—the lower the impurities of the child nodes, the larger the information gain. However, for simplicity and to reduce the combinatorial search space, most libraries (including scikit-learn) implement binary decision trees. This means that each parent node is split into two child nodes, 𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 and 𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 : 𝐼𝐼𝐼𝐼(𝐷𝐷𝑝𝑝 , 𝑓𝑓) = 𝐼𝐼(𝐷𝐷𝑝𝑝 ) −

𝑁𝑁𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑁𝑁𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 𝐼𝐼(𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) − 𝐼𝐼(𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 ) 𝑁𝑁𝑝𝑝 𝑁𝑁𝑝𝑝

The three impurity measures or splitting criteria that are commonly used in binary decision trees are Gini impurity ( 𝐼𝐼𝐺𝐺 ), entropy (𝐼𝐼𝐻𝐻 ), and the classification error (𝐼𝐼𝐸𝐸 ). Let's start with the definition of entropy for all non-empty classes ( 𝑝𝑝(𝑖𝑖|𝑡𝑡) ≠ 0): 𝑐𝑐

𝐼𝐼𝐻𝐻 (𝑡𝑡) = − ∑ 𝑝𝑝(𝑖𝑖|𝑡𝑡) 𝑙𝑙𝑙𝑙𝑙𝑙2 𝑝𝑝(𝑖𝑖|𝑡𝑡) 𝑖𝑖=1

Here, 𝑝𝑝(𝑖𝑖|𝑡𝑡) is the proportion of the examples that belong to class i for a particular node, t. The entropy is therefore 0 if all examples at a node belong to the same class, and the entropy is maximal if we have a uniform class distribution. For example, in a binary class setting, the entropy is 0 if 𝑝𝑝(𝑖𝑖 = 1|𝑡𝑡) = 1 or 𝑝𝑝(𝑖𝑖 = 0|𝑡𝑡) = 0 . If the classes are distributed uniformly with 𝑝𝑝(𝑖𝑖 = 1|𝑡𝑡) = 0.5 and 𝑝𝑝(𝑖𝑖 = 0|𝑡𝑡) = 0.5, the entropy is 1. Therefore, we can say that the entropy criterion attempts to maximize the mutual information in the tree. The Gini impurity can be understood as a criterion to minimize the probability of misclassification: 𝑐𝑐

𝑐𝑐

𝑖𝑖=1

𝑖𝑖=1

𝐼𝐼𝐺𝐺 (𝑡𝑡) = ∑ 𝑝𝑝(𝑖𝑖|𝑡𝑡)(1 − 𝑝𝑝(𝑖𝑖|𝑡𝑡)) = 1 − ∑ 𝑝𝑝(𝑖𝑖|𝑡𝑡)2

Similar to entropy, the Gini impurity is maximal if the classes are perfectly mixed, for example, in a binary class setting (c = 2): 𝑐𝑐

𝐼𝐼𝐺𝐺 (𝑡𝑡) = 1 − ∑ 0.52 = 0.5 𝑖𝑖=1

[ 92 ]

Chapter 3

However, in practice, both Gini impurity and entropy typically yield very similar results, and it is often not worth spending much time on evaluating trees using different impurity criteria rather than experimenting with different pruning cut-offs. Another impurity measure is the classification error: 𝐼𝐼𝐸𝐸 (𝑡𝑡) = 1 − max{ 𝑝𝑝(𝑖𝑖|𝑡𝑡)}

This is a useful criterion for pruning but not recommended for growing a decision tree, since it is less sensitive to changes in the class probabilities of the nodes. We can illustrate this by looking at the two possible splitting scenarios shown in the following figure:

We start with a dataset, 𝐷𝐷𝑝𝑝 , at the parent node, which consists of 40 examples from class 1 and 40 examples from class 2 that we split into two datasets, 𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 and 𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 . The information gain using the classification error as a splitting criterion would be the same ( 𝐼𝐼𝐼𝐼𝐸𝐸 = 0.25) in both scenarios, A and B: 𝐼𝐼𝐸𝐸 (𝐷𝐷𝑝𝑝 ) = 1 − 0.5 = 0.5

𝐴𝐴: 𝐼𝐼𝐸𝐸 (𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) = 1 −

3 = 0.25 4

𝐴𝐴: 𝐼𝐼𝐸𝐸 (𝐷𝐷𝑅𝑅𝑅𝑅𝑅𝑅ℎ𝑡𝑡 ) = 1 −

3 = 0.25 4

4 4 𝐴𝐴: 𝐼𝐼𝐼𝐼𝐸𝐸 = 0.5 − 0.25 − 0.25 = 0.25 8 8

𝐵𝐵: 𝐼𝐼𝐸𝐸 (𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) = 1 −

4 1 = 6 3

𝐵𝐵: 𝐼𝐼𝐸𝐸 (𝐷𝐷𝑅𝑅𝑅𝑅𝑅𝑅ℎ𝑡𝑡 ) = 1 − 1 = 0

6 1 𝐵𝐵: 𝐼𝐼𝐼𝐼𝐸𝐸 = 0.5 − × − 0 = 0.25 8 3 [ 93 ]

A Tour of Machine Learning Classifiers Using scikit-learn

However, the Gini impurity would favor the split in scenario B ( 𝐼𝐼𝐼𝐼𝐺𝐺 = 0.16̿ ) over scenario A ( 𝐼𝐼𝐼𝐼𝐺𝐺 = 0.125 ), which is indeed purer: 𝐼𝐼𝐺𝐺 (𝐷𝐷𝑝𝑝 ) = 1 − (0.52 + 0.52 ) = 0.5

3 2 1 2 3 𝐴𝐴: 𝐼𝐼𝐺𝐺 (𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) = 1 − (( ) + ( ) ) = = 0.375 4 4 8

1 2 3 2 3 𝐴𝐴: 𝐼𝐼𝐺𝐺 (𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 ) = 1 − (( ) + ( ) ) = = 0.375 4 4 8

𝐴𝐴: 𝐼𝐼𝐼𝐼𝐺𝐺 = 0.5 −

4 4 0.375 − 0.375 = 0.125 8 8

2 2 4 2 4 𝐵𝐵: 𝐼𝐼𝐺𝐺 (𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) = 1 − (( ) + ( ) ) = = 0. 4̅ 6 6 9 𝐵𝐵: 𝐼𝐼𝐺𝐺 (𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 ) = 1 − (12 + 02 ) = 0

6 ̅̅̅̅ 𝐵𝐵: 𝐼𝐼𝐼𝐼𝐺𝐺 = 0.5 − 0. 4̅ − 0 = 0. 16 8

Similarly, the entropy criterion would also favor scenario B ( 𝐼𝐼𝐼𝐼𝐻𝐻 = 0.31 ) over scenario A ( 𝐼𝐼𝐼𝐼𝐻𝐻 = 0.19): 𝐼𝐼𝐻𝐻 (𝐷𝐷𝑝𝑝 ) = −(0.5 𝑙𝑙𝑙𝑙𝑙𝑙2 (0.5) + 0.5 𝑙𝑙𝑙𝑙𝑙𝑙2 (0.5)) = 1

3 3 1 1 𝐴𝐴: 𝐼𝐼𝐻𝐻 (𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) = − ( 𝑙𝑙𝑙𝑙𝑙𝑙2 ( ) + 𝑙𝑙𝑙𝑙𝑔𝑔2 ( )) = 0.81 4 4 4 4

1 1 3 3 𝐴𝐴: 𝐼𝐼𝐻𝐻 (𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 ) = − ( 𝑙𝑙𝑙𝑙𝑙𝑙2 ( ) + 𝑙𝑙𝑙𝑙𝑙𝑙2 ( )) = 0.81 4 4 4 4 𝐴𝐴: 𝐼𝐼𝐼𝐼𝐻𝐻 = 1 −

4 4 0.81 − 0.81 = 0.19 8 8

2 2 4 4 𝐵𝐵: 𝐼𝐼𝐻𝐻 (𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) = − ( 𝑙𝑙𝑙𝑙𝑙𝑙2 ( ) + 𝑙𝑙𝑙𝑙𝑙𝑙2 ( )) = 0.92 6 6 6 6 𝐵𝐵: 𝐼𝐼𝐻𝐻 (𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 ) = 0

6 𝐵𝐵: 𝐼𝐼𝐼𝐼𝐻𝐻 = 1 − 0.92 − 0 = 0.31 8 [ 94 ]

Chapter 3

For a more visual comparison of the three different impurity criteria that we discussed previously, let's plot the impurity indices for the probability range [0, 1] for class 1. Note that we will also add a scaled version of the entropy (entropy / 2) to observe that the Gini impurity is an intermediate measure between entropy and the classification error. The code is as follows: >>> >>> >>> ... >>> ... >>> ... >>> >>> >>> >>> >>> >>> >>> ... ... ... ... ... ... ... ... >>> ... >>> >>> >>> >>> >>>

import matplotlib.pyplot as plt import numpy as np def gini(p): return (p)*(1 - (p)) + (1 - p)*(1 - (1-p)) def entropy(p): return - p*np.log2(p) - (1 - p)*np.log2((1 - p)) def error(p): return 1 - np.max([p, 1 - p]) x = np.arange(0.0, 1.0, 0.01) ent = [entropy(p) if p != 0 else None for p in x] sc_ent = [e*0.5 if e else None for e in ent] err = [error(i) for i in x] fig = plt.figure() ax = plt.subplot(111) for i, lab, ls, c, in zip([ent, sc_ent, gini(x), err], ['Entropy', 'Entropy (scaled)', 'Gini impurity', 'Misclassification error'], ['-', '-', '--', '-.'], ['black', 'lightgray', 'red', 'green', 'cyan']): line = ax.plot(x, i, label=lab, linestyle=ls, lw=2, color=c) ax.legend(loc='upper center', bbox_to_anchor=(0.5, 1.15), ncol=5, fancybox=True, shadow=False) ax.axhline(y=0.5, linewidth=1, color='k', linestyle='--') ax.axhline(y=1.0, linewidth=1, color='k', linestyle='--') plt.ylim([0, 1.1]) plt.xlabel('p(i=1)') plt.ylabel('impurity index')

[ 95 ]

A Tour of Machine Learning Classifiers Using scikit-learn

The plot produced by the preceding code example is as follows:

Building a decision tree

Decision trees can build complex decision boundaries by dividing the feature space into rectangles. However, we have to be careful since the deeper the decision tree, the more complex the decision boundary becomes, which can easily result in overfitting. Using scikit-learn, we will now train a decision tree with a maximum depth of 4, using Gini impurity as a criterion for impurity. Although feature scaling may be desired for visualization purposes, note that feature scaling is not a requirement for decision tree algorithms. The code is as follows: >>> >>> ... ... >>> >>>

from sklearn.tree import DecisionTreeClassifier tree_model = DecisionTreeClassifier(criterion='gini', max_depth=4, random_state=1) tree_model.fit(X_train, y_train) X_combined = np.vstack((X_train, X_test))

[ 96 ]

Chapter 3 >>> >>> ... ... ... >>> >>> >>> >>> >>>

y_combined = np.hstack((y_train, y_test)) plot_decision_regions(X_combined, y_combined, classifier=tree_model, test_idx=range(105, 150)) plt.xlabel('petal length [cm]') plt.ylabel('petal width [cm]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

After executing the code example, we get the typical axis-parallel decision boundaries of the decision tree:

A nice feature in scikit-learn is that it allows us to readily visualize the decision tree model after training via the following code: >>> from sklearn import tree >>> tree.plot_tree(tree_model) >>> plt.show()

[ 97 ]

A Tour of Machine Learning Classifiers Using scikit-learn

However, nicer visualizations can be obtained by using the Graphviz program as a backend for plotting scikit-learn decision trees. This program is freely available from http://www.graphviz.org and is supported by Linux, Windows, and macOS. In addition to Graphviz, we will use a Python library called PyDotPlus, which has capabilities similar to Graphviz and allows us to convert .dot data files into a decision tree image file. After you have installed Graphviz (by following the instructions on http://www.graphviz.org/download), you can install PyDotPlus directly via the pip installer, for example, by executing the following command in your command-line terminal: > pip3 install pydotplus

Installing PyDotPlus prerequisites Note that on some systems, you may have to install the PyDotPlus prerequisites manually by executing the following commands: pip3 install graphviz pip3 install pyparsing

The following code will create an image of our decision tree in PNG format in our local directory:

[ 98 ]

Chapter 3 >>> >>> >>> ... ... ... ... ... ... ... ... >>> >>>

from pydotplus import graph_from_dot_data from sklearn.tree import export_graphviz dot_data = export_graphviz(tree_model, filled=True, rounded=True, class_names=['Setosa', 'Versicolor', 'Virginica'], feature_names=['petal length', 'petal width'], out_file=None) graph = graph_from_dot_data(dot_data) graph.write_png('tree.png')

By using the out_file=None setting, we directly assigned the DOT data to a dot_data variable, instead of writing an intermediate tree.dot file to disk. The arguments for filled, rounded, class_names, and feature_names are optional, but make the resulting image file visually more appealing by adding color, rounding the box edges, showing the name of the majority class label at each node, and displaying the feature name in each splitting criterion. These settings resulted in the following decision tree image:

[ 99 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Looking at the decision tree figure, we can now nicely trace back the splits that the decision tree determined from our training dataset. We started with 105 examples at the root and split them into two child nodes with 35 and 70 examples, using the petal width cut-off ≤ 0.75 cm. After the first split, we can see that the left child node is already pure and only contains examples from the Iris-setosa class (Gini impurity = 0). The further splits on the right are then used to separate the examples from the Iris-versicolor and Iris-virginica class. Looking at this tree, and the decision region plot of the tree, we can see that the decision tree does a very good job of separating the flower classes. Unfortunately, scikit-learn currently does not implement functionality to manually post-prune a decision tree. However, we could go back to our previous code example, change the max_depth of our decision tree to 3, and compare it to our current model, but we leave this as an exercise for the interested reader.

Combining multiple decision trees via random forests

Ensemble methods have gained huge popularity in applications of machine learning during the last decade due to their good classification performance and robustness toward overfitting. While we are going to cover different ensemble methods, including bagging and boosting, later in Chapter 7, Combining Different Models for Ensemble Learning, let's discuss the decision tree-based random forest algorithm, which is known for its good scalability and ease of use. A random forest can be considered as an ensemble of decision trees. The idea behind a random forest is to average multiple (deep) decision trees that individually suffer from high variance to build a more robust model that has a better generalization performance and is less susceptible to overfitting. The random forest algorithm can be summarized in four simple steps: 1. Draw a random bootstrap sample of size n (randomly choose n examples from the training dataset with replacement). 2. Grow a decision tree from the bootstrap sample. At each node: a. Randomly select d features without replacement.

b. Split the node using the feature that provides the best split according to the objective function, for instance, maximizing the information gain. 3. Repeat the steps 1-2 k times. 4. Aggregate the prediction by each tree to assign the class label by majority vote. Majority voting will be discussed in more detail in Chapter 7, Combining Different Models for Ensemble Learning. [ 100 ]

Chapter 3

We should note one slight modification in step 2 when we are training the individual decision trees: instead of evaluating all features to determine the best split at each node, we only consider a random subset of those. Sampling with and without replacement In case you are not familiar with the terms sampling "with" and "without" replacement, let's walk through a simple thought experiment. let's assume that we are playing a lottery game where we randomly draw numbers from an urn. We start with an urn that holds five unique numbers, 0, 1, 2, 3, and 4, and we draw exactly one number each turn. In the first round, the chance of drawing a particular number from the urn would be 1/5. Now, in sampling without replacement, we do not put the number back into the urn after each turn. Consequently, the probability of drawing a particular number from the set of remaining numbers in the next round depends on the previous round. For example, if we have a remaining set of numbers 0, 1, 2, and 4, the chance of drawing number 0 would become 1/4 in the next turn. However, in random sampling with replacement, we always return the drawn number to the urn so that the probability of drawing a particular number at each turn does not change; we can draw the same number more than once. In other words, in sampling with replacement, the samples (numbers) are independent and have a covariance of zero. For example, the results from five rounds of drawing random numbers could look like this: • Random sampling without replacement: 2, 1, 3, 4, 0 • Random sampling with replacement: 1, 3, 3, 4, 1

Although random forests don't offer the same level of interpretability as decision trees, a big advantage of random forests is that we don't have to worry so much about choosing good hyperparameter values. We typically don't need to prune the random forest since the ensemble model is quite robust to noise from the individual decision trees. The only parameter that we really need to care about in practice is the number of trees, k, (step 3) that we choose for the random forest. Typically, the larger the number of trees, the better the performance of the random forest classifier at the expense of an increased computational cost.

[ 101 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Although it is less common in practice, other hyperparameters of the random forest classifier that can be optimized—using techniques that we will discuss in Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning—are the size, n, of the bootstrap sample (step 1), and the number of features, d, that are randomly chosen for each split (step 2.a), respectively. Via the sample size, n, of the bootstrap sample, we control the bias-variance tradeoff of the random forest. Decreasing the size of the bootstrap sample increases the diversity among the individual trees, since the probability that a particular training example is included in the bootstrap sample is lower. Thus, shrinking the size of the bootstrap samples may increase the randomness of the random forest, and it can help to reduce the effect of overfitting. However, smaller bootstrap samples typically result in a lower overall performance of the random forest, and a small gap between training and test performance, but a low test performance overall. Conversely, increasing the size of the bootstrap sample may increase the degree of overfitting. Because the bootstrap samples, and consequently the individual decision trees, become more similar to each other, they learn to fit the original training dataset more closely. In most implementations, including the RandomForestClassifier implementation in scikit-learn, the size of the bootstrap sample is chosen to be equal to the number of training examples in the original training dataset, which usually provides a good bias-variance tradeoff. For the number of features, d, at each split, we want to choose a value that is smaller than the total number of features in the training dataset. A reasonable default that is used in scikit-learn and other implementations is 𝑑𝑑 = √𝑚𝑚 , where m is the number of features in the training dataset.

Conveniently, we don't have to construct the random forest classifier from individual decision trees by ourselves because there is already an implementation in scikit-learn that we can use: >>> >>> ... ... ... >>> >>> ... >>> >>> >>> >>> >>>

from sklearn.ensemble import RandomForestClassifier forest = RandomForestClassifier(criterion='gini', n_estimators=25, random_state=1, n_jobs=2) forest.fit(X_train, y_train) plot_decision_regions(X_combined, y_combined, classifier=forest, test_idx=range(105,150)) plt.xlabel('petal length [cm]') plt.ylabel('petal width [cm]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

[ 102 ]

Chapter 3

After executing the preceding code, we should see the decision regions formed by the ensemble of trees in the random forest, as shown in the following plot:

Using the preceding code, we trained a random forest from 25 decision trees via the n_estimators parameter and used the Gini impurity measure as a criterion to split the nodes. Although we are growing a very small random forest from a very small training dataset, we used the n_jobs parameter for demonstration purposes, which allows us to parallelize the model training using multiple cores of our computer (here, two cores).

K-nearest neighbors – a lazy learning algorithm

The last supervised learning algorithm that we want to discuss in this chapter is the k-nearest neighbor (KNN) classifier, which is particularly interesting because it is fundamentally different from the learning algorithms that we have discussed so far. KNN is a typical example of a lazy learner. It is called "lazy" not because of its apparent simplicity, but because it doesn't learn a discriminative function from the training data but memorizes the training dataset instead.

[ 103 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Parametric versus nonparametric models Machine learning algorithms can be grouped into parametric and nonparametric models. Using parametric models, we estimate parameters from the training dataset to learn a function that can classify new data points without requiring the original training dataset anymore. Typical examples of parametric models are the perceptron, logistic regression, and the linear SVM. In contrast, nonparametric models can't be characterized by a fixed set of parameters, and the number of parameters grows with the training data. Two examples of nonparametric models that we have seen so far are the decision tree classifier/ random forest and the kernel SVM. KNN belongs to a subcategory of nonparametric models that is described as instance-based learning. Models based on instance-based learning are characterized by memorizing the training dataset, and lazy learning is a special case of instancebased learning that is associated with no (zero) cost during the learning process.

The KNN algorithm itself is fairly straightforward and can be summarized by the following steps: 1. Choose the number of k and a distance metric. 2. Find the k-nearest neighbors of the data record that we want to classify. 3. Assign the class label by majority vote. The following figure illustrates how a new data point (?) is assigned the triangle class label based on majority voting among its five nearest neighbors.

[ 104 ]

Chapter 3

Based on the chosen distance metric, the KNN algorithm finds the k examples in the training dataset that are closest (most similar) to the point that we want to classify. The class label of the data point is then determined by a majority vote among its k nearest neighbors. The main advantage of such a memory-based approach is that the classifier immediately adapts as we collect new training data. However, the downside is that the computational complexity for classifying new examples grows linearly with the number of examples in the training dataset in the worst-case scenario—unless the dataset has very few dimensions (features) and the algorithm has been implemented using efficient data structures such as k-d trees (An Algorithm for Finding Best Matches in Logarithmic Expected Time, J. H. Friedman, J. L. Bentley, and R.A. Finkel, ACM transactions on mathematical software (TOMS), 3(3): 209–226, 1977). Furthermore, we can't discard training examples since no training step is involved. Thus, storage space can become a challenge if we are working with large datasets.

[ 105 ]

A Tour of Machine Learning Classifiers Using scikit-learn

By executing the following code, we will now implement a KNN model in scikitlearn using a Euclidean distance metric: >>> >>> ... >>> >>> ... >>> >>> >>> >>> >>>

from sklearn.neighbors import KNeighborsClassifier knn = KNeighborsClassifier(n_neighbors=5, p=2, metric='minkowski') knn.fit(X_train_std, y_train) plot_decision_regions(X_combined_std, y_combined, classifier=knn, test_idx=range(105,150)) plt.xlabel('petal length [standardized]') plt.ylabel('petal width [standardized]') plt.legend(loc='upper left') plt.tight_layout() plt.show()

By specifying five neighbors in the KNN model for this dataset, we obtain a relatively smooth decision boundary, as shown in the following plot:

Resolving ties In the case of a tie, the scikit-learn implementation of the KNN algorithm will prefer the neighbors with a closer distance to the data record to be classified. If the neighbors have similar distances, the algorithm will choose the class label that comes first in the training dataset.

[ 106 ]

Chapter 3

The right choice of k is crucial to finding a good balance between overfitting and underfitting. We also have to make sure that we choose a distance metric that is appropriate for the features in the dataset. Often, a simple Euclidean distance measure is used for real-value examples, for example, the flowers in our Iris dataset, which have features measured in centimeters. However, if we are using a Euclidean distance measure, it is also important to standardize the data so that each feature contributes equally to the distance. The minkowski distance that we used in the previous code is just a generalization of the Euclidean and Manhattan distance, which can be written as follows: 𝑝𝑝

𝑑𝑑(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = √∑|𝑥𝑥𝑘𝑘(𝑖𝑖) 𝑘𝑘

𝑝𝑝

(𝑗𝑗) 𝑥𝑥𝑘𝑘 | ,

It becomes the Euclidean distance if we set the parameter p=2 or the Manhattan distance at p=1. Many other distance metrics are available in scikit-learn and can be provided to the metric parameter. They are listed at http://scikit-learn.org/ stable/modules/generated/sklearn.neighbors.DistanceMetric.html. The curse of dimensionality It is important to mention that KNN is very susceptible to overfitting due to the curse of dimensionality. The curse of dimensionality describes the phenomenon where the feature space becomes increasingly sparse for an increasing number of dimensions of a fixed-size training dataset. We can think of even the closest neighbors as being too far away in a highdimensional space to give a good estimate. We discussed the concept of regularization in the section about logistic regression as one way to avoid overfitting. However, in models where regularization is not applicable, such as decision trees and KNN, we can use feature selection and dimensionality reduction techniques to help us to avoid the curse of dimensionality. This will be discussed in more detail in the next chapter.

[ 107 ]

A Tour of Machine Learning Classifiers Using scikit-learn

Summary

In this chapter, you learned about many different machine learning algorithms that are used to tackle linear and nonlinear problems. You have seen that decision trees are particularly attractive if we care about interpretability. Logistic regression is not only a useful model for online learning via SGD, but also allows us to predict the probability of a particular event. Although SVMs are powerful linear models that can be extended to nonlinear problems via the kernel trick, they have many parameters that have to be tuned in order to make good predictions. In contrast, ensemble methods, such as random forests, don't require much parameter tuning and don't overfit as easily as decision trees, which makes them attractive models for many practical problem domains. The KNN classifier offers an alternative approach to classification via lazy learning that allows us to make predictions without any model training, but with a more computationally expensive prediction step. However, even more important than the choice of an appropriate learning algorithm is the available data in our training dataset. No algorithm will be able to make good predictions without informative and discriminatory features. In the next chapter, we will discuss important topics regarding the preprocessing of data, feature selection, and dimensionality reduction, which means that we will need to build powerful machine learning models. Later, in Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, we will see how we can evaluate and compare the performance of our models and learn useful tricks to fine-tune the different algorithms.

[ 108 ]

Building Good Training Datasets – Data Preprocessing The quality of the data and the amount of useful information that it contains are key factors that determine how well a machine learning algorithm can learn. Therefore, it is absolutely critical to ensure that we examine and preprocess a dataset before we feed it to a learning algorithm. In this chapter, we will discuss the essential data preprocessing techniques that will help us to build good machine learning models. The topics that we will cover in this chapter are as follows: • Removing and imputing missing values from the dataset • Getting categorical data into shape for machine learning algorithms • Selecting relevant features for the model construction

Dealing with missing data

It is not uncommon in real-world applications for our training examples to be missing one or more values for various reasons. There could have been an error in the data collection process, certain measurements may not be applicable, or particular fields could have been simply left blank in a survey, for example. We typically see missing values as blank spaces in our data table or as placeholder strings such as NaN, which stands for "not a number," or NULL (a commonly used indicator of unknown values in relational databases). Unfortunately, most computational tools are unable to handle such missing values or will produce unpredictable results if we simply ignore them. Therefore, it is crucial that we take care of those missing values before we proceed with further analyses.

[ 109 ]

Building Good Training Datasets – Data Preprocessing

In this section, we will work through several practical techniques for dealing with missing values by removing entries from our dataset or imputing missing values from other training examples and features.

Identifying missing values in tabular data

Before we discuss several techniques for dealing with missing values, let's create a simple example data frame from a comma-separated values (CSV) file to get a better grasp of the problem: >>> import pandas as pd >>> from io import StringIO >>> ... ... ... ... >>> >>> >>> >>> >>> 0 1 2

csv_data = \ '''A,B,C,D 1.0,2.0,3.0,4.0 5.0,6.0,,8.0 10.0,11.0,12.0,''' # If you are using Python 2.7, you need # to convert the string to unicode: # csv_data = unicode(csv_data) df = pd.read_csv(StringIO(csv_data)) df A B C D 1.0 2.0 3.0 4.0 5.0 6.0 NaN 8.0 10.0 11.0 12.0 NaN

Using the preceding code, we read CSV-formatted data into a pandas DataFrame via the read_csv function and noticed that the two missing cells were replaced by NaN. The StringIO function in the preceding code example was simply used for the purposes of illustration. It allowed us to read the string assigned to csv_data into a pandas DataFrame as if it was a regular CSV file on our hard drive. For a larger DataFrame, it can be tedious to look for missing values manually; in this case, we can use the isnull method to return a DataFrame with Boolean values that indicate whether a cell contains a numeric value (False) or if data is missing (True). Using the sum method, we can then return the number of missing values per column as follows: >>> df.isnull().sum() A 0 B 0 C 1 [ 110 ]

Chapter 4 D 1 dtype: int64

This way, we can count the number of missing values per column; in the following subsections, we will take a look at different strategies for how to deal with this missing data. Convenient data handling with pandas' data frames Although scikit-learn was originally developed for working with NumPy arrays only, it can sometimes be more convenient to preprocess data using pandas' DataFrame. Nowadays, most scikit-learn functions support DataFrame objects as inputs, but since NumPy array handling is more mature in the scikit-learn API, it is recommended to use NumPy arrays when possible. Note that you can always access the underlying NumPy array of a DataFrame via the values attribute before you feed it into a scikit-learn estimator: >>> df.values array([[ 1., [ 5., [ 10.,

2., 6., 11.,

3., nan, 12.,

4.], 8.], nan]])

Eliminating training examples or features with missing values

One of the easiest ways to deal with missing data is simply to remove the corresponding features (columns) or training examples (rows) from the dataset entirely; rows with missing values can easily be dropped via the dropna method: >>> df.dropna(axis=0) A B C D 0 1.0 2.0 3.0 4.0

Similarly, we can drop columns that have at least one NaN in any row by setting the axis argument to 1: >>> df.dropna(axis=1) A B 0 1.0 2.0 1 5.0 6.0 2 10.0 11.0 [ 111 ]

Building Good Training Datasets – Data Preprocessing

The dropna method supports several additional parameters that can come in handy: # only drop rows where all columns are NaN # (returns the whole array here since we don't # have a row with all values NaN) >>> df.dropna(how='all') A B C D 0 1.0 2.0 3.0 4.0 1 5.0 6.0 NaN 8.0 2 10.0 11.0 12.0 NaN

# drop rows that have fewer than 4 real values >>> df.dropna(thresh=4) A B C D 0 1.0 2.0 3.0 4.0

# only drop rows where NaN appear in specific columns (here: 'C') >>> df.dropna(subset=['C']) A B C D 0 1.0 2.0 3.0 4.0 2 10.0 11.0 12.0 NaN

Although the removal of missing data seems to be a convenient approach, it also comes with certain disadvantages; for example, we may end up removing too many samples, which will make a reliable analysis impossible. Or, if we remove too many feature columns, we will run the risk of losing valuable information that our classifier needs to discriminate between classes. In the next section, we will look at one of the most commonly used alternatives for dealing with missing values: interpolation techniques.

Imputing missing values

Often, the removal of training examples or dropping of entire feature columns is simply not feasible, because we might lose too much valuable data. In this case, we can use different interpolation techniques to estimate the missing values from the other training examples in our dataset. One of the most common interpolation techniques is mean imputation, where we simply replace the missing value with the mean value of the entire feature column. A convenient way to achieve this is by using the SimpleImputer class from scikit-learn, as shown in the following code:

[ 112 ]

Chapter 4 >>> >>> >>> >>> >>> >>>

from sklearn.impute import SimpleImputer import numpy as np imr = SimpleImputer(missing_values=np.nan, strategy='mean') imr = imr.fit(df.values) imputed_data = imr.transform(df.values) imputed_data

array([[ 1., [ 5., [ 10.,

2., 6., 11.,

3., 7.5, 12.,

4.], 8.], 6.]])

Here, we replaced each NaN value with the corresponding mean, which is separately calculated for each feature column. Other options for the strategy parameter are median or most_frequent, where the latter replaces the missing values with the most frequent values. This is useful for imputing categorical feature values, for example, a feature column that stores an encoding of color names, such as red, green, and blue. We will encounter examples of such data later in this chapter. Alternatively, an even more convenient way to impute missing values is by using pandas' fillna method and providing an imputation method as an argument. For example, using pandas, we could achieve the same mean imputation directly in the DataFrame object via the following command: >>> df.fillna(df.mean())

Understanding the scikit-learn estimator API

In the previous section, we used the SimpleImputer class from scikit-learn to impute missing values in our dataset. The SimpleImputer class belongs to the so-called transformer classes in scikit-learn, which are used for data transformation. The two essential methods of those estimators are fit and transform. The fit method is used to learn the parameters from the training data, and the transform method uses those parameters to transform the data. Any data array that is to be transformed needs to have the same number of features as the data array that was used to fit the model.

[ 113 ]

Building Good Training Datasets – Data Preprocessing

The following figure illustrates how a transformer, fitted on the training data, is used to transform a training dataset as well as a new test dataset:

The classifiers that we used in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, belong to the so-called estimators in scikit-learn, with an API that is conceptually very similar to the transformer class. Estimators have a predict method but can also have a transform method, as you will see later in this chapter. As you may recall, we also used the fit method to learn the parameters of a model when we trained those estimators for classification. However, in supervised learning tasks, we additionally provide the class labels for fitting the model, which can then be used to make predictions about new, unlabeled data examples via the predict method, as illustrated in the following figure:

[ 114 ]

Chapter 4

Handling categorical data

So far, we have only been working with numerical values. However, it is not uncommon for real-world datasets to contain one or more categorical feature columns. In this section, we will make use of simple yet effective examples to see how to deal with this type of data in numerical computing libraries. When we are talking about categorical data, we have to further distinguish between ordinal and nominal features. Ordinal features can be understood as categorical values that can be sorted or ordered. For example, t-shirt size would be an ordinal feature, because we can define an order: XL > L > M. In contrast, nominal features don't imply any order and, to continue with the previous example, we could think of t-shirt color as a nominal feature since it typically doesn't make sense to say that, for example, red is larger than blue.

[ 115 ]

Building Good Training Datasets – Data Preprocessing

Categorical data encoding with pandas

Before we explore different techniques for handling such categorical data, let's create a new DataFrame to illustrate the problem: >>> >>> ... ... ... >>> >>> 0 1 2

import pandas as pd df = pd.DataFrame([ ['green', 'M', 10.1, 'class2'], ['red', 'L', 13.5, 'class1'], ['blue', 'XL', 15.3, 'class2']]) df.columns = ['color', 'size', 'price', 'classlabel'] df color size price classlabel green M 10.1 class2 red L 13.5 class1 blue XL 15.3 class2

As we can see in the preceding output, the newly created DataFrame contains a nominal feature (color), an ordinal feature (size), and a numerical feature (price) column. The class labels (assuming that we created a dataset for a supervised learning task) are stored in the last column. The learning algorithms for classification that we discuss in this book do not use ordinal information in class labels.

Mapping ordinal features

To make sure that the learning algorithm interprets the ordinal features correctly, we need to convert the categorical string values into integers. Unfortunately, there is no convenient function that can automatically derive the correct order of the labels of our size feature, so we have to define the mapping manually. In the following simple example, let's assume that we know the numerical difference between features, for example, XL = L + 1 = M + 2: >>> size_mapping = {'XL': 3, ... 'L': 2, ... 'M': 1} >>> df['size'] = df['size'].map(size_mapping) >>> df color size price classlabel 0 green 1 10.1 class2 1 red 2 13.5 class1 2 blue 3 15.3 class2

[ 116 ]

Chapter 4

If we want to transform the integer values back to the original string representation at a later stage, we can simply define a reverse-mapping dictionary, inv_size_ mapping = {v: k for k, v in size_mapping.items()}, which can then be used via the pandas map method on the transformed feature column and is similar to the size_mapping dictionary that we used previously. We can use it as follows: >>> inv_size_mapping = {v: k for k, v in size_mapping.items()} >>> df['size'].map(inv_size_mapping) 0 M 1 L 2 XL Name: size, dtype: object

Encoding class labels

Many machine learning libraries require that class labels are encoded as integer values. Although most estimators for classification in scikit-learn convert class labels to integers internally, it is considered good practice to provide class labels as integer arrays to avoid technical glitches. To encode the class labels, we can use an approach similar to the mapping of ordinal features discussed previously. We need to remember that class labels are not ordinal, and it doesn't matter which integer number we assign to a particular string label. Thus, we can simply enumerate the class labels, starting at 0: >>> import numpy as np >>> class_mapping = {label: idx for idx, label in ... enumerate(np.unique(df['classlabel']))} >>> class_mapping {'class1': 0, 'class2': 1}

Next, we can use the mapping dictionary to transform the class labels into integers: >>> df['classlabel'] = df['classlabel'].map(class_mapping) >>> df color size price classlabel 0 green 1 10.1 1 1 red 2 13.5 0 2 blue 3 15.3 1

We can reverse the key-value pairs in the mapping dictionary as follows to map the converted class labels back to the original string representation: >>> inv_class_mapping = {v: k for k, v in class_mapping.items()} >>> df['classlabel'] = df['classlabel'].map(inv_class_mapping) >>> df

[ 117 ]

Building Good Training Datasets – Data Preprocessing

0 1 2

color green red blue

size 1 2 3

price 10.1 13.5 15.3

classlabel class2 class1 class2

Alternatively, there is a convenient LabelEncoder class directly implemented in scikit-learn to achieve this: >>> from sklearn.preprocessing import LabelEncoder >>> class_le = LabelEncoder() >>> y = class_le.fit_transform(df['classlabel'].values) >>> y array([1, 0, 1])

Note that the fit_transform method is just a shortcut for calling fit and transform separately, and we can use the inverse_transform method to transform the integer class labels back into their original string representation: >>> class_le.inverse_transform(y) array(['class2', 'class1', 'class2'], dtype=object)

Performing one-hot encoding on nominal features

In the previous Mapping ordinal features section, we used a simple dictionarymapping approach to convert the ordinal size feature into integers. Since scikitlearn's estimators for classification treat class labels as categorical data that does not imply any order (nominal), we used the convenient LabelEncoder to encode the string labels into integers. It may appear that we could use a similar approach to transform the nominal color column of our dataset, as follows: >>> X = df[['color', 'size', 'price']].values >>> color_le = LabelEncoder() >>> X[:, 0] = color_le.fit_transform(X[:, 0]) >>> X array([[1, 1, 10.1], [2, 2, 13.5], [0, 3, 15.3]], dtype=object)

After executing the preceding code, the first column of the NumPy array, X, now holds the new color values, which are encoded as follows: • blue = 0 • green = 1 • red = 2 [ 118 ]

Chapter 4

If we stop at this point and feed the array to our classifier, we will make one of the most common mistakes in dealing with categorical data. Can you spot the problem? Although the color values don't come in any particular order, a learning algorithm will now assume that green is larger than blue, and red is larger than green. Although this assumption is incorrect, the algorithm could still produce useful results. However, those results would not be optimal. A common workaround for this problem is to use a technique called one-hot encoding. The idea behind this approach is to create a new dummy feature for each unique value in the nominal feature column. Here, we would convert the color feature into three new features: blue, green, and red. Binary values can then be used to indicate the particular color of an example; for example, a blue example can be encoded as blue=1, green=0, red=0. To perform this transformation, we can use the OneHotEncoder that is implemented in scikit-learn's preprocessing module: >>> >>> >>> >>>

from sklearn.preprocessing import OneHotEncoder X = df[['color', 'size', 'price']].values color_ohe = OneHotEncoder() color_ohe.fit_transform(X[:, 0].reshape(-1, 1)).toarray() array([[0., 1., 0.], [0., 0., 1.], [1., 0., 0.]])

Note that we applied the OneHotEncoder to only a single column, (X[:, 0].reshape(-1, 1))), to avoid modifying the other two columns in the array as well. If we want to selectively transform columns in a multi-feature array, we can use the ColumnTransformer, which accepts a list of (name, transformer, column(s)) tuples as follows: >>> >>> >>> ... ... ... >>>

from sklearn.compose import ColumnTransformer X = df[['color', 'size', 'price']].values c_transf = ColumnTransformer([ ('onehot', OneHotEncoder(), [0]), ('nothing', 'passthrough', [1, 2]) ]) c_transf.fit_transform(X).astype(float) array([[0.0, 1.0, 0.0, 1, 10.1], [0.0, 0.0, 1.0, 2, 13.5], [1.0, 0.0, 0.0, 3, 15.3]])

In the preceding code example, we specified that we want to modify only the first column and leave the other two columns untouched via the 'passthrough' argument.

[ 119 ]

Building Good Training Datasets – Data Preprocessing

An even more convenient way to create those dummy features via one-hot encoding is to use the get_dummies method implemented in pandas. Applied to a DataFrame, the get_dummies method will only convert string columns and leave all other columns unchanged: >>> pd.get_dummies(df[['price', 'color', 'size']]) price size color_blue color_green color_red 0 10.1 1 0 1 0 1 13.5 2 0 0 1 2 15.3 3 1 0 0

When we are using one-hot encoding datasets, we have to keep in mind that this introduces multicollinearity, which can be an issue for certain methods (for instance, methods that require matrix inversion). If features are highly correlated, matrices are computationally difficult to invert, which can lead to numerically unstable estimates. To reduce the correlation among variables, we can simply remove one feature column from the one-hot encoded array. Note that we do not lose any important information by removing a feature column, though; for example, if we remove the column color_blue, the feature information is still preserved since if we observe color_green=0 and color_red=0, it implies that the observation must be blue. If we use the get_dummies function, we can drop the first column by passing a True argument to the drop_first parameter, as shown in the following code example: >>> pd.get_dummies(df[['price', 'color', 'size']], ... drop_first=True) price size color_green color_red 0 10.1 1 1 0 1 13.5 2 0 1 2 15.3 3 0 0

In order to drop a redundant column via the OneHotEncoder, we need to set drop='first' and set categories='auto' as follows: >>> >>> ... ... ... >>>

color_ohe = OneHotEncoder(categories='auto', drop='first') c_transf = ColumnTransformer([ ('onehot', color_ohe, [0]), ('nothing', 'passthrough', [1, 2]) ]) c_transf.fit_transform(X).astype(float)

array([[ [ [

1. , 0. , 0. ,

0. , 1. , 0. ,

1. , 2. , 3. ,

10.1], 13.5], 15.3]])

[ 120 ]

Chapter 4

Optional: encoding ordinal features If we are unsure about the numerical differences between the categories of ordinal features, or the difference between two ordinal values is not defined, we can also encode them using a threshold encoding with 0/1 values. For example, we can split the feature size with values M, L, and XL into two new features, "x > M" and "x > L". Let's consider the original DataFrame: >>> df = pd.DataFrame([['green', 'M', 10.1, ... 'class2'], ... ['red', 'L', 13.5, ... 'class1'], ... ['blue', 'XL', 15.3, ... 'class2']]) >>> df.columns = ['color', 'size', 'price', ... 'classlabel'] >>> df

We can use the apply method of pandas' DataFrames to write custom lambda expressions in order to encode these variables using the value-threshold approach: >>> ... >>> ... >>> >>>

df['x > M'] = df['size'].apply( lambda x: 1 if x in {'L', 'XL'} else 0) df['x > L'] = df['size'].apply( lambda x: 1 if x == 'XL' else 0) del df['size'] df

Partitioning a dataset into separate training and test datasets

We briefly introduced the concept of partitioning a dataset into separate datasets for training and testing in Chapter 1, Giving Computers the Ability to Learn from Data, and Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn. Remember that comparing predictions to true labels in the test set can be understood as the unbiased performance evaluation of our model before we let it loose on the real world. In this section, we will prepare a new dataset, the Wine dataset. After we have preprocessed the dataset, we will explore different techniques for feature selection to reduce the dimensionality of a dataset. [ 121 ]

Building Good Training Datasets – Data Preprocessing

The Wine dataset is another open source dataset that is available from the UCI machine learning repository (https://archive.ics.uci.edu/ml/datasets/Wine); it consists of 178 wine examples with 13 features describing their different chemical properties. Obtaining the Wine dataset You can find a copy of the Wine dataset (and all other datasets used in this book) in the code bundle of this book, which you can use if you are working offline or the dataset at https:// archive.ics.uci.edu/ml/machine-learning-databases/ wine/wine.data is temporarily unavailable on the UCI server.

For instance, to load the Wine dataset from a local directory, you can replace this line: df = pd.read_csv( 'https://archive.ics.uci.edu/ml/' 'machine-learning-databases/wine/wine.data', header=None)

with the following one: df = pd.read_csv( 'your/local/path/to/wine.data', header=None)

Using the pandas library, we will directly read in the open source Wine dataset from the UCI machine learning repository: >>> df_wine = pd.read_csv('https://archive.ics.uci.edu/' ... 'ml/machine-learning-databases/' ... 'wine/wine.data', header=None) >>> df_wine.columns = ['Class label', 'Alcohol', ... 'Malic acid', 'Ash', ... 'Alcalinity of ash', 'Magnesium', ... 'Total phenols', 'Flavanoids', ... 'Nonflavanoid phenols', ... 'Proanthocyanins', ... 'Color intensity', 'Hue', ... 'OD280/OD315 of diluted wines', ... 'Proline'] >>> print('Class labels', np.unique(df_wine['Class label'])) Class labels [1 2 3] >>> df_wine.head()

[ 122 ]

Chapter 4

The 13 different features in the Wine dataset, describing the chemical properties of the 178 wine examples, are listed in the following table:

The examples belong to one of three different classes, 1, 2, and 3, which refer to the three different types of grape grown in the same region in Italy but derived from different wine cultivars, as described in the dataset summary (https://archive. ics.uci.edu/ml/machine-learning-databases/wine/wine.names). A convenient way to randomly partition this dataset into separate test and training datasets is to use the train_test_split function from scikit-learn's model_selection submodule: >>> from sklearn.model_selection import train_test_split >>> X, y = df_wine.iloc[:, 1:].values, df_wine.iloc[:, 0].values >>> X_train, X_test, y_train, y_test =\ ... train_test_split(X, y, ... test_size=0.3, ... random_state=0, ... stratify=y)

First, we assigned the NumPy array representation of the feature columns 1-13 to the variable X and we assigned the class labels from the first column to the variable y. Then, we used the train_test_split function to randomly split X and y into separate training and test datasets. By setting test_size=0.3, we assigned 30 percent of the wine examples to X_test and y_test, and the remaining 70 percent of the examples were assigned to X_train and y_train, respectively. Providing the class label array y as an argument to stratify ensures that both training and test datasets have the same class proportions as the original dataset.

[ 123 ]

Building Good Training Datasets – Data Preprocessing

Choosing an appropriate ratio for partitioning a dataset into training and test datasets If we are dividing a dataset into training and test datasets, we have to keep in mind that we are withholding valuable information that the learning algorithm could benefit from. Thus, we don't want to allocate too much information to the test set. However, the smaller the test set, the more inaccurate the estimation of the generalization error. Dividing a dataset into training and test datasets is all about balancing this tradeoff. In practice, the most commonly used splits are 60:40, 70:30, or 80:20, depending on the size of the initial dataset. However, for large datasets, 90:10 or 99:1 splits are also common and appropriate. For example, if the dataset contains more than 100,000 training examples, it might be fine to withhold only 10,000 examples for testing in order to get a good estimate of the generalization performance. More information and illustrations can be found in section one of my article Model evaluation, model selection, and algorithm selection in machine learning, which is freely available at https://arxiv.org/pdf/1811.12808.pdf. Moreover, instead of discarding the allocated test data after model training and evaluation, it is a common practice to retrain a classifier on the entire dataset, as it can improve the predictive performance of the model. While this approach is generally recommended, it could lead to worse generalization performance if the dataset is small and the test dataset contains outliers, for example. Also, after refitting the model on the whole dataset, we don't have any independent data left to evaluate its performance.

Bringing features onto the same scale

Feature scaling is a crucial step in our preprocessing pipeline that can easily be forgotten. Decision trees and random forests are two of the very few machine learning algorithms where we don't need to worry about feature scaling. Those algorithms are scale invariant. However, the majority of machine learning and optimization algorithms behave much better if features are on the same scale, as we saw in Chapter 2, Training Simple Machine Learning Algorithms for Classification, when we implemented the gradient descent optimization algorithm. The importance of feature scaling can be illustrated by a simple example. Let's assume that we have two features where one feature is measured on a scale from 1 to 10 and the second feature is measured on a scale from 1 to 100,000, respectively. [ 124 ]

Chapter 4

When we think of the squared error function in Adaline from Chapter 2, Training Simple Machine Learning Algorithms for Classification, it makes sense to say that the algorithm will mostly be busy optimizing the weights according to the larger errors in the second feature. Another example is the k-nearest neighbors (KNN) algorithm with a Euclidean distance measure: the computed distances between examples will be dominated by the second feature axis. Now, there are two common approaches to bringing different features onto the same scale: normalization and standardization. Those terms are often used quite loosely in different fields, and the meaning has to be derived from the context. Most often, normalization refers to the rescaling of the features to a range of [0, 1], which is a special case of min-max scaling. To normalize our data, we can simply apply the (𝑖𝑖) , of an example, min-max scaling to each feature column, where the new value, 𝑥𝑥𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 𝑥𝑥 (𝑖𝑖) , can be calculated as follows: (𝑖𝑖)

𝑥𝑥𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 =

𝑥𝑥 (𝑖𝑖) − 𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 − 𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚

Here, 𝑥𝑥 (𝑖𝑖) is a particular example, 𝑥𝑥𝑚𝑚𝑚𝑚𝑛𝑛 is the smallest value in a feature column, and 𝑥𝑥𝑚𝑚𝑚𝑚𝑚𝑚 is the largest value.

The min-max scaling procedure is implemented in scikit-learn and can be used as follows: >>> >>> >>> >>>

from sklearn.preprocessing import MinMaxScaler mms = MinMaxScaler() X_train_norm = mms.fit_transform(X_train) X_test_norm = mms.transform(X_test)

Although normalization via min-max scaling is a commonly used technique that is useful when we need values in a bounded interval, standardization can be more practical for many machine learning algorithms, especially for optimization algorithms such as gradient descent. The reason is that many linear models, such as the logistic regression and SVM from Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, initialize the weights to 0 or small random values close to 0. Using standardization, we center the feature columns at mean 0 with standard deviation 1 so that the feature columns have the same parameters as a standard normal distribution (zero mean and unit variance), which makes it easier to learn the weights. Furthermore, standardization maintains useful information about outliers and makes the algorithm less sensitive to them in contrast to min-max scaling, which scales the data to a limited range of values.

[ 125 ]

Building Good Training Datasets – Data Preprocessing

The procedure for standardization can be expressed by the following equation: (𝑖𝑖)

𝑥𝑥𝑠𝑠𝑠𝑠𝑠𝑠 =

𝑥𝑥 (𝑖𝑖) − 𝜇𝜇𝑥𝑥 𝜎𝜎𝑥𝑥

Here, 𝜇𝜇𝑥𝑥 is the sample mean of a particular feature column, and 𝜎𝜎𝑥𝑥 is the corresponding standard deviation.

The following table illustrates the difference between the two commonly used feature scaling techniques, standardization and normalization, on a simple example dataset consisting of numbers 0 to 5: Input

Standardized

Min-max normalized

0.0

-1.46385

0.0

1.0

-0.87831

0.2

2.0

-0.29277

0.4

3.0

0.29277

0.6

4.0

0.87831

0.8

5.0

1.46385

1.0

You can perform the standardization and normalization shown in the table manually by executing the following code examples: >>> ex = np.array([0, 1, 2, 3, 4, 5]) >>> print('standardized:', (ex - ex.mean()) / ex.std()) standardized: [-1.46385011 -0.87831007 -0.29277002 0.29277002 0.87831007 1.46385011] >>> print('normalized:', (ex - ex.min()) / (ex.max() - ex.min())) normalized: [ 0. 0.2 0.4 0.6 0.8 1. ]

Similar to the MinMaxScaler class, scikit-learn also implements a class for standardization: >>> >>> >>> >>>

from sklearn.preprocessing import StandardScaler stdsc = StandardScaler() X_train_std = stdsc.fit_transform(X_train) X_test_std = stdsc.transform(X_test)

[ 126 ]

Chapter 4

Again, it is also important to highlight that we fit the StandardScaler class only once—on the training data—and use those parameters to transform the test dataset or any new data point. Other, more advanced methods for feature scaling are available from scikitlearn, such as the RobustScaler. The RobustScaler is especially helpful and recommended if we are working with small datasets that contain many outliers. Similarly, if the machine learning algorithm applied to this dataset is prone to overfitting, the RobustScaler can be a good choice. Operating on each feature column independently, the RobustScaler removes the median value and scales the dataset according to the 1st and 3rd quartile of the dataset (that is, the 25th and 75th quantile, respectively) such that more extreme values and outliers become less pronounced. The interested reader can find more information about the RobustScaler in the official scikit-learn documentation at https://scikit-learn. org/stable/modules/generated/sklearn.preprocessing.RobustScaler.html.

Selecting meaningful features

If we notice that a model performs much better on a training dataset than on the test dataset, this observation is a strong indicator of overfitting. As we discussed in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, overfitting means the model fits the parameters too closely with regard to the particular observations in the training dataset, but does not generalize well to new data; we say that the model has a high variance. The reason for the overfitting is that our model is too complex for the given training data. Common solutions to reduce the generalization error are as follows: • Collect more training data • Introduce a penalty for complexity via regularization • Choose a simpler model with fewer parameters • Reduce the dimensionality of the data Collecting more training data is often not applicable. In Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, we will learn about a useful technique to check whether more training data is helpful. In the following sections, we will look at common ways to reduce overfitting by regularization and dimensionality reduction via feature selection, which leads to simpler models by requiring fewer parameters to be fitted to the data.

[ 127 ]

Building Good Training Datasets – Data Preprocessing

L1 and L2 regularization as penalties against model complexity

You will recall from Chapter 3, A Tour of Machine Learning Classifiers Using scikitlearn, that L2 regularization is one approach to reduce the complexity of a model by penalizing large individual weights. We defined the squared L2 norm of our weight vector, w, as follows:

𝐿𝐿2:

‖𝒘𝒘‖22

𝑚𝑚

= ∑ 𝑤𝑤𝑗𝑗2 𝑗𝑗=1

Another approach to reduce the model complexity is the related L1 regularization:

𝐿𝐿1:

𝑚𝑚

‖𝒘𝒘‖1 = ∑|𝑤𝑤𝑗𝑗 | 𝑗𝑗=1

Here, we simply replaced the square of the weights by the sum of the absolute values of the weights. In contrast to L2 regularization, L1 regularization usually yields sparse feature vectors and most feature weights will be zero. Sparsity can be useful in practice if we have a high-dimensional dataset with many features that are irrelevant, especially in cases where we have more irrelevant dimensions than training examples. In this sense, L1 regularization can be understood as a technique for feature selection.

A geometric interpretation of L2 regularization As mentioned in the previous section, L2 regularization adds a penalty term to the cost function that effectively results in less extreme weight values compared to a model trained with an unregularized cost function.

[ 128 ]

Chapter 4

To better understand how L1 regularization encourages sparsity, let's take a step back and take a look at a geometric interpretation of regularization. Let's plot the contours of a convex cost function for two weight coefficients, 𝑤𝑤1 and 𝑤𝑤2.

Here, we will consider the sum of squared errors (SSE) cost function that we used for Adaline in Chapter 2, Training Simple Machine Learning Algorithms for Classification, since it is spherical and easier to draw than the cost function of logistic regression; however, the same concepts apply. Remember that our goal is to find the combination of weight coefficients that minimize the cost function for the training data, as shown in the following figure (the point in the center of the ellipses):

We can think of regularization as adding a penalty term to the cost function to encourage smaller weights; in other words, we penalize large weights. Thus, by increasing the regularization strength via the regularization parameter, 𝜆𝜆 , we shrink the weights toward zero and decrease the dependence of our model on the training data. Let's illustrate this concept in the following figure for the L2 penalty term:

[ 129 ]

Building Good Training Datasets – Data Preprocessing

The quadratic L2 regularization term is represented by the shaded ball. Here, our weight coefficients cannot exceed our regularization budget—the combination of the weight coefficients cannot fall outside the shaded area. On the other hand, we still want to minimize the cost function. Under the penalty constraint, our best effort is to choose the point where the L2 ball intersects with the contours of the unpenalized cost function. The larger the value of the regularization parameter, 𝜆𝜆, gets, the faster the penalized cost grows, which leads to a narrower L2 ball. For example, if we increase the regularization parameter towards infinity, the weight coefficients will become effectively zero, denoted by the center of the L2 ball. To summarize the main message of the example, our goal is to minimize the sum of the unpenalized cost plus the penalty term, which can be understood as adding bias and preferring a simpler model to reduce the variance in the absence of sufficient training data to fit the model.

[ 130 ]

Chapter 4

Sparse solutions with L1 regularization

Now, let's discuss L1 regularization and sparsity. The main concept behind L1 regularization is similar to what we discussed in the previous section. However, since the L1 penalty is the sum of the absolute weight coefficients (remember that the L2 term is quadratic), we can represent it as a diamond-shape budget, as shown in the following figure:

In the preceding figure, we can see that the contour of the cost function touches the L1 diamond at 𝑤𝑤1 = 0. Since the contours of an L1 regularized system are sharp, it is more likely that the optimum—that is, the intersection between the ellipses of the cost function and the boundary of the L1 diamond—is located on the axes, which encourages sparsity. L1 regularization and sparsity The mathematical details of why L1 regularization can lead to sparse solutions are beyond the scope of this book. If you are interested, an excellent explanation of L2 versus L1 regularization can be found in Section 3.4, The Elements of Statistical Learning, Trevor Hastie, Robert Tibshirani, and Jerome Friedman, Springer Science+Business Media, 2009.

[ 131 ]

Building Good Training Datasets – Data Preprocessing

For regularized models in scikit-learn that support L1 regularization, we can simply set the penalty parameter to 'l1' to obtain a sparse solution: >>> from sklearn.linear_model import LogisticRegression >>> LogisticRegression(penalty='l1', ... solver='liblinear', ... multi_class='ovr')

Note that we also need to select a different optimization algorithm (for example, solver='liblinear'), since 'lbfgs' currently does not support L1-regularized loss optimization. Applied to the standardized Wine data, the L1 regularized logistic regression would yield the following sparse solution: >>> lr = LogisticRegression(penalty='l1', ... C=1.0, ... solver='liblinear', ... multi_class='ovr') # Note that C=1.0 is the default. You can increase # or decrease it to make the regularization effect # stronger or weaker, respectively. >>> lr.fit(X_train_std, y_train) >>> print('Training accuracy:', lr.score(X_train_std, y_train)) Training accuracy: 1.0 >>> print('Test accuracy:', lr.score(X_test_std, y_test)) Test accuracy: 1.0

Both training and test accuracies (both 100 percent) indicate that our model does a perfect job on both datasets. When we access the intercept terms via the lr.intercept_ attribute, we can see that the array returns three values: >>> lr.intercept_ array([-1.26346036, -1.21584018, -2.3697841 ])

Since we fit the LogisticRegression object on a multiclass dataset via the one-vs.rest (OvR) approach, the first intercept belongs to the model that fits class 1 versus classes 2 and 3, the second value is the intercept of the model that fits class 2 versus classes 1 and 3, and the third value is the intercept of the model that fits class 3 versus classes 1 and 2: >>> lr.coef_ array([[ 1.24590762, 0.18070219, 0.74375939, -1.16141503, 0. , 0. , 1.16926815, 0. , 0. , 0. , 0. , 0.54784923, 2.51028042], [-1.53680415, -0.38795309, -0.99494046, 0.36508729, -0.05981561, 0. , 0.6681573 , 0. , 0. , -1.93426485, 1.23265994, 0. , [ 132 ]

Chapter 4 -2.23137595], [ 0.13547047, 0.16873019, 0.35728003, 0. , 0. , 0. , -2.43713947, 0. , 0. , 1.56351492, -0.81894749, -0.49308407, 0. ]])

The weight array that we accessed via the lr.coef_ attribute contains three rows of weight coefficients, one weight vector for each class. Each row consists of 13 weights, where each weight is multiplied by the respective feature in the 13-dimensional Wine dataset to calculate the net input: 𝑚𝑚

𝑧𝑧 = 𝑤𝑤0 𝑥𝑥0 + ⋯ + 𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚 = ∑ 𝑥𝑥𝑗𝑗 𝑤𝑤𝑗𝑗 = 𝒘𝒘𝑇𝑇 𝒙𝒙 𝑗𝑗=0

Accessing the bias and weight parameters of scikit-learn estimators In scikit-learn, the intercept_ corresponds to 𝑤𝑤0 and coef_ corresponds to the values 𝑤𝑤𝑗𝑗 for j > 0.

As a result of L1 regularization, which, as mentioned, serves as a method for feature selection, we just trained a model that is robust to the potentially irrelevant features in this dataset. Strictly speaking, though, the weight vectors from the previous example are not necessarily sparse because they contain more non-zero than zero entries. However, we could enforce sparsity (more zero entries) by further increasing the regularization strength—that is, choosing lower values for the C parameter. In the last example on regularization in this chapter, we will vary the regularization strength and plot the regularization path—the weight coefficients of the different features for different regularization strengths: >>> import matplotlib.pyplot as plt >>> fig = plt.figure() >>> ax = plt.subplot(111) >>> colors = ['blue', 'green', 'red', 'cyan', ... 'magenta', 'yellow', 'black', ... 'pink', 'lightgreen', 'lightblue', ... 'gray', 'indigo', 'orange'] >>> weights, params = [], [] >>> for c in np.arange(-4., 6.):

[ 133 ]

Building Good Training Datasets – Data Preprocessing ... ... ... ... ... ...

lr = LogisticRegression(penalty='l1', C=10.**c, solver='liblinear', multi_class='ovr', random_state=0) lr.fit(X_train_std, y_train) weights.append(lr.coef_[1]) params.append(10**c)

>>> weights = np.array(weights) >>> ... ... ... >>> >>> >>> >>> >>> >>> >>> ... ... >>>

for column, color in zip(range(weights.shape[1]), colors): plt.plot(params, weights[:, column], label=df_wine.columns[column + 1], color=color) plt.axhline(0, color='black', linestyle='--', linewidth=3) plt.xlim([10**(-5), 10**5]) plt.ylabel('weight coefficient') plt.xlabel('C') plt.xscale('log') plt.legend(loc='upper left') ax.legend(loc='upper center', bbox_to_anchor=(1.38, 1.03), ncol=1, fancybox=True) plt.show()

The resulting plot provides us with further insights into the behavior of L1 regularization. As we can see, all feature weights will be zero if we penalize the model with a strong regularization parameter (C < 0.01); C is the inverse of the regularization parameter, 𝜆𝜆:

[ 134 ]

Chapter 4

Sequential feature selection algorithms

An alternative way to reduce the complexity of the model and avoid overfitting is dimensionality reduction via feature selection, which is especially useful for unregularized models. There are two main categories of dimensionality reduction techniques: feature selection and feature extraction. Via feature selection, we select a subset of the original features, whereas in feature extraction, we derive information from the feature set to construct a new feature subspace. In this section, we will take a look at a classic family of feature selection algorithms. In the next chapter, Chapter 5, Compressing Data via Dimensionality Reduction, we will learn about different feature extraction techniques to compress a dataset onto a lower-dimensional feature subspace. Sequential feature selection algorithms are a family of greedy search algorithms that are used to reduce an initial d-dimensional feature space to a k-dimensional feature subspace where k self.k_features: scores = [] subsets = [] for p in combinations(self.indices_, r=dim - 1): score = self._calc_score(X_train, y_train, X_test, y_test, p) scores.append(score) subsets.append(p) best = np.argmax(scores) self.indices_ = subsets[best] self.subsets_.append(self.indices_) dim -= 1 self.scores_.append(scores[best]) self.k_score_ = self.scores_[-1] return self def transform(self, X): return X[:, self.indices_] def _calc_score(self, X_train, y_train, X_test, y_test, indices): self.estimator.fit(X_train[:, indices], y_train) y_pred = self.estimator.predict(X_test[:, indices]) score = self.scoring(y_test, y_pred) return score

In the preceding implementation, we defined the k_features parameter to specify the desired number of features we want to return. By default, we use the accuracy_score from scikit-learn to evaluate the performance of a model (an estimator for classification) on the feature subsets. [ 137 ]

Building Good Training Datasets – Data Preprocessing

Inside the while loop of the fit method, the feature subsets created by the itertools.combination function are evaluated and reduced until the feature subset has the desired dimensionality. In each iteration, the accuracy score of the best subset is collected in a list, self.scores_, based on the internally created test dataset, X_test. We will use those scores later to evaluate the results. The column indices of the final feature subset are assigned to self.indices_, which we can use via the transform method to return a new data array with the selected feature columns. Note that, instead of calculating the criterion explicitly inside the fit method, we simply removed the feature that is not contained in the best performing feature subset. Now, let's see our SBS implementation in action using the KNN classifier from scikit-learn: >>> import matplotlib.pyplot as plt >>> from sklearn.neighbors import KNeighborsClassifier >>> knn = KNeighborsClassifier(n_neighbors=5) >>> sbs = SBS(knn, k_features=1) >>> sbs.fit(X_train_std, y_train)

Although our SBS implementation already splits the dataset into a test and training dataset inside the fit function, we still fed the training dataset, X_train, to the algorithm. The SBS fit method will then create new training subsets for testing (validation) and training, which is why this test set is also called the validation dataset. This approach is necessary to prevent our original test set from becoming part of the training data. Remember that our SBS algorithm collects the scores of the best feature subset at each stage, so let's move on to the more exciting part of our implementation and plot the classification accuracy of the KNN classifier that was calculated on the validation dataset. The code is as follows: >>> k_feat = [len(k) for k in sbs.subsets_] >>> >>> >>> >>> >>> >>> >>>

plt.plot(k_feat, sbs.scores_, marker='o') plt.ylim([0.7, 1.02]) plt.ylabel('Accuracy') plt.xlabel('Number of features') plt.grid() plt.tight_layout() plt.show()

[ 138 ]

Chapter 4

As we can see in the following figure, the accuracy of the KNN classifier improved on the validation dataset as we reduced the number of features, which is likely due to a decrease in the curse of dimensionality that we discussed in the context of the KNN algorithm in Chapter 3, A Tour of Machine Learning Classifiers Using scikitlearn. Also, we can see in the following plot that the classifier achieved 100 percent accuracy for k={3, 7, 8, 9, 10, 11, 12}:

To satisfy our own curiosity, let's see what the smallest feature subset (k=3), which yielded such a good performance on the validation dataset, looks like: >>> k3 = list(sbs.subsets_[10]) >>> print(df_wine.columns[1:][k3]) Index(['Alcohol', 'Malic acid', 'OD280/OD315 of diluted wines'], dtype='object')

Using the preceding code, we obtained the column indices of the three-feature subset from the 11th position in the sbs.subsets_ attribute and returned the corresponding feature names from the column index of the pandas Wine DataFrame. Next, let's evaluate the performance of the KNN classifier on the original test dataset: >>> knn.fit(X_train_std, y_train) >>> print('Training accuracy:', knn.score(X_train_std, y_train)) Training accuracy: 0.967741935484 >>> print('Test accuracy:', knn.score(X_test_std, y_test)) Test accuracy: 0.962962962963 [ 139 ]

Building Good Training Datasets – Data Preprocessing

In the preceding code section, we used the complete feature set and obtained approximately 97 percent accuracy on the training dataset and approximately 96 percent accuracy on the test dataset, which indicates that our model already generalizes well to new data. Now, let's use the selected three-feature subset and see how well KNN performs: >>> knn.fit(X_train_std[:, k3], y_train) >>> print('Training accuracy:', ... knn.score(X_train_std[:, k3], y_train)) Training accuracy: 0.951612903226 >>> print('Test accuracy:', ... knn.score(X_test_std[:, k3], y_test)) Test accuracy: 0.925925925926

When using less than a quarter of the original features in the Wine dataset, the prediction accuracy on the test dataset declined slightly. This may indicate that those three features do not provide less discriminatory information than the original dataset. However, we also have to keep in mind that the Wine dataset is a small dataset and is very susceptible to randomness—that is, the way we split the dataset into training and test subsets, and how we split the training dataset further into a training and validation subset. While we did not increase the performance of the KNN model by reducing the number of features, we shrank the size of the dataset, which can be useful in realworld applications that may involve expensive data collection steps. Also, by substantially reducing the number of features, we obtain simpler models, which are easier to interpret. Feature selection algorithms in scikit-learn There are many more feature selection algorithms available via scikit-learn. Those include recursive backward elimination based on feature weights, tree-based methods to select features by importance, and univariate statistical tests. A comprehensive discussion of the different feature selection methods is beyond the scope of this book, but a good summary with illustrative examples can be found at http://scikit-learn.org/ stable/modules/feature_selection.html. You can find implementations of several different flavors of sequential feature selection related to the simple SBS that we implemented previously in the Python package mlxtend at http://rasbt. github.io/mlxtend/user_guide/feature_selection/ SequentialFeatureSelector/.

[ 140 ]

Chapter 4

Assessing feature importance with random forests

In previous sections, you learned how to use L1 regularization to zero out irrelevant features via logistic regression and how to use the SBS algorithm for feature selection and apply it to a KNN algorithm. Another useful approach for selecting relevant features from a dataset is using a random forest, an ensemble technique that was introduced in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn. Using a random forest, we can measure the feature importance as the averaged impurity decrease computed from all decision trees in the forest, without making any assumptions about whether our data is linearly separable or not. Conveniently, the random forest implementation in scikit-learn already collects the feature importance values for us so that we can access them via the feature_importances_ attribute after fitting a RandomForestClassifier. By executing the following code, we will now train a forest of 500 trees on the Wine dataset and rank the 13 features by their respective importance measures—remember from our discussion in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn that we don't need to use standardized or normalized features in tree-based models: >>> from sklearn.ensemble import RandomForestClassifier >>> feat_labels = df_wine.columns[1:] >>> forest = RandomForestClassifier(n_estimators=500, ... random_state=1) >>> forest.fit(X_train, y_train) >>> importances = forest.feature_importances_ >>> indices = np.argsort(importances)[::-1] >>> for f in range(X_train.shape[1]): ... print("%2d) %-*s %f" % (f + 1, 30, ... feat_labels[indices[f]], ... importances[indices[f]])) >>> plt.title('Feature Importance') >>> plt.bar(range(X_train.shape[1]), ... importances[indices], ... align='center') >>> plt.xticks(range(X_train.shape[1]), ... feat_labels[indices] rotation=90) >>> plt.xlim([-1, X_train.shape[1]]) >>> plt.tight_layout() [ 141 ]

Building Good Training Datasets – Data Preprocessing >>> plt.show() 1) Proline 2) Flavanoids 3) Color intensity 4) OD280/OD315 of diluted wines 5) Alcohol 6) Hue 7) Total phenols 8) Magnesium 9) Malic acid 10) Proanthocyanins 11) Alcalinity of ash 12) Nonflavanoid phenols 13) Ash

0.185453 0.174751 0.143920 0.136162 0.118529 0.058739 0.050872 0.031357 0.025648 0.025570 0.022366 0.013354 0.013279

After executing the code, we created a plot that ranks the different features in the Wine dataset by their relative importance; note that the feature importance values are normalized so that they sum up to 1.0:

We can conclude that the proline and flavonoid levels, the color intensity, the OD280/OD315 diffraction, and the alcohol concentration of wine are the most discriminative features in the dataset based on the average impurity decrease in the 500 decision trees. Interestingly, two of the top-ranked features in the plot are also in the three-feature subset selection from the SBS algorithm that we implemented in the previous section (alcohol concentration and OD280/OD315 of diluted wines).

[ 142 ]

Chapter 4

However, as far as interpretability is concerned, the random forest technique comes with an important gotcha that is worth mentioning. If two or more features are highly correlated, one feature may be ranked very highly while the information on the other feature(s) may not be fully captured. On the other hand, we don't need to be concerned about this problem if we are merely interested in the predictive performance of a model rather than the interpretation of feature importance values. To conclude this section about feature importance values and random forests, it is worth mentioning that scikit-learn also implements a SelectFromModel object that selects features based on a user-specified threshold after model fitting, which is useful if we want to use the RandomForestClassifier as a feature selector and intermediate step in a scikit-learn Pipeline object, which allows us to connect different preprocessing steps with an estimator, as you will see in Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning. For example, we could set the threshold to 0.1 to reduce the dataset to the five most important features using the following code: >>> from sklearn.feature_selection import SelectFromModel >>> sfm = SelectFromModel(forest, threshold=0.1, prefit=True) >>> X_selected = sfm.transform(X_train) >>> print('Number of features that meet this threshold', ... 'criterion:', X_selected.shape[1]) Number of features that meet this threshold criterion: 5 >>> ... ... ... 1) 2) 3) 4) 5)

for f in range(X_selected.shape[1]): print("%2d) %-*s %f" % (f + 1, 30, feat_labels[indices[f]], importances[indices[f]])) Proline 0.185453 Flavanoids 0.174751 Color intensity 0.143920 OD280/OD315 of diluted wines 0.136162 Alcohol 0.118529

Summary

We started this chapter by looking at useful techniques to make sure that we handle missing data correctly. Before we feed data to a machine learning algorithm, we also have to make sure that we encode categorical variables correctly, and in this chapter, we saw how we can map ordinal and nominal feature values to integer representations. [ 143 ]

Building Good Training Datasets – Data Preprocessing

Moreover, we briefly discussed L1 regularization, which can help us to avoid overfitting by reducing the complexity of a model. As an alternative approach to removing irrelevant features, we used a sequential feature selection algorithm to select meaningful features from a dataset. In the next chapter, you will learn about yet another useful approach to dimensionality reduction: feature extraction. It allows us to compress features onto a lower-dimensional subspace, rather than removing features entirely as in feature selection.

[ 144 ]

Compressing Data via Dimensionality Reduction In Chapter 4, Building Good Training Datasets – Data Preprocessing, you learned about the different approaches for reducing the dimensionality of a dataset using different feature selection techniques. An alternative approach to feature selection for dimensionality reduction is feature extraction. In this chapter, you will learn about three fundamental techniques that will help you to summarize the information content of a dataset by transforming it onto a new feature subspace of lower dimensionality than the original one. Data compression is an important topic in machine learning, and it helps us to store and analyze the increasing amounts of data that are produced and collected in the modern age of technology. In this chapter, we will cover the following topics: • Principal component analysis (PCA) for unsupervised data compression • Linear discriminant analysis (LDA) as a supervised dimensionality reduction technique for maximizing class separability • Nonlinear dimensionality reduction via kernel principal component analysis (KPCA)

Unsupervised dimensionality reduction via principal component analysis

Similar to feature selection, we can use different feature extraction techniques to reduce the number of features in a dataset. The difference between feature selection and feature extraction is that while we maintain the original features when we use feature selection algorithms, such as sequential backward selection, we use feature extraction to transform or project the data onto a new feature space.

[ 145 ]

Compressing Data via Dimensionality Reduction

In the context of dimensionality reduction, feature extraction can be understood as an approach to data compression with the goal of maintaining most of the relevant information. In practice, feature extraction is not only used to improve storage space or the computational efficiency of the learning algorithm, but can also improve the predictive performance by reducing the curse of dimensionality—especially if we are working with non-regularized models.

The main steps behind principal component analysis

In this section, we will discuss PCA, an unsupervised linear transformation technique that is widely used across different fields, most prominently for feature extraction and dimensionality reduction. Other popular applications of PCA include exploratory data analyses and the denoising of signals in stock market trading, and the analysis of genome data and gene expression levels in the field of bioinformatics. PCA helps us to identify patterns in data based on the correlation between features. In a nutshell, PCA aims to find the directions of maximum variance in highdimensional data and projects the data onto a new subspace with equal or fewer dimensions than the original one. The orthogonal axes (principal components) of the new subspace can be interpreted as the directions of maximum variance given the constraint that the new feature axes are orthogonal to each other, as illustrated in the following figure:

In the preceding figure, 𝑥𝑥1 and 𝑥𝑥2 are the original feature axes, and PC1 and PC2 are the principal components. [ 146 ]

Chapter 5

If we use PCA for dimensionality reduction, we construct a 𝑑𝑑 × 𝑘𝑘-dimensional transformation matrix, W, that allows us to map a vector, x, the features of a training example, onto a new k-dimensional feature subspace that has fewer dimensions than the original d-dimensional feature space. For instance, the process is as follows. Suppose we have a feature vector, x:

𝒙𝒙 = [𝑥𝑥1 , 𝑥𝑥2 , … , 𝑥𝑥𝑑𝑑 ],

𝒙𝒙 ∈ ℝ𝑑𝑑

which is then transformed by a transformation matrix, 𝑾𝑾 ∈ ℝ𝑑𝑑×𝑘𝑘 :

resulting in the output vector:

𝒙𝒙𝒙𝒙 = 𝒛𝒛

𝒛𝒛 = [𝑧𝑧1 , 𝑧𝑧2 , … , 𝑧𝑧𝑘𝑘 ],

𝒛𝒛 ∈ ℝ𝑘𝑘

As a result of transforming the original d-dimensional data onto this new k-dimensional subspace (typically k >> >>> ... >>> >>>

# Make a list of (eigenvalue, eigenvector) tuples eigen_pairs = [(np.abs(eigen_vals[i]), eigen_vecs[:, i]) for i in range(len(eigen_vals))] # Sort the (eigenvalue, eigenvector) tuples from high to low eigen_pairs.sort(key=lambda k: k[0], reverse=True)

Next, we collect the two eigenvectors that correspond to the two largest eigenvalues, to capture about 60 percent of the variance in this dataset. Note that two eigenvectors have been chosen for the purpose of illustration, since we are going to plot the data via a two-dimensional scatter plot later in this subsection. In practice, the number of principal components has to be determined by a tradeoff between computational efficiency and the performance of the classifier: >>> w = np.hstack((eigen_pairs[0][1][:, np.newaxis], ... eigen_pairs[1][1][:, np.newaxis])) >>> print('Matrix W:\n', w) Matrix W: [[-0.13724218 0.50303478] [ 0.24724326 0.16487119] [-0.02545159 0.24456476] [ 0.20694508 -0.11352904] [-0.15436582 0.28974518] [-0.39376952 0.05080104] [-0.41735106 -0.02287338] [ 0.30572896 0.09048885] [-0.30668347 0.00835233] [ 0.07554066 0.54977581] [-0.32613263 -0.20716433] [-0.36861022 -0.24902536] [-0.29669651 0.38022942]]

By executing the preceding code, we have created a 13 × 2-dimensional projection matrix, W, from the top two eigenvectors.

[ 153 ]

Compressing Data via Dimensionality Reduction

Mirrored projections Depending on which versions of NumPy and LAPACK you are using, you may obtain the matrix, W, with its signs flipped. Please note that this is not an issue; if v is an eigenvector of a matrix, Σ, we have:

Σ𝒗𝒗 = 𝜆𝜆𝒗𝒗

Here, v is the eigenvector, and –v is also an eigenvector, which we can show as follows. Using basic algebra, we can multiply both sides of the equation by a scalar, 𝛼𝛼:

𝛼𝛼Σ𝒗𝒗 = 𝛼𝛼𝛼𝛼𝒗𝒗

Since matrix multiplication is associative for scalar multiplication, we can then rearrange this to the following:

Σ(𝛼𝛼𝒗𝒗) = 𝜆𝜆(𝛼𝛼𝒗𝒗)

Now, we can see that 𝛼𝛼𝒗𝒗 is an eigenvector with the same eigenvalue, 𝜆𝜆, for both 𝛼𝛼 = 1 and 𝛼𝛼 = −1. Hence, both v and –v are eigenvectors.

Using the projection matrix, we can now transform an example, x (represented as a 13-dimensional row vector), onto the PCA subspace (the principal components one and two) obtaining 𝒙𝒙′, now a two-dimensional example vector consisting of two new features:

𝒙𝒙′ = 𝒙𝒙𝒙𝒙

>>> X_train_std[0].dot(w) array([ 2.38299011, 0.45458499])

Similarly, we can transform the entire 124 × 13-dimensional training dataset onto the two principal components by calculating the matrix dot product:

𝑿𝑿′ = 𝑿𝑿𝑿𝑿

>>> X_train_pca = X_train_std.dot(w)

Lastly, let's visualize the transformed Wine training dataset, now stored as an 124 × 2 -dimensional matrix, in a two-dimensional scatterplot: [ 154 ]

Chapter 5 >>> >>> >>> ... ... ... >>> >>> >>> >>> >>>

colors = ['r', 'b', 'g'] markers = ['s', 'x', 'o'] for l, c, m in zip(np.unique(y_train), colors, markers): plt.scatter(X_train_pca[y_train==l, 0], X_train_pca[y_train==l, 1], c=c, label=l, marker=m) plt.xlabel('PC 1') plt.ylabel('PC 2') plt.legend(loc='lower left') plt.tight_layout() plt.show()

As we can see in the resulting plot, the data is more spread along the x-axis—the first principal component—than the second principal component (y-axis), which is consistent with the explained variance ratio plot that we created in the previous subsection. However, we can tell that a linear classifier will likely be able to separate the classes well:

Although we encoded the class label information for the purpose of illustration in the preceding scatter plot, we have to keep in mind that PCA is an unsupervised technique that doesn't use any class label information.

Principal component analysis in scikit-learn

Although the verbose approach in the previous subsection helped us to follow the inner workings of PCA, we will now discuss how to use the PCA class implemented in scikit-learn. [ 155 ]

Compressing Data via Dimensionality Reduction

The PCA class is another one of scikit-learn's transformer classes, where we first fit the model using the training data before we transform both the training data and the test dataset using the same model parameters. Now, let's use the PCA class from scikitlearn on the Wine training dataset, classify the transformed examples via logistic regression, and visualize the decision regions via the plot_decision_regions function that we defined in Chapter 2, Training Simple Machine Learning Algorithms for Classification: from matplotlib.colors import ListedColormap def plot_decision_regions(X, y, classifier, resolution=0.02): # setup marker generator and color map markers = ('s', 'x', 'o', '^', 'v') colors = ('red', 'blue', 'lightgreen', 'gray', 'cyan') cmap = ListedColormap(colors[:len(np.unique(y))]) # plot the decision surface x1_min, x1_max = X[:, 0].min() - 1, X[:, 0].max() + 1 x2_min, x2_max = X[:, 1].min() - 1, X[:, 1].max() + 1 xx1, xx2 = np.meshgrid(np.arange(x1_min, x1_max, resolution), np.arange(x2_min, x2_max, resolution)) Z = classifier.predict(np.array([xx1.ravel(), xx2.ravel()]).T) Z = Z.reshape(xx1.shape) plt.contourf(xx1, xx2, Z, alpha=0.4, cmap=cmap) plt.xlim(xx1.min(), xx1.max()) plt.ylim(xx2.min(), xx2.max()) # plot examples by class for idx, cl in enumerate(np.unique(y)): plt.scatter(x=X[y == cl, 0], y=X[y == cl, 1], alpha=0.6, color=cmap(idx), edgecolor='black', marker=markers[idx], label=cl)

For your convenience, you can place the plot_decision_regions code shown above into a separate code file in your current working directory, for example, plot_decision_regions_script.py, and import it into your current Python session. >>> >>> >>> >>>

from sklearn.linear_model import LogisticRegression from sklearn.decomposition import PCA # initializing the PCA transformer and # logistic regression estimator: [ 156 ]

Chapter 5 >>> >>> ... ... >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>>

pca = PCA(n_components=2) lr = LogisticRegression(multi_class='ovr', random_state=1, solver='lbfgs') # dimensionality reduction: X_train_pca = pca.fit_transform(X_train_std) X_test_pca = pca.transform(X_test_std) # fitting the logistic regression model on the reduced dataset: lr.fit(X_train_pca, y_train) plot_decision_regions(X_train_pca, y_train, classifier=lr) plt.xlabel('PC 1') plt.ylabel('PC 2') plt.legend(loc='lower left') plt.tight_layout() plt.show()

By executing the preceding code, we should now see the decision regions for the training data reduced to two principal component axes:

When we compare PCA projections via scikit-learn with our own PCA implementation, it can happen that the resulting plots are mirror images of each other. Note that this is not due to an error in either of those two implementations; the reason for this difference is that, depending on the eigensolver, eigenvectors can have either negative or positive signs.

[ 157 ]

Compressing Data via Dimensionality Reduction

Not that it matters, but we could simply revert the mirror image by multiplying the data by –1 if we wanted to; note that eigenvectors are typically scaled to unit length 1. For the sake of completeness, let's plot the decision regions of the logistic regression on the transformed test dataset to see if it can separate the classes well: >>> >>> >>> >>> >>> >>>

plot_decision_regions(X_test_pca, y_test, classifier=lr) plt.xlabel('PC1') plt.ylabel('PC2') plt.legend(loc='lower left') plt.tight_layout() plt.show()

After we plot the decision regions for the test dataset by executing the preceding code, we can see that logistic regression performs quite well on this small two-dimensional feature subspace and only misclassifies a few examples in the test dataset:

If we are interested in the explained variance ratios of the different principal components, we can simply initialize the PCA class with the n_components parameter set to None, so all principal components are kept and the explained variance ratio can then be accessed via the explained_variance_ratio_ attribute: >>> pca = PCA(n_components=None) >>> X_train_pca = pca.fit_transform(X_train_std) >>> pca.explained_variance_ratio_ array([ 0.36951469, 0.18434927, 0.11815159, 0.07334252, 0.06422108, 0.05051724, 0.03954654, 0.02643918, 0.02389319, 0.01629614, 0.01380021, 0.01172226, 0.00820609]) [ 158 ]

Chapter 5

Note that we set n_components=None when we initialized the PCA class so that it will return all principal components in a sorted order, instead of performing a dimensionality reduction.

Supervised data compression via linear discriminant analysis

LDA can be used as a technique for feature extraction to increase the computational efficiency and reduce the degree of overfitting due to the curse of dimensionality in non-regularized models. The general concept behind LDA is very similar to PCA, but whereas PCA attempts to find the orthogonal component axes of maximum variance in a dataset, the goal in LDA is to find the feature subspace that optimizes class separability. In the following sections, we will discuss the similarities between LDA and PCA in more detail and walk through the LDA approach step by step.

Principal component analysis versus linear discriminant analysis

Both PCA and LDA are linear transformation techniques that can be used to reduce the number of dimensions in a dataset; the former is an unsupervised algorithm, whereas the latter is supervised. Thus, we might think that LDA is a superior feature extraction technique for classification tasks compared to PCA. However, A.M. Martinez reported that preprocessing via PCA tends to result in better classification results in an image recognition task in certain cases, for instance, if each class consists of only a small number of examples (PCA Versus LDA, A. M. Martinez and A. C. Kak, IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(2): 228-233, 2001). Fisher LDA LDA is sometimes also called Fisher's LDA. Ronald A. Fisher initially formulated Fisher's Linear Discriminant for two-class classification problems in 1936 (The Use of Multiple Measurements in Taxonomic Problems, R. A. Fisher, Annals of Eugenics, 7(2): 179188, 1936). Fisher's linear discriminant was later generalized for multi-class problems by C. Radhakrishna Rao under the assumption of equal class covariances and normally distributed classes in 1948, which we now call LDA (The Utilization of Multiple Measurements in Problems of Biological Classification, C. R. Rao, Journal of the Royal Statistical Society. Series B (Methodological), 10(2): 159-203, 1948).

[ 159 ]

Compressing Data via Dimensionality Reduction

The following figure summarizes the concept of LDA for a two-class problem. Examples from class 1 are shown as circles, and examples from class 2 are shown as crosses:

A linear discriminant, as shown on the x-axis (LD 1), would separate the two normal distributed classes well. Although the exemplary linear discriminant shown on the y-axis (LD 2) captures a lot of the variance in the dataset, it would fail as a good linear discriminant since it does not capture any of the class-discriminatory information. One assumption in LDA is that the data is normally distributed. Also, we assume that the classes have identical covariance matrices and that the training examples are statistically independent of each other. However, even if one, or more, of those assumptions is (slightly) violated, LDA for dimensionality reduction can still work reasonably well (Pattern Classification 2nd Edition, R. O. Duda, P. E. Hart, and D. G. Stork, New York, 2001).

The inner workings of linear discriminant analysis

Before we dive into the code implementation, let's briefly summarize the main steps that are required to perform LDA: 1. Standardize the d-dimensional dataset (d is the number of features). 2. For each class, compute the d-dimensional mean vector. 3. Construct the between-class scatter matrix, 𝑺𝑺𝐵𝐵, and the within-class scatter matrix, 𝑺𝑺𝑤𝑤.

[ 160 ]

Chapter 5

4. Compute the eigenvectors and corresponding eigenvalues of the matrix, 𝑺𝑺−1 𝑊𝑊 𝑺𝑺𝐵𝐵.

5. Sort the eigenvalues by decreasing order to rank the corresponding eigenvectors.

6. Choose the k eigenvectors that correspond to the k largest eigenvalues to construct a 𝑑𝑑 × 𝑘𝑘-dimensional transformation matrix, W; the eigenvectors are the columns of this matrix.

7. Project the examples onto the new feature subspace using the transformation matrix, W.

As we can see, LDA is quite similar to PCA in the sense that we are decomposing matrices into eigenvalues and eigenvectors, which will form the new lowerdimensional feature space. However, as mentioned before, LDA takes class label information into account, which is represented in the form of the mean vectors computed in step 2. In the following sections, we will discuss these seven steps in more detail, accompanied by illustrative code implementations.

Computing the scatter matrices

Since we already standardized the features of the Wine dataset in the PCA section at the beginning of this chapter, we can skip the first step and proceed with the calculation of the mean vectors, which we will use to construct the within-class scatter matrix and between-class scatter matrix, respectively. Each mean vector, 𝒎𝒎𝑖𝑖, stores the mean feature value, 𝜇𝜇𝑚𝑚, with respect to the examples of class i:

𝒎𝒎𝑖𝑖 = This results in three mean vectors:

1 ∑ 𝒙𝒙𝑚𝑚 𝑛𝑛𝑖𝑖 𝒙𝒙∈𝐷𝐷𝑖𝑖

𝜇𝜇𝑖𝑖,𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎ℎ𝑜𝑜𝑜𝑜 𝑇𝑇 𝜇𝜇𝑖𝑖,𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝒎𝒎𝑖𝑖 = [ ] ⋮ 𝜇𝜇𝑖𝑖,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝

>>> np.set_printoptions(precision=4) >>> mean_vecs = [] >>> for label in range(1,4): ... mean_vecs.append(np.mean( [ 161 ]

𝑖𝑖 ∈ {1, 2, 3}

Compressing Data via Dimensionality Reduction ... X_train_std[y_train==label], axis=0)) ... print('MV %s: %s\n' %(label, mean_vecs[label-1])) MV 1: [ 0.9066 -0.3497 0.3201 -0.7189 0.5056 0.8807 0.9589 -0.5516 0.5416 0.2338 0.5897 0.6563 1.2075] MV 2: [-0.8749 -0.2848 -0.3735 0.3157 -0.3848 -0.0433 0.0635 -0.0946 0.0703 -0.8286 0.3144 0.3608 -0.7253] MV 3: [ 0.1992 0.866 0.1682 0.4148 -0.0451 -1.0286 -1.2876 0.8287 -0.7795 0.9649 -1.209 -1.3622 -0.4013]

Using the mean vectors, we can now compute the within-class scatter matrix, 𝑺𝑺𝑊𝑊 : 𝑐𝑐

𝑺𝑺𝑊𝑊 = ∑ 𝑺𝑺𝑖𝑖 𝑖𝑖=1

This is calculated by summing up the individual scatter matrices, 𝑺𝑺𝑖𝑖, of each individual class i:

𝑺𝑺𝑖𝑖 = ∑ (𝒙𝒙 − 𝒎𝒎𝑖𝑖 )(𝒙𝒙 − 𝒎𝒎𝑖𝑖 )𝑇𝑇 𝒙𝒙∈𝐷𝐷𝑖𝑖

>>> d = 13 # number of features >>> S_W = np.zeros((d, d)) >>> for label, mv in zip(range(1, 4), mean_vecs): ... class_scatter = np.zeros((d, d)) >>> for row in X_train_std[y_train == label]: ... row, mv = row.reshape(d, 1), mv.reshape(d, 1) ... class_scatter += (row - mv).dot((row - mv).T) ... S_W += class_scatter >>> print('Within-class scatter matrix: %sx%s' % ( ... S_W.shape[0], S_W.shape[1])) Within-class scatter matrix: 13x13

The assumption that we are making when we are computing the scatter matrices is that the class labels in the training dataset are uniformly distributed. However, if we print the number of class labels, we see that this assumption is violated: >>> print('Class label distribution: %s' ... % np.bincount(y_train)[1:]) Class label distribution: [41 50 33]

[ 162 ]

Chapter 5

Thus, we want to scale the individual scatter matrices, 𝑺𝑺𝑖𝑖, before we sum them up as scatter matrix 𝑺𝑺𝑊𝑊. When we divide the scatter matrices by the number of classexamples, 𝑛𝑛𝑖𝑖, we can see that computing the scatter matrix is in fact the same as computing the covariance matrix, Σ𝑖𝑖—the covariance matrix is a normalized version of the scatter matrix:

Σ𝑖𝑖 =

1 1 𝑺𝑺𝑖𝑖 = ∑ (𝒙𝒙 − 𝒎𝒎𝑖𝑖 )(𝒙𝒙 − 𝒎𝒎𝑖𝑖 )𝑇𝑇 𝑛𝑛𝑖𝑖 𝑛𝑛𝑖𝑖 𝒙𝒙∈𝐷𝐷𝑖𝑖

The code for computing the scaled within-class scatter matrix is as follows: >>> d = 13 # number of features >>> S_W = np.zeros((d, d)) >>> for label,mv in zip(range(1, 4), mean_vecs): ... class_scatter = np.cov(X_train_std[y_train==label].T) ... S_W += class_scatter >>> print('Scaled within-class scatter matrix: %sx%s' ... % (S_W.shape[0], S_W.shape[1])) Scaled within-class scatter matrix: 13x13

After we compute the scaled within-class scatter matrix (or covariance matrix), we can move on to the next step and compute the between-class scatter matrix 𝑺𝑺𝑩𝑩: 𝒄𝒄

𝑺𝑺𝑩𝑩 = ∑ 𝑛𝑛𝑖𝑖 (𝒎𝒎𝑖𝑖 − 𝒎𝒎)(𝒎𝒎𝑖𝑖 − 𝒎𝒎)𝑇𝑇 𝑖𝑖=1

Here, m is the overall mean that is computed, including examples from all c classes: >>> mean_overall = np.mean(X_train_std, axis=0) >>> d = 13 # number of features >>> S_B = np.zeros((d, d)) >>> for i, mean_vec in enumerate(mean_vecs): ... n = X_train_std[y_train == i + 1, :].shape[0] ... mean_vec = mean_vec.reshape(d, 1) # make column vector ... mean_overall = mean_overall.reshape(d, 1) ... S_B += n * (mean_vec - mean_overall).dot( ... (mean_vec - mean_overall).T) >>> print('Between-class scatter matrix: %sx%s' % ( ... S_B.shape[0], S_B.shape[1])) Between-class scatter matrix: 13x13

[ 163 ]

Compressing Data via Dimensionality Reduction

Selecting linear discriminants for the new feature subspace

The remaining steps of the LDA are similar to the steps of the PCA. However, instead of performing the eigendecomposition on the covariance matrix, we solve the generalized eigenvalue problem of the matrix, 𝑺𝑺−1 𝑊𝑊 𝑺𝑺𝐵𝐵: >>> eigen_vals, eigen_vecs =\ ... np.linalg.eig(np.linalg.inv(S_W).dot(S_B))

After we compute the eigenpairs, we can sort the eigenvalues in descending order: >>> ... >>> ... >>> >>> ...

eigen_pairs = [(np.abs(eigen_vals[i]), eigen_vecs[:,i]) for i in range(len(eigen_vals))] eigen_pairs = sorted(eigen_pairs, key=lambda k: k[0], reverse=True) print('Eigenvalues in descending order:\n') for eigen_val in eigen_pairs: print(eigen_val[0])

Eigenvalues in descending order: 349.617808906 172.76152219 3.78531345125e-14 2.11739844822e-14 1.51646188942e-14 1.51646188942e-14 1.35795671405e-14 1.35795671405e-14 7.58776037165e-15 5.90603998447e-15 5.90603998447e-15 2.25644197857e-15 0.0

[ 164 ]

Chapter 5

In LDA, the number of linear discriminants is at most c−1, where c is the number of class labels, since the in-between scatter matrix, 𝑺𝑺𝑩𝑩, is the sum of c matrices with rank one or less. We can indeed see that we only have two nonzero eigenvalues (the eigenvalues 3-13 are not exactly zero, but this is due to the floating-point arithmetic in NumPy). Collinearity Note that in the rare case of perfect collinearity (all aligned example points fall on a straight line), the covariance matrix would have rank one, which would result in only one eigenvector with a nonzero eigenvalue.

To measure how much of the class-discriminatory information is captured by the linear discriminants (eigenvectors), let's plot the linear discriminants by decreasing eigenvalues, similar to the explained variance plot that we created in the PCA section. For simplicity, we will call the content of class-discriminatory information discriminability: >>> >>> >>> >>> ... >>> ... >>> >>> >>> >>> >>> >>>

tot = sum(eigen_vals.real) discr = [(i / tot) for i in sorted(eigen_vals.real, reverse=True)] cum_discr = np.cumsum(discr) plt.bar(range(1, 14), discr, alpha=0.5, align='center', label='Individual "discriminability"') plt.step(range(1, 14), cum_discr, where='mid', label='Cumulative "discriminability"') plt.ylabel('"Discriminability" ratio') plt.xlabel('Linear Discriminants') plt.ylim([-0.1, 1.1]) plt.legend(loc='best') plt.tight_layout() plt.show()

[ 165 ]

Compressing Data via Dimensionality Reduction

As we can see in the resulting figure, the first two linear discriminants alone capture 100 percent of the useful information in the Wine training dataset:

Let's now stack the two most discriminative eigenvector columns to create the transformation matrix, W: >>> w = np.hstack((eigen_pairs[0][1][:, np.newaxis].real, ... eigen_pairs[1][1][:, np.newaxis].real)) >>> print('Matrix W:\n', w) Matrix W: [[-0.1481 -0.4092] [ 0.0908 -0.1577] [-0.0168 -0.3537] [ 0.1484 0.3223] [-0.0163 -0.0817] [ 0.1913 0.0842] [-0.7338 0.2823] [-0.075 -0.0102] [ 0.0018 0.0907] [ 0.294 -0.2152] [-0.0328 0.2747] [-0.3547 -0.0124] [-0.3915 -0.5958]]

[ 166 ]

Chapter 5

Projecting examples onto the new feature space

Using the transformation matrix, W, that we created in the previous subsection, we can now transform the training dataset by multiplying the matrices:

>>> >>> >>> >>> ... ... ... >>> >>> >>> >>> >>>

𝑿𝑿′ = 𝑿𝑿𝑿𝑿

X_train_lda = X_train_std.dot(w) colors = ['r', 'b', 'g'] markers = ['s', 'x', 'o'] for l, c, m in zip(np.unique(y_train), colors, markers): plt.scatter(X_train_lda[y_train==l, 0], X_train_lda[y_train==l, 1] * (-1), c=c, label=l, marker=m) plt.xlabel('LD 1') plt.ylabel('LD 2') plt.legend(loc='lower right') plt.tight_layout() plt.show()

As we can see in the resulting plot, the three Wine classes are now perfectly linearly separable in the new feature subspace:

[ 167 ]

Compressing Data via Dimensionality Reduction

LDA via scikit-learn

That step-by-step implementation was a good exercise to understand the inner workings of an LDA and understand the differences between LDA and PCA. Now, let's look at the LDA class implemented in scikit-learn: >>> # the following import statement is one line >>> from sklearn.discriminant_analysis import LinearDiscriminantAnalysis as LDA >>> lda = LDA(n_components=2) >>> X_train_lda = lda.fit_transform(X_train_std, y_train)

Next, let's see how the logistic regression classifier handles the lower-dimensional training dataset after the LDA transformation: >>> ... >>> >>> >>> >>> >>> >>> >>>

lr = LogisticRegression(multi_class='ovr', random_state=1, solver='lbfgs') lr = lr.fit(X_train_lda, y_train) plot_decision_regions(X_train_lda, y_train, classifier=lr) plt.xlabel('LD 1') plt.ylabel('LD 2') plt.legend(loc='lower left') plt.tight_layout() plt.show()

Looking at the resulting plot, we can see that the logistic regression model misclassifies one of the examples from class 2:

[ 168 ]

Chapter 5

By lowering the regularization strength, we could probably shift the decision boundaries so that the logistic regression model classifies all examples in the training dataset correctly. However, and more importantly, let's take a look at the results on the test dataset: >>> >>> >>> >>> >>> >>> >>>

X_test_lda = lda.transform(X_test_std) plot_decision_regions(X_test_lda, y_test, classifier=lr) plt.xlabel('LD 1') plt.ylabel('LD 2') plt.legend(loc='lower left') plt.tight_layout() plt.show()

As we can see in the following plot, the logistic regression classifier is able to get a perfect accuracy score for classifying the examples in the test dataset by only using a two-dimensional feature subspace, instead of the original 13 Wine features:

Using kernel principal component analysis for nonlinear mappings

Many machine learning algorithms make assumptions about the linear separability of the input data. You have learned that the perceptron even requires perfectly linearly separable training data to converge. Other algorithms that we have covered so far assume that the lack of perfect linear separability is due to noise: Adaline, logistic regression, and the (standard) SVM to just name a few. [ 169 ]

Compressing Data via Dimensionality Reduction

However, if we are dealing with nonlinear problems, which we may encounter rather frequently in real-world applications, linear transformation techniques for dimensionality reduction, such as PCA and LDA, may not be the best choice. In this section, we will take a look at a kernelized version of PCA, or KPCA, which relates to the concepts of kernel SVM that you will remember from Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn. Using KPCA, we will learn how to transform data that is not linearly separable onto a new, lower-dimensional subspace that is suitable for linear classifiers.

Kernel functions and the kernel trick

As you will remember from our discussion about kernel SVMs in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, we can tackle nonlinear problems by projecting them onto a new feature space of higher dimensionality where the classes become linearly separable. To transform the examples 𝒙𝒙 ∈ ℝ𝑑𝑑 onto this higher k-dimensional subspace, we defined a nonlinear mapping function, 𝜙𝜙:

𝜙𝜙: ℝ𝑑𝑑 ⟶ ℝ𝑘𝑘

(𝑘𝑘 ≫ 𝑑𝑑)

We can think of 𝜙𝜙 as a function that creates nonlinear combinations of the original features to map the original d-dimensional dataset onto a larger, k-dimensional feature space.

[ 170 ]

Chapter 5

For example, if we had a feature vector 𝒙𝒙 ∈ ℝ𝑑𝑑 (x is a column vector consisting of d features) with two dimensions (d = 2), a potential mapping onto a 3D-space could be:

𝒙𝒙 = [𝑥𝑥1 , 𝑥𝑥2 ]𝑇𝑇 ↓ 𝜙𝜙

𝒛𝒛 = [𝑥𝑥12 , √2𝑥𝑥1 𝑥𝑥2 , 𝑥𝑥22 ]

𝑇𝑇

In other words, we perform a nonlinear mapping via KPCA that transforms the data onto a higher-dimensional space. We then use standard PCA in this higherdimensional space to project the data back onto a lower-dimensional space where the examples can be separated by a linear classifier (under the condition that the examples can be separated by density in the input space). However, one downside of this approach is that it is computationally very expensive, and this is where we use the kernel trick. Using the kernel trick, we can compute the similarity between two high-dimension feature vectors in the original feature space. Before we proceed with more details about the kernel trick to tackle this computationally expensive problem, let's think back to the standard PCA approach that we implemented at the beginning of this chapter. We computed the covariance between two features, k and j, as follows: 𝑛𝑛

1 (𝑖𝑖) (𝑖𝑖) 𝜎𝜎𝑗𝑗𝑗𝑗 = ∑(𝑥𝑥𝑗𝑗 − 𝜇𝜇𝑗𝑗 )(𝑥𝑥𝑘𝑘 − 𝜇𝜇𝑘𝑘 ) 𝑛𝑛 𝑖𝑖=1

Since the standardizing of features centers them at mean zero, for instance, 𝜇𝜇𝑗𝑗 = 0 and 𝜇𝜇𝑘𝑘 = 0, we can simplify this equation as follows: 𝑛𝑛

1 (𝑖𝑖) (𝑖𝑖) 𝜎𝜎𝑗𝑗𝑗𝑗 = ∑ 𝑥𝑥𝑗𝑗 𝑥𝑥𝑘𝑘 𝑛𝑛 𝑖𝑖=1

[ 171 ]

Compressing Data via Dimensionality Reduction

Note that the preceding equation refers to the covariance between two features; now, let's write the general equation to calculate the covariance matrix, Σ: 𝑛𝑛

1 𝑇𝑇 Σ = ∑ 𝒙𝒙(𝑖𝑖) 𝒙𝒙(𝑖𝑖) 𝑛𝑛 𝑖𝑖=1

Bernhard Scholkopf generalized this approach (Kernel principal component analysis, B. Scholkopf, A. Smola, and K.R. Muller, pages 583-588, 1997) so that we can replace the dot products between examples in the original feature space with the nonlinear feature combinations via 𝜙𝜙: 𝑛𝑛

1 𝑇𝑇 Σ = ∑ 𝜙𝜙(𝒙𝒙(𝑖𝑖) )𝜙𝜙(𝒙𝒙(𝑖𝑖) ) 𝑛𝑛 𝑖𝑖=1

To obtain the eigenvectors—the principal components—from this covariance matrix, we have to solve the following equation:

Σ𝒗𝒗 = 𝜆𝜆𝒗𝒗 𝑛𝑛

1 𝑇𝑇 ⇒ ∑ 𝜙𝜙(𝒙𝒙(𝑖𝑖) )𝜙𝜙(𝒙𝒙(𝑖𝑖) ) 𝒗𝒗 = 𝜆𝜆𝒗𝒗 𝑛𝑛 𝑖𝑖=1

𝑛𝑛

𝑛𝑛

𝑖𝑖=1

𝑖𝑖=1

1 1 𝑇𝑇 ⇒ 𝑣𝑣 = ∑ 𝜙𝜙(𝒙𝒙(𝑖𝑖) )𝜙𝜙(𝒙𝒙(𝑖𝑖) ) 𝒗𝒗 = ∑ 𝒂𝒂(𝒊𝒊) 𝜙𝜙(𝒙𝒙(𝑖𝑖) ) 𝑛𝑛 𝑛𝑛𝑛𝑛

Here, 𝜆𝜆 and v are the eigenvalues and eigenvectors of the covariance matrix, Σ, and a can be obtained by extracting the eigenvectors of the kernel (similarity) matrix, K, as you will see in the following paragraphs.

[ 172 ]

Chapter 5

Deriving the kernel matrix The derivation of the kernel matrix can be shown as follows. First, let's write the covariance matrix as in matrix notation, where 𝜙𝜙(𝑋𝑋) is an 𝑛𝑛 × 𝑘𝑘-dimensional matrix: 𝑛𝑛

1 1 𝑇𝑇 Σ = ∑ 𝜙𝜙(𝒙𝒙(𝑖𝑖) )𝜙𝜙(𝒙𝒙(𝑖𝑖) ) = 𝜙𝜙(𝑿𝑿)𝑇𝑇 𝜙𝜙(𝑿𝑿) 𝑛𝑛 𝑛𝑛 𝑖𝑖=1

Now, we can write the eigenvector equation as follows: 𝑛𝑛

1 𝑣𝑣 = ∑ 𝑎𝑎(𝒊𝒊) 𝜙𝜙(𝒙𝒙(𝑖𝑖) ) = 𝜆𝜆𝜙𝜙(𝑿𝑿)𝑇𝑇 𝒂𝒂 𝑛𝑛 𝑖𝑖=1

Since Σ𝒗𝒗 = 𝜆𝜆𝒗𝒗, we get:

1 𝜙𝜙(𝑿𝑿)𝑇𝑇 𝜙𝜙(𝑿𝑿)𝜙𝜙(𝑿𝑿)𝑇𝑇 𝒂𝒂 = 𝜆𝜆𝜙𝜙(𝑿𝑿)𝑇𝑇 𝒂𝒂 𝑛𝑛

Multiplying it by 𝜙𝜙(𝑿𝑿) on both sides yields the following result: 1 𝜙𝜙(𝑿𝑿)𝜙𝜙(𝑿𝑿)𝑇𝑇 𝜙𝜙(𝑿𝑿)𝜙𝜙(𝑿𝑿)𝑇𝑇 𝒂𝒂 = 𝜆𝜆𝜆𝜆(𝑿𝑿)𝜙𝜙(𝑿𝑿)𝑇𝑇 𝒂𝒂 𝑛𝑛 1 ⇒ 𝜙𝜙(𝑿𝑿)𝜙𝜙(𝑿𝑿)𝑇𝑇 𝒂𝒂 = 𝜆𝜆𝒂𝒂 𝑛𝑛 1 ⇒ 𝑲𝑲𝑲𝑲 = 𝜆𝜆𝒂𝒂 𝑛𝑛

Here, K is the similarity (kernel) matrix: 𝑲𝑲 = 𝜙𝜙(𝑿𝑿)𝜙𝜙(𝑿𝑿)𝑇𝑇

As you might recall from the Solving nonlinear problems using a kernel SVM section in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, we use the kernel trick to avoid calculating the pairwise dot products of the examples, x, under 𝜙𝜙 explicitly by using a kernel function, 𝜅𝜅, so that we don't need to calculate the eigenvectors explicitly: 𝑇𝑇

𝜅𝜅(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = 𝜙𝜙(𝒙𝒙(𝑖𝑖) ) 𝜙𝜙(𝒙𝒙(𝑗𝑗) )

[ 173 ]

Compressing Data via Dimensionality Reduction

In other words, what we obtain after KPCA are the examples already projected onto the respective components, rather than constructing a transformation matrix as in the standard PCA approach. Basically, the kernel function (or simply kernel) can be understood as a function that calculates a dot product between two vectors—a measure of similarity. The most commonly used kernels are as follows: • The polynomial kernel: 𝑝𝑝

𝑇𝑇

𝜅𝜅(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = (𝒙𝒙(𝑖𝑖) 𝒙𝒙(𝑗𝑗) + 𝜃𝜃)

Here, 𝜃𝜃 is the threshold and p is the power that has to be specified by the user.

• The hyperbolic tangent (sigmoid) kernel:

𝑇𝑇

𝜅𝜅(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = tanh (𝜂𝜂𝒙𝒙(𝑖𝑖) 𝒙𝒙(𝑗𝑗) + 𝜃𝜃)

• The radial basis function (RBF) or Gaussian kernel, which we will use in the following examples in the next subsection:

2

‖𝒙𝒙(𝑖𝑖) − 𝒙𝒙(𝑗𝑗) ‖ 𝜅𝜅(𝒙𝒙 , 𝒙𝒙 ) = exp (− ) 2𝜎𝜎 2 (𝑖𝑖)

(𝑗𝑗)

1 It is often written in the following form, introducing the variable 𝛾𝛾 = 2𝜎𝜎 2 . 2

𝜅𝜅(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = exp (−𝛾𝛾‖𝒙𝒙(𝑖𝑖) − 𝒙𝒙(𝑗𝑗) ‖ )

To summarize what we have learned so far, we can define the following three steps to implement an RBF KPCA: 1. We compute the kernel (similarity) matrix, K, where we need to calculate the following: 2

𝜅𝜅(𝒙𝒙(𝑖𝑖) , 𝒙𝒙(𝑗𝑗) ) = exp (−𝛾𝛾‖𝒙𝒙(𝑖𝑖) − 𝒙𝒙(𝑗𝑗) ‖ ) [ 174 ]

Chapter 5

We do this for each pair of examples:

𝜅𝜅(𝒙𝒙(1) , 𝒙𝒙(1) )

𝜅𝜅(𝒙𝒙(1) , 𝒙𝒙(2) )

(2) (1) (2) (2) 𝑲𝑲 = 𝜅𝜅(𝒙𝒙 , 𝒙𝒙 ) (𝒙𝒙 , 𝒙𝒙 ) ⋮ ⋮ (𝑛𝑛) (1) (𝑛𝑛) (2) 𝜅𝜅(𝒙𝒙 , 𝒙𝒙 ) 𝜅𝜅(𝒙𝒙 , 𝒙𝒙 ) [

⋯

𝜅𝜅(𝒙𝒙(1) , 𝒙𝒙(𝑛𝑛) )

⋯ 𝜅𝜅(𝒙𝒙(2) , 𝒙𝒙(𝑛𝑛) ) ⋱ ⋮ ⋯ 𝜅𝜅(𝒙𝒙(𝑛𝑛) , 𝒙𝒙(𝑛𝑛) )]

For example, if our dataset contains 100 training examples, the symmetric kernel matrix of the pairwise similarities would be 100 × 100-dimensional.

2. We center the kernel matrix, K, using the following equation: 𝑲𝑲′ = 𝑲𝑲 − 𝟏𝟏𝒏𝒏 𝑲𝑲 − 𝑲𝑲𝟏𝟏𝒏𝒏 + 𝟏𝟏𝒏𝒏 𝑲𝑲𝟏𝟏𝒏𝒏

Here, 𝟏𝟏𝒏𝒏 is an 𝑛𝑛 × 𝑛𝑛-dimensional matrix (the same dimensions as the kernel 1 matrix) where all values are equal to . 𝑛𝑛 3. We collect the top k eigenvectors of the centered kernel matrix based on their corresponding eigenvalues, which are ranked by decreasing magnitude. In contrast to standard PCA, the eigenvectors are not the principal component axes, but the examples already projected onto these axes. At this point, you may be wondering why we need to center the kernel matrix in the second step. We previously assumed that we are working with standardized data, where all features have mean zero when we formulate the covariance matrix and replace the dot-products with the nonlinear feature combinations via 𝜙𝜙. Thus, the centering of the kernel matrix in the second step becomes necessary, since we do not compute the new feature space explicitly so that we cannot guarantee that the new feature space is also centered at zero. In the next section, we will put those three steps into action by implementing a KPCA in Python.

Implementing a kernel principal component analysis in Python

In the previous subsection, we discussed the core concepts behind KPCA. Now, we are going to implement an RBF KPCA in Python following the three steps that summarized the KPCA approach. Using some SciPy and NumPy helper functions, we will see that implementing a KPCA is actually really simple:

[ 175 ]

Compressing Data via Dimensionality Reduction from scipy.spatial.distance import pdist, squareform from scipy import exp from scipy.linalg import eigh import numpy as np def rbf_kernel_pca(X, gamma, n_components): """ RBF kernel PCA implementation. Parameters -----------X: {NumPy ndarray}, shape = [n_examples, n_features] gamma: float Tuning parameter of the RBF kernel n_components: int Number of principal components to return Returns -----------X_pc: {NumPy ndarray}, shape = [n_examples, k_features] Projected dataset """ # Calculate pairwise squared Euclidean distances # in the MxN dimensional dataset. sq_dists = pdist(X, 'sqeuclidean') # Convert pairwise distances into a square matrix. mat_sq_dists = squareform(sq_dists) # Compute the symmetric kernel matrix. K = exp(-gamma * mat_sq_dists) # Center the kernel matrix. N = K.shape[0] one_n = np.ones((N,N)) / N K = K - one_n.dot(K) - K.dot(one_n) + one_n.dot(K).dot(one_n) # Obtaining eigenpairs from the centered kernel matrix # scipy.linalg.eigh returns them in ascending order eigvals, eigvecs = eigh(K) eigvals, eigvecs = eigvals[::-1], eigvecs[:, ::-1] [ 176 ]

Chapter 5

# Collect the top k eigenvectors (projected examples) X_pc = np.column_stack([eigvecs[:, i] for i in range(n_components)]) return X_pc

One downside of using an RBF KPCA for dimensionality reduction is that we have to specify the 𝛾𝛾 parameter a priori. Finding an appropriate value for 𝛾𝛾 requires experimentation and is best done using algorithms for parameter tuning, for example, performing a grid search, which we will discuss in more detail in Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning.

Example 1 – separating half-moon shapes

Now, let us apply our rbf_kernel_pca on some nonlinear example datasets. We will start by creating a two-dimensional dataset of 100 example points representing two half-moon shapes: >>> >>> >>> ... >>> ... >>> >>>

from sklearn.datasets import make_moons X, y = make_moons(n_samples=100, random_state=123) plt.scatter(X[y==0, 0], X[y==0, 1], color='red', marker='^', alpha=0.5) plt.scatter(X[y==1, 0], X[y==1, 1], color='blue', marker='o', alpha=0.5) plt.tight_layout() plt.show()

For the purposes of illustration, the half-moon of triangle symbols will represent one class, and the half-moon depicted by the circle symbols will represent the examples from another class:

[ 177 ]

Compressing Data via Dimensionality Reduction

Clearly, these two half-moon shapes are not linearly separable, and our goal is to unfold the half-moons via KPCA so that the dataset can serve as a suitable input for a linear classifier. But first, let's see how the dataset looks if we project it onto the principal components via standard PCA: >>> >>> >>> >>> >>> ... >>> ... >>> ... >>> ... >>> >>> >>> >>> >>> >>> >>>

from sklearn.decomposition import PCA scikit_pca = PCA(n_components=2) X_spca = scikit_pca.fit_transform(X) fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(7,3)) ax[0].scatter(X_spca[y==0, 0], X_spca[y==0, 1], color='red', marker='^', alpha=0.5) ax[0].scatter(X_spca[y==1, 0], X_spca[y==1, 1], color='blue', marker='o', alpha=0.5) ax[1].scatter(X_spca[y==0, 0], np.zeros((50,1))+0.02, color='red', marker='^', alpha=0.5) ax[1].scatter(X_spca[y==1, 0], np.zeros((50,1))-0.02, color='blue', marker='o', alpha=0.5) ax[0].set_xlabel('PC1') ax[0].set_ylabel('PC2') ax[1].set_ylim([-1, 1]) ax[1].set_yticks([]) ax[1].set_xlabel('PC1') plt.tight_layout() plt.show()

Clearly, we can see in the resulting figure that a linear classifier would be unable to perform well on the dataset transformed via standard PCA:

[ 178 ]

Chapter 5

Note that when we plotted the first principal component only (right subplot), we shifted the triangular examples slightly upward and the circular examples slightly downward to better visualize the class overlap. As the left subplot shows, the original half-moon shapes are only slightly sheared and flipped across the vertical center—this transformation would not help a linear classifier in discriminating between circles and triangles. Similarly, the circles and triangles corresponding to the two half-moon shapes are not linearly separable if we project the dataset onto a one-dimensional feature axis, as shown in the right subplot. PCA versus LDA Please remember that PCA is an unsupervised method and does not use class label information in order to maximize the variance in contrast to LDA. Here, the triangle and circle symbols were just added for visualization purposes to indicate the degree of separation.

Now, let's try out our kernel PCA function, rbf_kernel_pca, which we implemented in the previous subsection: >>> >>> >>> ... >>> ... >>> ... >>> ... >>> >>> >>> >>> >>> >>> >>>

X_kpca = rbf_kernel_pca(X, gamma=15, n_components=2) fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(7, 3)) ax[0].scatter(X_kpca[y==0, 0], X_kpca[y==0, 1], color='red', marker='^', alpha=0.5) ax[0].scatter(X_kpca[y==1, 0], X_kpca[y==1, 1], color='blue', marker='o', alpha=0.5) ax[1].scatter(X_kpca[y==0, 0], np.zeros((50,1))+0.02, color='red', marker='^', alpha=0.5) ax[1].scatter(X_kpca[y==1, 0], np.zeros((50,1))-0.02, color='blue', marker='o', alpha=0.5) ax[0].set_xlabel('PC1') ax[0].set_ylabel('PC2') ax[1].set_ylim([-1, 1]) ax[1].set_yticks([]) ax[1].set_xlabel('PC1') plt.tight_layout() plt.show()

[ 179 ]

Compressing Data via Dimensionality Reduction

We can now see that the two classes (circles and triangles) are linearly well separated so that we have a suitable training dataset for linear classifiers:

Unfortunately, there is no universal value for the tuning parameter, 𝛾𝛾, that works well for different datasets. Finding a 𝛾𝛾 value that is appropriate for a given problem requires experimentation. In Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, we will discuss techniques that can help us to automate the task of optimizing such tuning parameters. Here, we will use values for 𝛾𝛾 that have been found to produce good results.

Example 2 – separating concentric circles

In the previous subsection, we saw how to separate half-moon shapes via KPCA. Since we put so much effort into understanding the concepts of KPCA, let's take a look at another interesting example of a nonlinear problem, concentric circles: >>> >>> ... ... >>> ... >>> ... >>> >>>

from sklearn.datasets import make_circles X, y = make_circles(n_samples=1000, random_state=123, noise=0.1, factor=0.2) plt.scatter(X[y == 0, 0], X[y == 0, 1], color='red', marker='^', alpha=0.5) plt.scatter(X[y == 1, 0], X[y == 1, 1], color='blue', marker='o', alpha=0.5) plt.tight_layout() plt.show()

[ 180 ]

Chapter 5

Again, we assume a two-class problem where the triangle shapes represent one class, and the circle shapes represent another class:

Let's start with the standard PCA approach to compare it to the results of the RBF kernel PCA: >>> >>> >>> >>> ... >>> ... >>> ... >>> ... >>> >>> >>> >>> >>> >>> >>>

scikit_pca = PCA(n_components=2) X_spca = scikit_pca.fit_transform(X) fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(7,3)) ax[0].scatter(X_spca[y==0, 0], X_spca[y==0, 1], color='red', marker='^', alpha=0.5) ax[0].scatter(X_spca[y==1, 0], X_spca[y==1, 1], color='blue', marker='o', alpha=0.5) ax[1].scatter(X_spca[y==0, 0], np.zeros((500,1))+0.02, color='red', marker='^', alpha=0.5) ax[1].scatter(X_spca[y==1, 0], np.zeros((500,1))-0.02, color='blue', marker='o', alpha=0.5) ax[0].set_xlabel('PC1') ax[0].set_ylabel('PC2') ax[1].set_ylim([-1, 1]) ax[1].set_yticks([]) ax[1].set_xlabel('PC1') plt.tight_layout() plt.show()

[ 181 ]

Compressing Data via Dimensionality Reduction

Again, we can see that standard PCA is not able to produce results suitable for training a linear classifier:

Given an appropriate value for 𝛾𝛾, let's see if we are luckier using the RBF KPCA implementation: >>> >>> >>> ... >>> ... >>> ... >>> ... >>> >>> >>> >>> >>> >>> >>>

X_kpca = rbf_kernel_pca(X, gamma=15, n_components=2) fig, ax = plt.subplots(nrows=1,ncols=2, figsize=(7,3)) ax[0].scatter(X_kpca[y==0, 0], X_kpca[y==0, 1], color='red', marker='^', alpha=0.5) ax[0].scatter(X_kpca[y==1, 0], X_kpca[y==1, 1], color='blue', marker='o', alpha=0.5) ax[1].scatter(X_kpca[y==0, 0], np.zeros((500,1))+0.02, color='red', marker='^', alpha=0.5) ax[1].scatter(X_kpca[y==1, 0], np.zeros((500,1))-0.02, color='blue', marker='o', alpha=0.5) ax[0].set_xlabel('PC1') ax[0].set_ylabel('PC2') ax[1].set_ylim([-1, 1]) ax[1].set_yticks([]) ax[1].set_xlabel('PC1') plt.tight_layout() plt.show()

Again, the RBF KPCA projected the data onto a new subspace where the two classes become linearly separable:

[ 182 ]

Chapter 5

Projecting new data points

In the two previous example applications of KPCA, the half-moon shapes and the concentric circles, we projected a single dataset onto a new feature. In real applications, however, we may have more than one dataset that we want to transform, for example, training and test data, and typically also new examples we will collect after the model building and evaluation. In this section, you will learn how to project data points that were not part of the training dataset. As you will remember from the standard PCA approach at the beginning of this chapter, we project data by calculating the dot product between a transformation matrix and the input examples; the columns of the projection matrix are the top k eigenvectors (v) that we obtained from the covariance matrix. Now, the question is how we can transfer this concept to KPCA. If we think back to the idea behind KPCA, we will remember that we obtained an eigenvector (a) of the centered kernel matrix (not the covariance matrix), which means that those are the examples that are already projected onto the principal component axis, v. Thus, if we want to project a new example, 𝒙𝒙′, onto this principal component axis, we will need to compute the following:

𝜙𝜙(𝒙𝒙′)𝑇𝑇 𝒗𝒗

Fortunately, we can use the kernel trick so that we don't have to calculate the projection, 𝜙𝜙(𝒙𝒙′)𝑇𝑇 𝒗𝒗, explicitly. However, it is worth noting that KPCA, in contrast to standard PCA, is a memory-based method, which means that we have to reuse the original training dataset each time to project new examples.

[ 183 ]

Compressing Data via Dimensionality Reduction

We have to calculate the pairwise RBF kernel (similarity) between each ith example in the training dataset and the new example, 𝒙𝒙′:

𝜙𝜙(𝒙𝒙′)𝑇𝑇 𝒗𝒗 = ∑ 𝑎𝑎(𝑖𝑖) 𝜙𝜙(𝒙𝒙′ )𝑻𝑻𝜙𝜙(𝒙𝒙(𝑖𝑖) ) 𝑖𝑖

= ∑ 𝑎𝑎(𝑖𝑖) 𝜅𝜅(𝒙𝒙′ , 𝒙𝒙(𝑖𝑖) ) 𝑖𝑖

Here, the eigenvectors, a, and eigenvalues, 𝜆𝜆, of the kernel matrix, K, satisfy the following condition in the equation:

𝑲𝑲𝑲𝑲 = 𝜆𝜆𝒂𝒂

After calculating the similarity between the new examples and the examples in the training dataset, we have to normalize the eigenvector, a, by its eigenvalue. Thus, let's modify the rbf_kernel_pca function that we implemented earlier so that it also returns the eigenvalues of the kernel matrix: from scipy.spatial.distance import pdist, squareform from scipy import exp from scipy.linalg import eigh import numpy as np def rbf_kernel_pca(X, gamma, n_components): """ RBF kernel PCA implementation. Parameters -----------X: {NumPy ndarray}, shape = [n_examples, n_features] gamma: float Tuning parameter of the RBF kernel n_components: int Number of principal components to return Returns -----------alphas {NumPy ndarray}, shape = [n_examples, k_features] Projected dataset lambdas: list [ 184 ]

Chapter 5 Eigenvalues """ # Calculate pairwise squared Euclidean distances # in the MxN dimensional dataset. sq_dists = pdist(X, 'sqeuclidean') # Convert pairwise distances into a square matrix. mat_sq_dists = squareform(sq_dists) # Compute the symmetric kernel matrix. K = exp(-gamma * mat_sq_dists) # Center the kernel matrix. N = K.shape[0] one_n = np.ones((N,N)) / N K = K - one_n.dot(K) - K.dot(one_n) + one_n.dot(K).dot(one_n) # Obtaining eigenpairs from the centered kernel matrix # scipy.linalg.eigh returns them in ascending order eigvals, eigvecs = eigh(K) eigvals, eigvecs = eigvals[::-1], eigvecs[:, ::-1] # Collect the top k eigenvectors (projected examples) alphas = np.column_stack([eigvecs[:, i] for i in range(n_components)]) # Collect the corresponding eigenvalues lambdas = [eigvals[i] for i in range(n_components)] return alphas, lambdas

Now, let's create a new half-moon dataset and project it onto a one-dimensional subspace using the updated RBF KPCA implementation: >>> X, y = make_moons(n_samples=100, random_state=123) >>> alphas, lambdas = rbf_kernel_pca(X, gamma=15, n_components=1)

To make sure that we implemented the code for projecting new examples, let's assume that the 26th point from the half-moon dataset is a new data point, 𝒙𝒙′, and our task is to project it onto this new subspace: >>> x_new = X[25] >>> x_new array([ 1.8713187 , 0.00928245]) >>> x_proj = alphas[25] # original projection

[ 185 ]

Compressing Data via Dimensionality Reduction >>> x_proj array([ 0.07877284]) >>> def project_x(x_new, X, gamma, alphas, lambdas): ... pair_dist = np.array([np.sum( ... (x_new-row)**2) for row in X]) ... k = np.exp(-gamma * pair_dist) ... return k.dot(alphas / lambdas)

By executing the following code, we are able to reproduce the original projection. Using the project_x function, we will be able to project any new data example as well. The code is as follows: >>> x_reproj = project_x(x_new, X, ... gamma=15, alphas=alphas, ... lambdas=lambdas) >>> x_reproj array([ 0.07877284])

Lastly, let's visualize the projection on the first principal component: >>> ... >>> ... >>> ... ... >>> ... ... >>> >>> >>> >>>

plt.scatter(alphas[y==0, 0], np.zeros((50)), color='red', marker='^',alpha=0.5) plt.scatter(alphas[y==1, 0], np.zeros((50)), color='blue', marker='o', alpha=0.5) plt.scatter(x_proj, 0, color='black', label='Original projection of point X[25]', marker='^', s=100) plt.scatter(x_reproj, 0, color='green', label='Remapped point X[25]', marker='x', s=500) plt.yticks([], []) plt.legend(scatterpoints=1) plt.tight_layout() plt.show()

As we can now also see in the following scatterplot, we mapped the example, 𝒙𝒙′, onto the first principal component correctly:

[ 186 ]

Chapter 5

Kernel principal component analysis in scikitlearn

For our convenience, scikit-learn implements a KPCA class in the sklearn. decomposition submodule. The usage is similar to the standard PCA class, and we can specify the kernel via the kernel parameter: >>> >>> >>> ... >>>

from sklearn.decomposition import KernelPCA X, y = make_moons(n_samples=100, random_state=123) scikit_kpca = KernelPCA(n_components=2, kernel='rbf', gamma=15) X_skernpca = scikit_kpca.fit_transform(X)

To check that we get results that are consistent with our own KPCA implementation, let's plot the transformed half-moon shape data onto the first two principal components: >>> ... >>> ... >>> >>> >>> >>>

plt.scatter(X_skernpca[y==0, 0], X_skernpca[y==0, 1], color='red', marker='^', alpha=0.5) plt.scatter(X_skernpca[y==1, 0], X_skernpca[y==1, 1], color='blue', marker='o', alpha=0.5) plt.xlabel('PC1') plt.ylabel('PC2') plt.tight_layout() plt.show()

[ 187 ]

Compressing Data via Dimensionality Reduction

As we can see, the results of scikit-learn's KernelPCA are consistent with our own implementation:

Manifold learning The scikit-learn library also implements advanced techniques for nonlinear dimensionality reduction that are beyond the scope of this book. The interested reader can find a nice overview of the current implementations in scikit-learn, complemented by illustrative examples, at http://scikit-learn.org/stable/ modules/manifold.html.

Summary

In this chapter, you learned about three different, fundamental dimensionality reduction techniques for feature extraction: standard PCA, LDA, and KPCA. Using PCA, we projected data onto a lower-dimensional subspace to maximize the variance along the orthogonal feature axes, while ignoring the class labels. LDA, in contrast to PCA, is a technique for supervised dimensionality reduction, which means that it considers class information in the training dataset to attempt to maximize the classseparability in a linear feature space.

[ 188 ]

Chapter 5

Lastly, you learned about a nonlinear feature extractor, KPCA. Using the kernel trick and a temporary projection into a higher-dimensional feature space, you were ultimately able to compress datasets consisting of nonlinear features onto a lowerdimensional subspace where the classes became linearly separable. Equipped with these essential preprocessing techniques, you are now well prepared to learn about the best practices for efficiently incorporating different preprocessing techniques and evaluating the performance of different models in the next chapter.

[ 189 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning In the previous chapters, you learned about the essential machine learning algorithms for classification and how to get our data into shape before we feed it into those algorithms. Now, it's time to learn about the best practices of building good machine learning models by fine-tuning the algorithms and evaluating the performance of the models. In this chapter, we will learn how to do the following: • Assess the performance of machine learning models • Diagnose the common problems of machine learning algorithms • Fine-tune machine learning models • Evaluate predictive models using different performance metrics

Streamlining workflows with pipelines

When we applied different preprocessing techniques in the previous chapters, such as standardization for feature scaling in Chapter 4, Building Good Training Datasets – Data Preprocessing, or principal component analysis for data compression in Chapter 5, Compressing Data via Dimensionality Reduction, you learned that we have to reuse the parameters that were obtained during the fitting of the training data to scale and compress any new data, such as the examples in the separate test dataset. In this section, you will learn about an extremely handy tool, the Pipeline class in scikitlearn. It allows us to fit a model including an arbitrary number of transformation steps and apply it to make predictions about new data.

[ 191 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

Loading the Breast Cancer Wisconsin dataset In this chapter, we will be working with the Breast Cancer Wisconsin dataset, which contains 569 examples of malignant and benign tumor cells. The first two columns in the dataset store the unique ID numbers of the examples and the corresponding diagnoses (M = malignant, B = benign), respectively. Columns 3-32 contain 30 real-valued features that have been computed from digitized images of the cell nuclei, which can be used to build a model to predict whether a tumor is benign or malignant. The Breast Cancer Wisconsin dataset has been deposited in the UCI Machine Learning Repository, and more detailed information about this dataset can be found at https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisco nsin+(Diagnostic). Obtaining the Breast Cancer Wisconsin dataset You can find a copy of the dataset (and all other datasets used in this book) in the code bundle of this book, which you can use if you are working offline or the UCI server at https:// archive.ics.uci.edu/ml/machine-learning-databases/ breast-cancer-wisconsin/wdbc.data is temporarily

unavailable. For instance, to load the dataset from a local directory, you can replace the following lines: df = pd.read_csv( 'https://archive.ics.uci.edu/ml/' 'machine-learning-databases' '/breast-cancer-wisconsin/wdbc.data', header=None)

with these: df = pd.read_csv( 'your/local/path/to/wdbc.data', header=None)

In this section, we will read in the dataset and split it into training and test datasets in three simple steps: 1. We will start by reading in the dataset directly from the UCI website using pandas: >>> import pandas as pd >>> df = pd.read_csv('https://archive.ics.uci.edu/ml/' ... 'machine-learning-databases' ... '/breast-cancer-wisconsin/wdbc.data', ... header=None) [ 192 ]

Chapter 6

2. Next, we will assign the 30 features to a NumPy array, X. Using a LabelEncoder object, we will transform the class labels from their original string representation ('M' and 'B') into integers: >>> from sklearn.preprocessing import LabelEncoder >>> X = df.loc[:, 2:].values >>> y = df.loc[:, 1].values >>> le = LabelEncoder() >>> y = le.fit_transform(y) >>> le.classes_ array(['B', 'M'], dtype=object)

After encoding the class labels (diagnosis) in an array, y, the malignant tumors are now represented as class 1, and the benign tumors are represented as class 0, respectively. We can double-check this mapping by calling the transform method of the fitted LabelEncoder on two dummy class labels: >>> le.transform(['M', 'B']) array([1, 0])

3. Before we construct our first model pipeline in the following subsection, let's divide the dataset into a separate training dataset (80 percent of the data) and a separate test dataset (20 percent of the data): >>> from sklearn.model_selection import train_test_split >>> X_train, X_test, y_train, y_test = \ ... train_test_split(X, y, ... test_size=0.20, ... stratify=y, ... random_state=1)

Combining transformers and estimators in a pipeline

In the previous chapter, you learned that many learning algorithms require input features on the same scale for optimal performance. Since the features in the Breast Cancer Wisconsin dataset are measured on various different scales, we will standardize the columns in the Breast Cancer Wisconsin dataset before we feed them to a linear classifier, such as logistic regression. Furthermore, let's assume that we want to compress our data from the initial 30 dimensions onto a lower twodimensional subspace via principal component analysis (PCA), a feature extraction technique for dimensionality reduction that was introduced in Chapter 5, Compressing Data via Dimensionality Reduction. [ 193 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

Instead of going through the model fitting and data transformation steps for the training and test datasets separately, we can chain the StandardScaler, PCA, and LogisticRegression objects in a pipeline: >>> from sklearn.preprocessing import StandardScaler >>> from sklearn.decomposition import PCA >>> from sklearn.linear_model import LogisticRegression >>> from sklearn.pipeline import make_pipeline >>> pipe_lr = make_pipeline(StandardScaler(), ... PCA(n_components=2), ... LogisticRegression(random_state=1, ... solver='lbfgs')) >>> pipe_lr.fit(X_train, y_train) >>> y_pred = pipe_lr.predict(X_test) >>> print('Test Accuracy: %.3f' % pipe_lr.score(X_test, y_test)) Test Accuracy: 0.956

The make_pipeline function takes an arbitrary number of scikit-learn transformers (objects that support the fit and transform methods as input), followed by a scikitlearn estimator that implements the fit and predict methods. In our preceding code example, we provided two transformers, StandardScaler and PCA, and a LogisticRegression estimator as inputs to the make_pipeline function, which constructs a scikit-learn Pipeline object from these objects. We can think of a scikit-learn Pipeline as a meta-estimator or wrapper around those individual transformers and estimators. If we call the fit method of Pipeline, the data will be passed down a series of transformers via fit and transform calls on these intermediate steps until it arrives at the estimator object (the final element in a pipeline). The estimator will then be fitted to the transformed training data. When we executed the fit method on the pipe_lr pipeline in the preceding code example, StandardScaler first performed fit and transform calls on the training data. Second, the transformed training data was passed on to the next object in the pipeline, PCA. Similar to the previous step, PCA also executed fit and transform on the scaled input data and passed it to the final element of the pipeline, the estimator. Finally, the LogisticRegression estimator was fit to the training data after it underwent transformations via StandardScaler and PCA. Again, we should note that there is no limit to the number of intermediate steps in a pipeline; however, the last pipeline element has to be an estimator. Similar to calling fit on a pipeline, pipelines also implement a predict method. If we feed a dataset to the predict call of a Pipeline object instance, the data will pass through the intermediate steps via transform calls. In the final step, the estimator object will then return a prediction on the transformed data. [ 194 ]

Chapter 6

The pipelines of the scikit-learn library are immensely useful wrapper tools, which we will use frequently throughout the rest of this book. To make sure that you've got a good grasp of how the Pipeline object works, please take a close look at the following illustration, which summarizes our discussion from the previous paragraphs:

Using k-fold cross-validation to assess model performance

One of the key steps in building a machine learning model is to estimate its performance on data that the model hasn't seen before. Let's assume that we fit our model on a training dataset and use the same data to estimate how well it performs on new data. We remember from the Tackling overfitting via regularization section in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, that a model can suffer from underfitting (high bias) if the model is too simple, or it can overfit the training data (high variance) if the model is too complex for the underlying training data. To find an acceptable bias-variance tradeoff, we need to evaluate our model carefully. In this section, you will learn about the common cross-validation techniques holdout cross-validation and k-fold cross-validation, which can help us to obtain reliable estimates of the model's generalization performance, that is, how well the model performs on unseen data. [ 195 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

The holdout method

A classic and popular approach for estimating the generalization performance of machine learning models is holdout cross-validation. Using the holdout method, we split our initial dataset into separate training and test datasets—the former is used for model training, and the latter is used to estimate its generalization performance. However, in typical machine learning applications, we are also interested in tuning and comparing different parameter settings to further improve the performance for making predictions on unseen data. This process is called model selection, with the name referring to a given classification problem for which we want to select the optimal values of tuning parameters (also called hyperparameters). However, if we reuse the same test dataset over and over again during model selection, it will become part of our training data and thus the model will be more likely to overfit. Despite this issue, many people still use the test dataset for model selection, which is not a good machine learning practice. A better way of using the holdout method for model selection is to separate the data into three parts: a training dataset, a validation dataset, and a test dataset. The training dataset is used to fit the different models, and the performance on the validation dataset is then used for the model selection. The advantage of having a test dataset that the model hasn't seen before during the training and model selection steps is that we can obtain a less biased estimate of its ability to generalize to new data. The following figure illustrates the concept of holdout cross-validation, where we use a validation dataset to repeatedly evaluate the performance of the model after training using different hyperparameter values. Once we are satisfied with the tuning of hyperparameter values, we estimate the model's generalization performance on the test dataset:

[ 196 ]

Chapter 6

A disadvantage of the holdout method is that the performance estimate may be very sensitive to how we partition the training dataset into the training and validation subsets; the estimate will vary for different examples of the data. In the next subsection, we will take a look at a more robust technique for performance estimation, k-fold cross-validation, where we repeat the holdout method k times on k subsets of the training data.

K-fold cross-validation

In k-fold cross-validation, we randomly split the training dataset into k folds without replacement, where k – 1 folds are used for the model training, and one fold is used for performance evaluation. This procedure is repeated k times so that we obtain k models and performance estimates. Sampling with and without replacement We looked at an example to illustrate sampling with and without replacement in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn. If you haven't read that chapter, or want a refresher, refer to the information box titled Sampling with and without replacement in the Combining multiple decision trees via random forests section.

We then calculate the average performance of the models based on the different, independent test folds to obtain a performance estimate that is less sensitive to the sub-partitioning of the training data compared to the holdout method. Typically, we use k-fold cross-validation for model tuning, that is, finding the optimal hyperparameter values that yield a satisfying generalization performance, which is estimated from evaluating the model performance on the test folds. Once we have found satisfactory hyperparameter values, we can retrain the model on the complete training dataset and obtain a final performance estimate using the independent test dataset. The rationale behind fitting a model to the whole training dataset after k-fold cross-validation is that providing more training examples to a learning algorithm usually results in a more accurate and robust model. Since k-fold cross-validation is a resampling technique without replacement, the advantage of this approach is that each example will be used for training and validation (as part of a test fold) exactly once, which yields a lower-variance estimate of the model performance than the holdout method. The following figure summarizes the concept behind k-fold cross-validation with k = 10. The training dataset is divided into 10 folds, and during the 10 iterations, nine folds are used for training, and one fold will be used as the test dataset for the model evaluation. [ 197 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

Also, the estimated performances, 𝐸𝐸𝑖𝑖 (for example, classification accuracy or error), for each fold are then used to calculate the estimated average performance, E, of the model:

A good standard value for k in k-fold cross-validation is 10, as empirical evidence shows. For instance, experiments by Ron Kohavi on various real-world datasets suggest that 10-fold cross-validation offers the best tradeoff between bias and variance (A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection, Kohavi, Ron, International Joint Conference on Artificial Intelligence (IJCAI), 14 (12): 1137-43, 1995). However, if we are working with relatively small training sets, it can be useful to increase the number of folds. If we increase the value of k, more training data will be used in each iteration, which results in a lower pessimistic bias toward estimating the generalization performance by averaging the individual model estimates. However, large values of k will also increase the runtime of the cross-validation algorithm and yield estimates with higher variance, since the training folds will be more similar to each other. On the other hand, if we are working with large datasets, we can choose a smaller value for k, for example, k = 5, and still obtain an accurate estimate of the average performance of the model while reducing the computational cost of refitting and evaluating the model on the different folds.

[ 198 ]

Chapter 6

Leave-one-out cross-validation A special case of k-fold cross-validation is the leave-one-out cross-validation (LOOCV) method. In LOOCV, we set the number of folds equal to the number of training examples (k = n) so that only one training example is used for testing during each iteration, which is a recommended approach for working with very small datasets.

A slight improvement over the standard k-fold cross-validation approach is stratified k-fold cross-validation, which can yield better bias and variance estimates, especially in cases of unequal class proportions, which has also been been shown in the same study by Ron Kohavi referenced previously in this section. In stratified crossvalidation, the class label proportions are preserved in each fold to ensure that each fold is representative of the class proportions in the training dataset, which we will illustrate by using the StratifiedKFold iterator in scikit-learn: >>> import numpy as np >>> from sklearn.model_selection import StratifiedKFold >>> kfold = StratifiedKFold(n_splits=10).split(X_train, y_train) >>> scores = [] >>> for k, (train, test) in enumerate(kfold): ... pipe_lr.fit(X_train[train], y_train[train]) ... score = pipe_lr.score(X_train[test], y_train[test]) ... scores.append(score) ... print('Fold: %2d, Class dist.: %s, Acc: %.3f' % (k+1, ... np.bincount(y_train[train]), score)) Fold: 1, Class dist.: [256 153], Acc: 0.935 Fold: 2, Class dist.: [256 153], Acc: 0.935 Fold: 3, Class dist.: [256 153], Acc: 0.957 Fold: 4, Class dist.: [256 153], Acc: 0.957 Fold: 5, Class dist.: [256 153], Acc: 0.935 Fold: 6, Class dist.: [257 153], Acc: 0.956 Fold: 7, Class dist.: [257 153], Acc: 0.978 Fold: 8, Class dist.: [257 153], Acc: 0.933 Fold: 9, Class dist.: [257 153], Acc: 0.956 Fold: 10, Class dist.: [257 153], Acc: 0.956 >>> print('\nCV accuracy: %.3f +/- %.3f' % ... (np.mean(scores), np.std(scores))) CV accuracy: 0.950 +/- 0.014

[ 199 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

First, we initialized the StratifiedKFold iterator from the sklearn.model_selection module with the y_train class labels in the training dataset, and we specified the number of folds via the n_splits parameter. When we used the kfold iterator to loop through the k folds, we used the returned indices in train to fit the logistic regression pipeline that we set up at the beginning of this chapter. Using the pipe_lr pipeline, we ensured that the examples were scaled properly (for instance, standardized) in each iteration. We then used the test indices to calculate the accuracy score of the model, which we collected in the scores list to calculate the average accuracy and the standard deviation of the estimate. Although the previous code example was useful to illustrate how k-fold crossvalidation works, scikit-learn also implements a k-fold cross-validation scorer, which allows us to evaluate our model using stratified k-fold cross-validation less verbosely: >>> from sklearn.model_selection import cross_val_score >>> scores = cross_val_score(estimator=pipe_lr, ... X=X_train, ... y=y_train, ... cv=10, ... n_jobs=1) >>> print('CV accuracy scores: %s' % scores) CV accuracy scores: [ 0.93478261 0.93478261 0.95652174 0.95652174 0.93478261 0.95555556 0.97777778 0.93333333 0.95555556 0.95555556] >>> print('CV accuracy: %.3f +/- %.3f' % (np.mean(scores), ... np.std(scores))) CV accuracy: 0.950 +/- 0.014

An extremely useful feature of the cross_val_score approach is that we can distribute the evaluation of the different folds across multiple central processing units (CPUs) on our machine. If we set the n_jobs parameter to 1, only one CPU will be used to evaluate the performances, just like in our StratifiedKFold example previously. However, by setting n_jobs=2, we could distribute the 10 rounds of cross-validation to two CPUs (if available on our machine), and by setting n_jobs=-1, we can use all available CPUs on our machine to do the computation in parallel.

[ 200 ]

Chapter 6

Estimating generalization performance Please note that a detailed discussion of how the variance of the generalization performance is estimated in crossvalidation is beyond the scope of this book, but you can refer to a comprehensive article about model evaluation and crossvalidation (Model evaluation, model selection, and algorithm selection in machine learning. Raschka S. arXiv preprint arXiv:1811.12808, 2018) that discusses these topics in more depth. The article is freely available from https://arxiv.org/abs/1811.12808. In addition, you can find a detailed discussion in this excellent article by M. Markatou and others (Analysis of Variance of Crossvalidation Estimators of the Generalization Error, M. Markatou, H. Tian, S. Biswas, and G. M. Hripcsak, Journal of Machine Learning Research, 6: 1127-1168, 2005). You can also read about alternative cross-validation techniques, such as the .632 Bootstrap cross-validation method (Improvements on Cross-validation: The .632+ Bootstrap Method, B. Efron and R. Tibshirani, Journal of the American Statistical Association, 92(438): 548-560, 1997).

Debugging algorithms with learning and validation curves

In this section, we will take a look at two very simple yet powerful diagnostic tools that can help us to improve the performance of a learning algorithm: learning curves and validation curves. In the next subsections, we will discuss how we can use learning curves to diagnose whether a learning algorithm has a problem with overfitting (high variance) or underfitting (high bias). Furthermore, we will take a look at validation curves that can help us to address the common issues of a learning algorithm.

Diagnosing bias and variance problems with learning curves

If a model is too complex for a given training dataset—there are too many degrees of freedom or parameters in this model—the model tends to overfit the training data and does not generalize well to unseen data. Often, it can help to collect more training examples to reduce the degree of overfitting. [ 201 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

However, in practice, it can often be very expensive or simply not feasible to collect more data. By plotting the model training and validation accuracies as functions of the training dataset size, we can easily detect whether the model suffers from high variance or high bias, and whether the collection of more data could help to address this problem. But before we discuss how to plot learning curves in scikit-learn, let's discuss those two common model issues by walking through the following illustration:

The graph in the upper-left shows a model with high bias. This model has both low training and cross-validation accuracy, which indicates that it underfits the training data. Common ways to address this issue are to increase the number of parameters of the model, for example, by collecting or constructing additional features, or by decreasing the degree of regularization, for example, in support vector machine (SVM) or logistic regression classifiers. The graph in the upper-right shows a model that suffers from high variance, which is indicated by the large gap between the training and cross-validation accuracy. To address this problem of overfitting, we can collect more training data, reduce the complexity of the model, or increase the regularization parameter, for example.

[ 202 ]

Chapter 6

For unregularized models, it can also help to decrease the number of features via feature selection (Chapter 4, Building Good Training Datasets – Data Preprocessing) or feature extraction (Chapter 5, Compressing Data via Dimensionality Reduction) to decrease the degree of overfitting. While collecting more training data usually tends to decrease the chance of overfitting, it may not always help, for example, if the training data is extremely noisy or the model is already very close to optimal. In the next subsection, we will see how to address those model issues using validation curves, but let's first see how we can use the learning curve function from scikit-learn to evaluate the model: >>> import matplotlib.pyplot as plt >>> from sklearn.model_selection import learning_curve >>> ... ... ... ... >>> ... ... ... ... ... ... ... >>> >>> >>> >>>

pipe_lr = make_pipeline(StandardScaler(), LogisticRegression(penalty='l2', random_state=1, solver='lbfgs', max_iter=10000)) train_sizes, train_scores, test_scores =\ learning_curve(estimator=pipe_lr, X=X_train, y=y_train, train_sizes=np.linspace( 0.1, 1.0, 10), cv=10, n_jobs=1) train_mean = np.mean(train_scores, axis=1) train_std = np.std(train_scores, axis=1) test_mean = np.mean(test_scores, axis=1) test_std = np.std(test_scores, axis=1)

>>> plt.plot(train_sizes, train_mean, ... color='blue', marker='o', ... markersize=5, label='Training accuracy') >>> plt.fill_between(train_sizes, ... train_mean + train_std, ... train_mean - train_std, ... alpha=0.15, color='blue') >>> plt.plot(train_sizes, test_mean, ... color='green', linestyle='--', ... marker='s', markersize=5, ... label='Validation accuracy')

[ 203 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning >>> ... ... ... >>> >>> >>> >>> >>> >>>

plt.fill_between(train_sizes, test_mean + test_std, test_mean - test_std, alpha=0.15, color='green') plt.grid() plt.xlabel('Number of training examples') plt.ylabel('Accuracy') plt.legend(loc='lower right') plt.ylim([0.8, 1.03]) plt.show()

Note that we passed max_iter=10000 as an additional argument when instantiating the LogisticRegression object (which uses 1,000 iterations as a default) to avoid convergence issues for the smaller dataset sizes or extreme regularization parameter values (covered in the next section). After we have successfully executed the preceding code, we will obtain the following learning curve plot:

Via the train_sizes parameter in the learning_curve function, we can control the absolute or relative number of training examples that are used to generate the learning curves. Here, we set train_sizes=np.linspace(0.1, 1.0, 10) to use 10 evenly spaced, relative intervals for the training dataset sizes. By default, the learning_curve function uses stratified k-fold cross-validation to calculate the cross-validation accuracy of a classifier, and we set k=10 via the cv parameter for 10-fold stratified cross-validation.

[ 204 ]

Chapter 6

Then, we simply calculated the average accuracies from the returned cross-validated training and test scores for the different sizes of the training dataset, which we plotted using Matplotlib's plot function. Furthermore, we added the standard deviation of the average accuracy to the plot using the fill_between function to indicate the variance of the estimate. As we can see in the preceding learning curve plot, our model performs quite well on both the training and validation datasets if it has seen more than 250 examples during training. We can also see that the training accuracy increases for training datasets with fewer than 250 examples, and the gap between validation and training accuracy widens—an indicator of an increasing degree of overfitting.

Addressing over- and underfitting with validation curves

Validation curves are a useful tool for improving the performance of a model by addressing issues such as overfitting or underfitting. Validation curves are related to learning curves, but instead of plotting the training and test accuracies as functions of the sample size, we vary the values of the model parameters, for example, the inverse regularization parameter, C, in logistic regression. Let's go ahead and see how we create validation curves via scikit-learn: >>> >>> >>> ... ... ... ... ... ... >>> >>> >>> >>> >>> ... ... >>> ... ... >>> ... ... ...

from sklearn.model_selection import validation_curve param_range = [0.001, 0.01, 0.1, 1.0, 10.0, 100.0] train_scores, test_scores = validation_curve( estimator=pipe_lr, X=X_train, y=y_train, param_name='logisticregression__C', param_range=param_range, cv=10) train_mean = np.mean(train_scores, axis=1) train_std = np.std(train_scores, axis=1) test_mean = np.mean(test_scores, axis=1) test_std = np.std(test_scores, axis=1) plt.plot(param_range, train_mean, color='blue', marker='o', markersize=5, label='Training accuracy') plt.fill_between(param_range, train_mean + train_std, train_mean - train_std, alpha=0.15, color='blue') plt.plot(param_range, test_mean, color='green', linestyle='--', marker='s', markersize=5, label='Validation accuracy') [ 205 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning >>> ... ... ... >>> >>> >>> >>> >>> >>> >>>

plt.fill_between(param_range, test_mean + test_std, test_mean - test_std, alpha=0.15, color='green') plt.grid() plt.xscale('log') plt.legend(loc='lower right') plt.xlabel('Parameter C') plt.ylabel('Accuracy') plt.ylim([0.8, 1.0]) plt.show()

Using the preceding code, we obtained the validation curve plot for the parameter C:

Similar to the learning_curve function, the validation_curve function uses stratified k-fold cross-validation by default to estimate the performance of the classifier. Inside the validation_curve function, we specified the parameter that we wanted to evaluate. In this case, it is C, the inverse regularization parameter of the LogisticRegression classifier, which we wrote as 'logisticregression__C' to access the LogisticRegression object inside the scikit-learn pipeline for a specified value range that we set via the param_range parameter. Similar to the learning curve example in the previous section, we plotted the average training and cross-validation accuracies and the corresponding standard deviations.

[ 206 ]

Chapter 6

Although the differences in the accuracy for varying values of C are subtle, we can see that the model slightly underfits the data when we increase the regularization strength (small values of C). However, for large values of C, it means lowering the strength of regularization, so the model tends to slightly overfit the data. In this case, the sweet spot appears to be between 0.01 and 0.1 of the C value.

Fine-tuning machine learning models via grid search

In machine learning, we have two types of parameters: those that are learned from the training data, for example, the weights in logistic regression, and the parameters of a learning algorithm that are optimized separately. The latter are the tuning parameters (or hyperparameters) of a model, for example, the regularization parameter in logistic regression or the depth parameter of a decision tree. In the previous section, we used validation curves to improve the performance of a model by tuning one of its hyperparameters. In this section, we will take a look at a popular hyperparameter optimization technique called grid search, which can further help to improve the performance of a model by finding the optimal combination of hyperparameter values.

Tuning hyperparameters via grid search

The grid search approach is quite simple: it's a brute-force exhaustive search paradigm where we specify a list of values for different hyperparameters, and the computer evaluates the model performance for each combination to obtain the optimal combination of values from this set: >>> from sklearn.model_selection import GridSearchCV >>> from sklearn.svm import SVC >>> pipe_svc = make_pipeline(StandardScaler(), ... SVC(random_state=1)) >>> param_range = [0.0001, 0.001, 0.01, 0.1, ... 1.0, 10.0, 100.0, 1000.0] >>> param_grid = [{'svc__C': param_range, ... 'svc__kernel': ['linear']}, ... {'svc__C': param_range, ... 'svc__gamma': param_range, ... 'svc__kernel': ['rbf']}] >>> gs = GridSearchCV(estimator=pipe_svc, ... param_grid=param_grid, [ 207 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning ... scoring='accuracy', ... cv=10, ... refit=True, ... n_jobs=-1) >>> gs = gs.fit(X_train, y_train) >>> print(gs.best_score_) 0.9846153846153847 >>> print(gs.best_params_) {'svc__C': 100.0, 'svc__gamma': 0.001, 'svc__kernel': 'rbf'}

Using the preceding code, we initialized a GridSearchCV object from the sklearn. model_selection module to train and tune an SVM pipeline. We set the param_grid parameter of GridSearchCV to a list of dictionaries to specify the parameters that we'd want to tune. For the linear SVM, we only evaluated the inverse regularization parameter, C; for the RBF kernel SVM, we tuned both the svc__C and svc__gamma parameters. Note that the svc__gamma parameter is specific to kernel SVMs. After we used the training data to perform the grid search, we obtained the score of the best-performing model via the best_score_ attribute and looked at its parameters, which can be accessed via the best_params_ attribute. In this particular case, the RBF kernel SVM model with svc__C = 100.0 yielded the best k-fold crossvalidation accuracy: 98.5 percent. Finally, we use the independent test dataset to estimate the performance of the best-selected model, which is available via the best_estimator_ attribute of the GridSearchCV object: >>> clf = gs.best_estimator_ >>> clf.fit(X_train, y_train) >>> print('Test accuracy: %.3f' % clf.score(X_test, y_test)) Test accuracy: 0.974

Please note that fitting a model with the best settings (gs.best_estimator_) on the training set manually via clf.fit(X_train, y_train) after completing the grid search is not necessary. The GridSearchCV class has a refit parameter, which will refit the gs.best_estimator_ to the whole training set automatically if we set refit=True (default).

[ 208 ]

Chapter 6

Randomized hyperparameter search Although grid search is a powerful approach for finding the optimal set of parameters, the evaluation of all possible parameter combinations is also computationally very expensive. An alternative approach for sampling different parameter combinations using scikit-learn is randomized search. Randomized search usually performs about as well as grid search but is much more cost- and time-effective. In particular, if we only sample 60 parameter combinations via randomized search, we already have a 95 percent probability of obtaining solutions within 5 percent of the optimal performance (Random search for hyper-parameter optimization. Bergstra J, Bengio Y. Journal of Machine Learning Research. pp. 281-305, 2012). Using the RandomizedSearchCV class in scikit-learn, we can draw random parameter combinations from sampling distributions with a specified budget. More details and examples of its usage can be found at http://scikit-learn. org/stable/modules/grid_search.html#randomizedparameter-optimization.

Algorithm selection with nested crossvalidation

Using k-fold cross-validation in combination with grid search is a useful approach for fine-tuning the performance of a machine learning model by varying its hyperparameter values, as we saw in the previous subsection. If we want to select among different machine learning algorithms, though, another recommended approach is nested cross-validation. In a nice study on the bias in error estimation, Sudhir Varma and Richard Simon concluded that the true error of the estimate is almost unbiased relative to the test dataset when nested cross-validation is used (Bias in Error Estimation When Using Cross-Validation for Model Selection, BMC Bioinformatics, S. Varma and R. Simon, 7(1): 91, 2006).

[ 209 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

In nested cross-validation, we have an outer k-fold cross-validation loop to split the data into training and test folds, and an inner loop is used to select the model using k-fold cross-validation on the training fold. After model selection, the test fold is then used to evaluate the model performance. The following figure explains the concept of nested cross-validation with only five outer and two inner folds, which can be useful for large datasets where computational performance is important; this particular type of nested cross-validation is also known as 5x2 cross-validation:

In scikit-learn, we can perform nested cross-validation as follows: >>> gs = GridSearchCV(estimator=pipe_svc, ... param_grid=param_grid, ... scoring='accuracy', ... cv=2) >>> scores = cross_val_score(gs, X_train, y_train, ... scoring='accuracy', cv=5) >>> print('CV accuracy: %.3f +/- %.3f' % (np.mean(scores), ... np.std(scores))) CV accuracy: 0.974 +/- 0.015

[ 210 ]

Chapter 6

The returned average cross-validation accuracy gives us a good estimate of what to expect if we tune the hyperparameters of a model and use it on unseen data. For example, we can use the nested cross-validation approach to compare an SVM model to a simple decision tree classifier; for simplicity, we will only tune its depth parameter: >>> from sklearn.tree import DecisionTreeClassifier >>> gs = GridSearchCV(estimator=DecisionTreeClassifier( ... random_state=0), ... param_grid=[{'max_depth': [1, 2, 3, ... 4, 5, 6, ... 7, None]}], ... scoring='accuracy', ... cv=2) >>> scores = cross_val_score(gs, X_train, y_train, ... scoring='accuracy', cv=5) >>> print('CV accuracy: %.3f +/- %.3f' % (np.mean(scores), ... np.std(scores))) CV accuracy: 0.934 +/- 0.016

As we can see, the nested cross-validation performance of the SVM model (97.4 percent) is notably better than the performance of the decision tree (93.4 percent), and thus, we'd expect that it might be the better choice to classify new data that comes from the same population as this particular dataset.

Looking at different performance evaluation metrics

In the previous sections and chapters, we evaluated different machine learning models using the prediction accuracy, which is a useful metric with which to quantify the performance of a model in general. However, there are several other performance metrics that can be used to measure a model's relevance, such as precision, recall, and the F1 score.

Reading a confusion matrix

Before we get into the details of different scoring metrics, let's take a look at a confusion matrix, a matrix that lays out the performance of a learning algorithm.

[ 211 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

A confusion matrix is simply a square matrix that reports the counts of the true positive (TP), true negative (TN), false positive (FP), and false negative (FN) predictions of a classifier, as shown in the following figure:

Although these metrics can be easily computed manually by comparing the true and predicted class labels, scikit-learn provides a convenient confusion_matrix function that we can use, as follows: >>> from sklearn.metrics import confusion_matrix >>> pipe_svc.fit(X_train, y_train) >>> y_pred = pipe_svc.predict(X_test) >>> confmat = confusion_matrix(y_true=y_test, y_pred=y_pred) >>> print(confmat) [[71 1] [ 2 40]]

The array that was returned after executing the code provides us with information about the different types of error the classifier made on the test dataset. We can map this information onto the confusion matrix illustration in the previous figure using Matplotlib's matshow function: >>> >>> >>> ... ... ... ... >>> >>> >>>

fig, ax = plt.subplots(figsize=(2.5, 2.5)) ax.matshow(confmat, cmap=plt.cm.Blues, alpha=0.3) for i in range(confmat.shape[0]): for j in range(confmat.shape[1]): ax.text(x=j, y=i, s=confmat[i, j], va='center', ha='center') plt.xlabel('Predicted label') plt.ylabel('True label') plt.show()

[ 212 ]

Chapter 6

Now, the following confusion matrix plot, with the added labels, should make the results a little bit easier to interpret:

Assuming that class 1 (malignant) is the positive class in this example, our model correctly classified 71 of the examples that belong to class 0 (TN) and 40 examples that belong to class 1 (TP), respectively. However, our model also incorrectly misclassified two examples from class 1 as class 0 (FN), and it predicted that one example is malignant although it is a benign tumor (FP). In the next subsection, we will learn how we can use this information to calculate various error metrics.

Optimizing the precision and recall of a classification model

Both the prediction error (ERR) and accuracy (ACC) provide general information about how many examples are misclassified. The error can be understood as the sum of all false predictions divided by the number of total predictions, and the accuracy is calculated as the sum of correct predictions divided by the total number of predictions, respectively: 𝐸𝐸𝐸𝐸𝐸𝐸 =

𝐹𝐹𝐹𝐹 + 𝐹𝐹𝐹𝐹 𝐹𝐹𝐹𝐹 + 𝐹𝐹𝐹𝐹 + 𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇

The prediction accuracy can then be calculated directly from the error: 𝐴𝐴𝐴𝐴𝐴𝐴 =

𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇 = 1 − 𝐸𝐸𝐸𝐸𝐸𝐸 𝐹𝐹𝐹𝐹 + 𝐹𝐹𝐹𝐹 + 𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇

[ 213 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

The true positive rate (TPR) and false positive rate (FPR) are performance metrics that are especially useful for imbalanced class problems: 𝐹𝐹𝐹𝐹𝐹𝐹 = 𝑇𝑇𝑇𝑇𝑇𝑇 =

𝐹𝐹𝐹𝐹 𝐹𝐹𝐹𝐹 = 𝑁𝑁 𝐹𝐹𝐹𝐹 + 𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇 = 𝑃𝑃 𝐹𝐹𝐹𝐹 + 𝑇𝑇𝑇𝑇

In tumor diagnosis, for example, we are more concerned about the detection of malignant tumors in order to help a patient with the appropriate treatment. However, it is also important to decrease the number of benign tumors incorrectly classified as malignant (FP) to not unnecessarily concern patients. In contrast to the FPR, the TPR provides useful information about the fraction of positive (or relevant) examples that were correctly identified out of the total pool of positives (P). The performance metrics precision (PRE) and recall (REC) are related to those TP and TN rates, and in fact, REC is synonymous with TPR: 𝑃𝑃𝑃𝑃𝑃𝑃

=

𝑅𝑅𝑅𝑅𝑅𝑅 = 𝑇𝑇𝑇𝑇𝑇𝑇

=

𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇 + 𝐹𝐹𝐹𝐹

𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇 = 𝑃𝑃 𝐹𝐹𝐹𝐹 + 𝑇𝑇𝑇𝑇

Revisiting the malignant tumor detection example, optimizing for recall helps with minimizing the chance of not detecting a malignant tumor. However, this comes at the cost of predicting malignant tumors in patients although the patients are healthy (a high number of FP). If we optimize for precision, on the other hand, we emphasize correctness if we predict that a patient has a malignant tumor. However, this comes at the cost of missing malignant tumors more frequently (a high number of FN). To balance the up- and down-sides of optimizing PRE and REC, often a combination of PRE and REC is used, the so-called F1 score:

𝐹𝐹1 = 2

𝑃𝑃𝑃𝑃𝑃𝑃 × 𝑅𝑅𝑅𝑅𝑅𝑅 𝑃𝑃𝑃𝑃𝑃𝑃 + 𝑅𝑅𝑅𝑅𝑅𝑅

[ 214 ]

Chapter 6

Further reading on precision and recall If you are interested in a more thorough discussion of the different performance metrics, such as precision and recall, read David M. W. Powers' technical report Evaluation: From Precision, Recall and F-Factor to ROC, Informedness, Markedness & Correlation, which is freely available at http://www.flinders.edu.au/ science_engineering/fms/School-CSEM/publications/ tech_reps-research_artfcts/TRRA_2007.pdf.

Those scoring metrics are all implemented in scikit-learn and can be imported from the sklearn.metrics module as shown in the following snippet: >>> from sklearn.metrics import precision_score >>> from sklearn.metrics import recall_score, f1_score >>> print('Precision: %.3f' % precision_score( ... y_true=y_test, y_pred=y_pred)) Precision: 0.976 >>> print('Recall: %.3f' % recall_score( ... y_true=y_test, y_pred=y_pred)) Recall: 0.952 >>> print('F1: %.3f' % f1_score( ... y_true=y_test, y_pred=y_pred)) F1: 0.964

Furthermore, we can use a different scoring metric than accuracy in the GridSearchCV via the scoring parameter. A complete list of the different values that are accepted by the scoring parameter can be found at http://scikit-learn.org/ stable/modules/model_evaluation.html. Remember that the positive class in scikit-learn is the class that is labeled as class 1. If we want to specify a different positive label, we can construct our own scorer via the make_scorer function, which we can then directly provide as an argument to the scoring parameter in GridSearchCV (in this example, using the f1_score as a metric): >>> >>> >>> ... ... ... ... >>>

from sklearn.metrics import make_scorer, f1_score c_gamma_range = [0.01, 0.1, 1.0, 10.0] param_grid = [{'svc__C': c_gamma_range, 'svc__kernel': ['linear']}, {'svc__C': c_gamma_range, 'svc__gamma': c_gamma_range, 'svc__kernel': ['rbf']}] scorer = make_scorer(f1_score, pos_label=0) [ 215 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning >>> gs = GridSearchCV(estimator=pipe_svc, ... param_grid=param_grid, ... scoring=scorer, ... cv=10) >>> gs = gs.fit(X_train, y_train) >>> print(gs.best_score_) 0.986202145696 >>> print(gs.best_params_) {'svc__C': 10.0, 'svc__gamma': 0.01, 'svc__kernel': 'rbf'}

Plotting a receiver operating characteristic

Receiver operating characteristic (ROC) graphs are useful tools to select models for classification based on their performance with respect to the FPR and TPR, which are computed by shifting the decision threshold of the classifier. The diagonal of a ROC graph can be interpreted as random guessing, and classification models that fall below the diagonal are considered as worse than random guessing. A perfect classifier would fall into the top-left corner of the graph with a TPR of 1 and an FPR of 0. Based on the ROC curve, we can then compute the so-called ROC area under the curve (ROC AUC) to characterize the performance of a classification model. Similar to ROC curves, we can compute precision-recall curves for different probability thresholds of a classifier. A function for plotting those precision-recall curves is also implemented in scikit-learn and is documented at http://scikit-

learn.org/stable/modules/generated/sklearn.metrics.precision_recall_ curve.html.

Executing the following code example, we will plot a ROC curve of a classifier that only uses two features from the Breast Cancer Wisconsin dataset to predict whether a tumor is benign or malignant. Although we are going to use the same logistic regression pipeline that we defined previously, we are only using two features this time. This is to make the classification task more challenging for the classifier, by witholding useful information contained in the other features, so that the resulting ROC curve becomes visually more interesting. For similar reasons, we are also reducing the number of folds in the StratifiedKFold validator to three. The code is as follows: >>> from sklearn.metrics import roc_curve, auc >>> from scipy import interp >>> pipe_lr = make_pipeline(StandardScaler(), ... PCA(n_components=2), ... LogisticRegression(penalty='l2', ... random_state=1, ... solver='lbfgs', ... C=100.0)) [ 216 ]

Chapter 6

>>> X_train2 = X_train[:, [4, 14]] >>> cv = list(StratifiedKFold(n_splits=3, ... random_state=1).split(X_train, ... y_train)) >>> fig = plt.figure(figsize=(7, 5)) >>> mean_tpr = 0.0 >>> mean_fpr = np.linspace(0, 1, 100) >>> all_tpr = [] >>> for i, (train, test) in enumerate(cv): ... probas = pipe_lr.fit( ... X_train2[train], ... y_train[train]).predict_proba(X_train2[test]) ... fpr, tpr, thresholds = roc_curve(y_train[test], ... probas[:, 1], ... pos_label=1) ... mean_tpr += interp(mean_fpr, fpr, tpr) ... mean_tpr[0] = 0.0 ... roc_auc = auc(fpr, tpr) ... plt.plot(fpr, ... tpr, ... label='ROC fold %d (area = %0.2f)' ... % (i+1, roc_auc)) >>> plt.plot([0, 1], ... [0, 1], ... linestyle='--', ... color=(0.6, 0.6, 0.6), ... label='Random guessing') >>> >>> >>> >>> ... >>> ... ... ... ... >>> >>> >>> >>> >>> >>>

mean_tpr /= len(cv) mean_tpr[-1] = 1.0 mean_auc = auc(mean_fpr, mean_tpr) plt.plot(mean_fpr, mean_tpr, 'k--', label='Mean ROC (area = %0.2f)' % mean_auc, lw=2) plt.plot([0, 0, 1], [0, 1, 1], linestyle=':', color='black', label='Perfect performance') plt.xlim([-0.05, 1.05]) plt.ylim([-0.05, 1.05]) plt.xlabel('False positive rate') plt.ylabel('True positive rate') plt.legend(loc="lower right") plt.show() [ 217 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

In the preceding code example, we used the already familiar StratifiedKFold class from scikit-learn and calculated the ROC performance of the LogisticRegression classifier in our pipe_lr pipeline using the roc_curve function from the sklearn. metrics module separately for each iteration. Furthermore, we interpolated the average ROC curve from the three folds via the interp function that we imported from SciPy and calculated the area under the curve via the auc function. The resulting ROC curve indicates that there is a certain degree of variance between the different folds, and the average ROC AUC (0.76) falls between a perfect score (1.0) and random guessing (0.5):

Note that if we are just interested in the ROC AUC score, we could also directly import the roc_auc_score function from the sklearn.metrics submodule, which can be used similarly to the other scoring functions (for example, precision_score) that were introduced in the previous sections. Reporting the performance of a classifier as the ROC AUC can yield further insights into a classifier's performance with respect to imbalanced samples. However, while the accuracy score can be interpreted as a single cut-off point on an ROC curve, A. P. Bradley showed that the ROC AUC and accuracy metrics mostly agree with each other: The use of the area under the ROC curve in the evaluation of machine learning algorithms, A. P. Bradley, Pattern Recognition, 30(7): 1145-1159, 1997.

[ 218 ]

Chapter 6

Scoring metrics for multiclass classification

The scoring metrics that we've discussed so far are specific to binary classification systems. However, scikit-learn also implements macro and micro averaging methods to extend those scoring metrics to multiclass problems via one-vs.-all (OvA) classification. The micro-average is calculated from the individual TPs, TNs, FPs, and FNs of the system. For example, the micro-average of the precision score in a k-class system can be calculated as follows:

𝑃𝑃𝑃𝑃𝑃𝑃𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 =

𝑇𝑇𝑇𝑇1 + ⋯ + 𝑇𝑇𝑇𝑇𝑘𝑘 𝑇𝑇𝑇𝑇1 + ⋯ + 𝑇𝑇𝑇𝑇𝑘𝑘 + 𝐹𝐹𝐹𝐹1 + ⋯ + 𝐹𝐹𝐹𝐹𝑘𝑘

The macro-average is simply calculated as the average scores of the different systems: 𝑃𝑃𝑃𝑃𝑃𝑃𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 =

𝑃𝑃𝑃𝑃𝑃𝑃1 + ⋯ + 𝑃𝑃𝑃𝑃𝑃𝑃𝑘𝑘 𝑘𝑘

Micro-averaging is useful if we want to weight each instance or prediction equally, whereas macro-averaging weights all classes equally to evaluate the overall performance of a classifier with regard to the most frequent class labels. If we are using binary performance metrics to evaluate multiclass classification models in scikit-learn, a normalized or weighted variant of the macro-average is used by default. The weighted macro-average is calculated by weighting the score of each class label by the number of true instances when calculating the average. The weighted macro-average is useful if we are dealing with class imbalances, that is, different numbers of instances for each label. While the weighted macro-average is the default for multiclass problems in scikitlearn, we can specify the averaging method via the average parameter inside the different scoring functions that we import from the sklearn.metrics module, for example, the precision_score or make_scorer functions: >>> pre_scorer = make_scorer(score_func=precision_score, ... pos_label=1, ... greater_is_better=True, ... average='micro')

[ 219 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

Dealing with class imbalance

We've mentioned class imbalances several times throughout this chapter, and yet we haven't actually discussed how to deal with such scenarios appropriately if they occur. Class imbalance is a quite common problem when working with real-world data—examples from one class or multiple classes are over-represented in a dataset. We can think of several domains where this may occur, such as spam filtering, fraud detection, or screening for diseases. Imagine that the Breast Cancer Wisconsin dataset that we've been working with in this chapter consisted of 90 percent healthy patients. In this case, we could achieve 90 percent accuracy on the test dataset by just predicting the majority class (benign tumor) for all examples, without the help of a supervised machine learning algorithm. Thus, training a model on such a dataset that achieves approximately 90 percent test accuracy would mean our model hasn't learned anything useful from the features provided in this dataset. In this section, we will briefly go over some of the techniques that could help with imbalanced datasets. But before we discuss different methods to approach this problem, let's create an imbalanced dataset from our dataset, which originally consisted of 357 benign tumors (class 0) and 212 malignant tumors (class 1): >>> X_imb = np.vstack((X[y == 0], X[y == 1][:40])) >>> y_imb = np.hstack((y[y == 0], y[y == 1][:40]))

In this code snippet, we took all 357 benign tumor examples and stacked them with the first 40 malignant examples to create a stark class imbalance. If we were to compute the accuracy of a model that always predicts the majority class (benign, class 0), we would achieve a prediction accuracy of approximately 90 percent: >>> y_pred = np.zeros(y_imb.shape[0]) >>> np.mean(y_pred == y_imb) * 100 89.92443324937027

Thus, when we fit classifiers on such datasets, it would make sense to focus on other metrics than accuracy when comparing different models, such as precision, recall, the ROC curve—whatever we care most about in our application. For instance, our priority might be to identify the majority of patients with malignant cancer to recommend an additional screening, so recall should be our metric of choice. In spam filtering, where we don't want to label emails as spam if the system is not very certain, precision might be a more appropriate metric. Aside from evaluating machine learning models, class imbalance influences a learning algorithm during model fitting itself. Since machine learning algorithms typically optimize a reward or cost function that is computed as a sum over the training examples that it sees during fitting, the decision rule is likely going to be biased toward the majority class. [ 220 ]

Chapter 6

In other words, the algorithm implicitly learns a model that optimizes the predictions based on the most abundant class in the dataset, in order to minimize the cost or maximize the reward during training. One way to deal with imbalanced class proportions during model fitting is to assign a larger penalty to wrong predictions on the minority class. Via scikit-learn, adjusting such a penalty is as convenient as setting the class_weight parameter to class_weight='balanced', which is implemented for most classifiers. Other popular strategies for dealing with class imbalance include upsampling the minority class, downsampling the majority class, and the generation of synthetic training examples. Unfortunately, there's no universally best solution or technique that works best across different problem domains. Thus, in practice, it is recommended to try out different strategies on a given problem, evaluate the results, and choose the technique that seems most appropriate. The scikit-learn library implements a simple resample function that can help with the upsampling of the minority class by drawing new samples from the dataset with replacement. The following code will take the minority class from our imbalanced Breast Cancer Wisconsin dataset (here, class 1) and repeatedly draw new samples from it until it contains the same number of examples as class label 0: >>> from sklearn.utils import resample >>> print('Number of class 1 examples before:', ... X_imb[y_imb == 1].shape[0]) Number of class 1 examples before: 40 >>> X_upsampled, y_upsampled = resample( ... X_imb[y_imb == 1], ... y_imb[y_imb == 1], ... replace=True, ... n_samples=X_imb[y_imb == 0].shape[0], ... random_state=123) >>> print('Number of class 1 examples after:', ... X_upsampled.shape[0]) Number of class 1 examples after: 357

After resampling, we can then stack the original class 0 samples with the upsampled class 1 subset to obtain a balanced dataset as follows: >>> X_bal = np.vstack((X[y == 0], X_upsampled)) >>> y_bal = np.hstack((y[y == 0], y_upsampled))

Consequently, a majority vote prediction rule would only achieve 50 percent accuracy: >>> y_pred = np.zeros(y_bal.shape[0]) >>> np.mean(y_pred == y_bal) * 100 50 [ 221 ]

Learning Best Practices for Model Evaluation and Hyperparameter Tuning

Similarly, we could downsample the majority class by removing training examples from the dataset. To perform downsampling using the resample function, we could simply swap the class 1 label with class 0 in the previous code example and vice versa. Generating new training data to address class-imbalance Another technique for dealing with class imbalance is the generation of synthetic training examples, which is beyond the scope of this book. Probably the most widely used algorithm for synthetic training data generation is Synthetic Minority Oversampling Technique (SMOTE), and you can learn more about this technique in the original research article by Nitesh Chawla and others: SMOTE: Synthetic Minority Over-sampling Technique, Journal of Artificial Intelligence Research, 16: 321-357, 2002. It is also highly recommended to check out imbalanced-learn, a Python library that is entirely focused on imbalanced datasets, including an implementation of SMOTE. You can learn more about imbalanced-learn at https://github.com/scikitlearn-contrib/imbalanced-learn.

Summary

At the beginning of this chapter, we discussed how to chain different transformation techniques and classifiers in convenient model pipelines that help us to train and evaluate machine learning models more efficiently. We then used those pipelines to perform k-fold cross-validation, one of the essential techniques for model selection and evaluation. Using k-fold cross-validation, we plotted learning and validation curves to diagnose common problems of learning algorithms, such as overfitting and underfitting. Using grid search, we further fine-tuned our model. We then used confusion matrices and various performance metrics to evaluate and optimize a model's performance for specific problem tasks. Finally, we concluded this chapter by discussing different methods for dealing with imbalanced data, which is a common problem in many real-world applications. Now, you should be well-equipped with the essential techniques to build supervised machine learning models for classification successfully. In the next chapter, we will look at ensemble methods: methods that allow us to combine multiple models and classification algorithms to boost the predictive performance of a machine learning system even further.

[ 222 ]

Combining Different Models for Ensemble Learning In the previous chapter, we focused on the best practices for tuning and evaluating different models for classification. In this chapter, we will build upon those techniques and explore different methods for constructing a set of classifiers that can often have a better predictive performance than any of its individual members. We will learn how to do the following: • Make predictions based on majority voting • Use bagging to reduce overfitting by drawing random combinations of the training dataset with repetition • Apply boosting to build powerful models from weak learners that learn from their mistakes

Learning with ensembles

The goal of ensemble methods is to combine different classifiers into a metaclassifier that has better generalization performance than each individual classifier alone. For example, assuming that we collected predictions from 10 experts, ensemble methods would allow us to strategically combine those predictions by the 10 experts to come up with a prediction that was more accurate and robust than the predictions by each individual expert. As you will see later in this chapter, there are several different approaches for creating an ensemble of classifiers. This section will introduce a basic explanation of how ensembles work and why they are typically recognized for yielding a good generalization performance.

[ 223 ]

Combining Different Models for Ensemble Learning

In this chapter, we will focus on the most popular ensemble methods that use the majority voting principle. Majority voting simply means that we select the class label that has been predicted by the majority of classifiers, that is, received more than 50 percent of the votes. Strictly speaking, the term "majority vote" refers to binary class settings only. However, it is easy to generalize the majority voting principle to multiclass settings, which is called plurality voting. Here, we select the class label that received the most votes (the mode). The following diagram illustrates the concept of majority and plurality voting for an ensemble of 10 classifiers, where each unique symbol (triangle, square, and circle) represents a unique class label:

Using the training dataset, we start by training m different classifiers ( 𝐶𝐶1 , … , 𝐶𝐶𝑚𝑚 ). Depending on the technique, the ensemble can be built from different classification algorithms, for example, decision trees, support vector machines, logistic regression classifiers, and so on. Alternatively, we can also use the same base classification algorithm, fitting different subsets of the training dataset. One prominent example of this approach is the random forest algorithm, which combines different decision tree classifiers. The following figure illustrates the concept of a general ensemble approach using majority voting:

[ 224 ]

Chapter 7

To predict a class label via simple majority or plurality voting, we can combine the predicted class labels of each individual classifier, 𝐶𝐶𝑗𝑗, and select the class label, 𝑦𝑦̂, that received the most votes: 𝑦𝑦̂ = mode{𝐶𝐶1 (𝒙𝒙), 𝐶𝐶2 (𝒙𝒙), … , 𝐶𝐶𝑚𝑚 (𝒙𝒙)}

(In statistics, the mode is the most frequent event or result in a set. For example, mode{1, 2, 1, 1, 2, 4, 5, 4} = 1.) For example, in a binary classification task where class1 = –1 and class2 = +1, we can write the majority vote prediction as follows: 𝑚𝑚

𝐶𝐶(𝒙𝒙) = sign [∑ 𝐶𝐶𝑗𝑗 (𝒙𝒙)] = { 𝑗𝑗

1 if ∑ 𝐶𝐶𝑗𝑗 (𝒙𝒙) ≥ 0 𝑗𝑗

−1 otherwise

To illustrate why ensemble methods can work better than individual classifiers alone, let's apply the simple concepts of combinatorics. For the following example, we will make the assumption that all n-base classifiers for a binary classification task have an equal error rate, 𝜀𝜀. Furthermore, we will assume that the classifiers are independent and the error rates are not correlated. Under those assumptions, we can simply express the error probability of an ensemble of base classifiers as a probability mass function of a binomial distribution: 𝑛𝑛

𝑛𝑛 𝑃𝑃(𝑦𝑦 ≥ 𝑘𝑘) = ∑ ⟨ ⟩ 𝜀𝜀 𝑘𝑘 (1 − 𝜀𝜀)𝑛𝑛−𝑘𝑘 = 𝜀𝜀ensemble 𝑘𝑘 𝑘𝑘

𝑛𝑛 Here, ⟨ ⟩ is the binomial coefficient n choose k. In other words, we compute the 𝑘𝑘 probability that the prediction of the ensemble is wrong. Now, let's take a look at a more concrete example of 11 base classifiers (n = 11), where each classifier has an error rate of 0.25 (𝜀𝜀 = 0.25): 11

𝑃𝑃(𝑦𝑦 ≥ 𝑘𝑘) = ∑ ⟨ 𝑘𝑘=6

11 ⟩ 0.25𝑘𝑘 (1 − 0.25)11−𝑘𝑘 = 0.034 𝑘𝑘

[ 225 ]

Combining Different Models for Ensemble Learning

The binomial coefficient The binomial coefficient refers to the number of ways we can choose subsets of k-unordered elements from a set of size n; thus, it is often called "n choose k." Since the order does not matter here, the binomial coefficient is also sometimes referred to as combination or combinatorial number, and in its unabbreviated form, it is written as follows:

𝑛𝑛! (𝑛𝑛 − 𝑘𝑘)! 𝑘𝑘!

Here, the symbol (!) stands for factorial—for example,

3! = 3 ∙ 2 ∙ 1 = 6.

As you can see, the error rate of the ensemble (0.034) is much lower than the error rate of each individual classifier (0.25) if all the assumptions are met. Note that, in this simplified illustration, a 50-50 split by an even number of classifiers, n, is treated as an error, whereas this is only true half of the time. To compare such an idealistic ensemble classifier to a base classifier over a range of different base error rates, let's implement the probability mass function in Python: >>> from scipy.special import comb >>> import math >>> def ensemble_error(n_classifier, error): ... k_start = int(math.ceil(n_classifier / 2.)) ... probs = [comb(n_classifier, k) * ... error**k * ... (1-error)**(n_classifier - k) ... for k in range(k_start, n_classifier + 1)] ... return sum(probs) >>> ensemble_error(n_classifier=11, error=0.25) 0.03432750701904297

After we have implemented the ensemble_error function, we can compute the ensemble error rates for a range of different base errors from 0.0 to 1.0 to visualize the relationship between ensemble and base errors in a line graph: >>> >>> >>> >>> ... >>> ...

import numpy as np import matplotlib.pyplot as plt error_range = np.arange(0.0, 1.01, 0.01) ens_errors = [ensemble_error(n_classifier=11, error=error) for error in error_range] plt.plot(error_range, ens_errors, label='Ensemble error',

[ 226 ]

Chapter 7 ... >>> ... ... >>> >>> >>> >>> >>>

linewidth=2) plt.plot(error_range, error_range, linestyle='--', label='Base error', linewidth=2) plt.xlabel('Base error') plt.ylabel('Base/Ensemble error') plt.legend(loc='upper left') plt.grid(alpha=0.5) plt.show()

As you can see in the resulting plot, the error probability of an ensemble is always better than the error of an individual base classifier, as long as the base classifiers perform better than random guessing (𝜀𝜀 < 0.5). Note that the y axis depicts the base error (dotted line) as well as the ensemble error (continuous line):

Combining classifiers via majority vote

After the short introduction to ensemble learning in the previous section, let's start with a warm-up exercise and implement a simple ensemble classifier for majority voting in Python. Plurality voting Although the majority voting algorithm that we will discuss in this section also generalizes to multiclass settings via plurality voting, the term "majority voting" will be used for simplicity, as is often the case in the literature.

[ 227 ]

Combining Different Models for Ensemble Learning

Implementing a simple majority vote classifier The algorithm that we are going to implement in this section will allow us to combine different classification algorithms associated with individual weights for confidence. Our goal is to build a stronger meta-classifier that balances out the individual classifiers' weaknesses on a particular dataset. In more precise mathematical terms, we can write the weighted majority vote as follows: 𝑚𝑚

𝑦𝑦̂ = arg max ∑ 𝑤𝑤𝑗𝑗 𝜒𝜒𝐴𝐴 (𝐶𝐶𝑗𝑗 (𝒙𝒙) = 𝑖𝑖) 𝑖𝑖

𝑗𝑗=1

Here, 𝑤𝑤𝑗𝑗 is a weight associated with a base classifier, 𝐶𝐶𝑗𝑗 ; 𝑦𝑦̂ is the predicted class label of the ensemble; A is the set of unique class labels; 𝜒𝜒𝐴𝐴 (Greek chi) is the characteristic function or indicator function, which returns 1 if the predicted class of the jth classifier matches i (𝐶𝐶𝑗𝑗 (𝒙𝒙) = 𝑖𝑖). For equal weights, we can simplify this equation and write it as follows:

𝑦𝑦̂ = mode{𝐶𝐶1 (𝒙𝒙), 𝐶𝐶2 (𝒙𝒙), … , 𝐶𝐶𝑚𝑚 (𝒙𝒙)}

To better understand the concept of weighting, we will now take a look at a more concrete example. Let's assume that we have an ensemble of three base classifiers, 𝐶𝐶𝑗𝑗 (𝑗𝑗 ∈ {0, 1}), and we want to predict the class label, 𝐶𝐶𝑗𝑗 (𝒙𝒙) ∈ {0,1}, of a given example, x. Two out of three base classifiers predict the class label 0, and one, 𝐶𝐶3, predicts that the example belongs to class 1. If we weight the predictions of each base classifier equally, the majority vote predicts that the example belongs to class 0:

𝐶𝐶1 (𝒙𝒙) → 0, 𝐶𝐶2 (𝒙𝒙) → 0, 𝐶𝐶3 (𝒙𝒙) → 1 𝑦𝑦̂ = mode{0,0,1} = 0

Now, let's assign a weight of 0.6 to 𝐶𝐶3, and let's weight 𝐶𝐶1 and 𝐶𝐶2 by a coefficient of 0.2: 𝑦𝑦̂

𝑚𝑚

=

arg max ∑ 𝑤𝑤𝑗𝑗 𝜒𝜒𝐴𝐴 (𝐶𝐶𝑗𝑗 (𝒙𝒙) = 𝑖𝑖)

=

arg max[0.2 × 𝑖𝑖0 + 0.2 × 𝑖𝑖0 + 0.6 × 𝑖𝑖1 ] = 1

𝑖𝑖

𝑗𝑗=1

𝑖𝑖

[ 228 ]

Chapter 7

More simply, since 3 × 0.2 = 0.6, we can say that the prediction made by 𝐶𝐶3 has three times more weight than the predictions by 𝐶𝐶1 or 𝐶𝐶2, which we can write as follows:

𝑦𝑦̂ = mode{0,0,1,1,1} = 1

To translate the concept of the weighted majority vote into Python code, we can use NumPy's convenient argmax and bincount functions: >>> import numpy as np >>> np.argmax(np.bincount([0, 0, 1], ... weights=[0.2, 0.2, 0.6])) 1

As you will remember from the discussion on logistic regression in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, certain classifiers in scikit-learn can also return the probability of a predicted class label via the predict_proba method. Using the predicted class probabilities instead of the class labels for majority voting can be useful if the classifiers in our ensemble are well calibrated. The modified version of the majority vote for predicting class labels from probabilities can be written as follows: 𝑚𝑚

𝑦𝑦̂ = arg max ∑ 𝑤𝑤𝑗𝑗 𝑝𝑝𝑖𝑖𝑖𝑖 𝑖𝑖

𝑗𝑗=1

Here, 𝑝𝑝𝑖𝑖𝑖𝑖 is the predicted probability of the jth classifier for class label i.

To continue with our previous example, let's assume that we have a binary classification problem with class labels 𝑖𝑖 ∈ {0, 1} and an ensemble of three classifiers, 𝐶𝐶𝑗𝑗 (𝑗𝑗 ∈ {1, 2, 3}). Let's assume that the classifiers 𝐶𝐶𝑗𝑗 return the following class membership probabilities for a particular example, x: 𝐶𝐶1 (𝒙𝒙) → [0.9, 0.1], 𝐶𝐶2 (𝒙𝒙) → [0.8, 0.2], 𝐶𝐶3 (𝒙𝒙) → [0.4, 0.6]

Using the same weights as previously (0.2, 0.2, and 0.6), we can then calculate the individual class probabilities as follows:

𝑝𝑝(𝑖𝑖0|𝒙𝒙) = 𝑝𝑝(𝑖𝑖1|𝒙𝒙) = 𝑦𝑦̂

=

0.2 × 0.9 + 0.2 × 0.8 + 0.6 × 0.4 = 0.58 0.2 × 0.1 + 0.2 × 0.2 + 0.6 × 0.6 = 0.42 arg max[𝑝𝑝(𝑖𝑖0 | 𝒙𝒙), 𝑝𝑝(𝑖𝑖1 | 𝒙𝒙)] = 0 𝑖𝑖

[ 229 ]

Combining Different Models for Ensemble Learning

To implement the weighted majority vote based on class probabilities, we can again make use of NumPy, using np.average and np.argmax: >>> ex = np.array([[0.9, 0.1], ... [0.8, 0.2], ... [0.4, 0.6]]) >>> p = np.average(ex, axis=0, weights=[0.2, 0.2, 0.6]) >>> p array([0.58, 0.42]) >>> np.argmax(p) 0

Putting everything together, let's now implement MajorityVoteClassifier in Python: from sklearn.base import BaseEstimator from sklearn.base import ClassifierMixin from sklearn.preprocessing import LabelEncoder from sklearn.base import clone from sklearn.pipeline import _name_estimators import numpy as np import operator class MajorityVoteClassifier(BaseEstimator, ClassifierMixin): """ A majority vote ensemble classifier Parameters ---------classifiers : array-like, shape = [n_classifiers] Different classifiers for the ensemble vote : str, {'classlabel', 'probability'} Default: 'classlabel' If 'classlabel' the prediction is based on the argmax of class labels. Else if 'probability', the argmax of the sum of probabilities is used to predict the class label (recommended for calibrated classifiers). weights : array-like, shape = [n_classifiers] Optional, default: None If a list of 'int' or 'float' values are provided, the classifiers are weighted by importance; Uses uniform weights if 'weights=None'. """ def __init__(self, classifiers, vote='classlabel', weights=None):

[ 230 ]

Chapter 7 self.classifiers = classifiers self.named_classifiers = {key: value for key, value in _name_estimators(classifiers)} self.vote = vote self.weights = weights def fit(self, X, y): """ Fit classifiers. Parameters ---------X : {array-like, sparse matrix}, shape = [n_examples, n_features] Matrix of training examples. y : array-like, shape = [n_examples] Vector of target class labels. Returns ------self : object """ if self.vote not in ('probability', 'classlabel'): raise ValueError("vote must be 'probability'" "or 'classlabel'; got (vote=%r)" % self.vote) if self.weights and len(self.weights) != len(self.classifiers): raise ValueError("Number of classifiers and weights" "must be equal; got %d weights," "%d classifiers" % (len(self.weights), len(self.classifiers))) # Use LabelEncoder to ensure class labels start # with 0, which is important for np.argmax # call in self.predict self.lablenc_ = LabelEncoder() self.lablenc_.fit(y) self.classes_ = self.lablenc_.classes_ self.classifiers_ = [] for clf in self.classifiers: fitted_clf = clone(clf).fit(X, self.lablenc_.transform(y)) self.classifiers_.append(fitted_clf) return self

[ 231 ]

Combining Different Models for Ensemble Learning

I've added a lot of comments to the code to explain the individual parts. However, before we implement the remaining methods, let's take a quick break and discuss some of the code that may look confusing at first. We used the BaseEstimator and ClassifierMixin parent classes to get some base functionality for free, including the get_params and set_params methods to set and return the classifier's parameters, as well as the score method to calculate the prediction accuracy. Next, we will add the predict method to predict the class label via a majority vote based on the class labels if we initialize a new MajorityVoteClassifier object with vote='classlabel'. Alternatively, we will be able to initialize the ensemble classifier with vote='probability' to predict the class label based on the class membership probabilities. Furthermore, we will also add a predict_proba method to return the averaged probabilities, which is useful when computing the receiver operating characteristic area under the curve (ROC AUC): def predict(self, X): """ Predict class labels for X. Parameters ---------X : {array-like, sparse matrix}, Shape = [n_examples, n_features] Matrix of training examples. Returns ---------maj_vote : array-like, shape = [n_examples] Predicted class labels. """ if self.vote == 'probability': maj_vote = np.argmax(self.predict_proba(X), axis=1) else: # 'classlabel' vote # Collect results from clf.predict calls predictions = np.asarray([clf.predict(X) for clf in self.classifiers_]).T maj_vote = np.apply_along_axis(lambda x: np.argmax( np.bincount(x, weights=self.weights)), axis=1, arr=predictions)

[ 232 ]

Chapter 7 maj_vote = self.lablenc_.inverse_transform(maj_vote) return maj_vote def predict_proba(self, X): """ Predict class probabilities for X. Parameters ---------X : {array-like, sparse matrix}, shape = [n_examples, n_features] Training vectors, where n_examples is the number of examples and n_features is the number of features. Returns ---------avg_proba : array-like, shape = [n_examples, n_classes] Weighted average probability for each class per example. """ probas = np.asarray([clf.predict_proba(X) for clf in self.classifiers_]) avg_proba = np.average(probas, axis=0, weights=self.weights) return avg_proba def get_params(self, deep=True): """ Get classifier parameter names for GridSearch""" if not deep: return super(MajorityVoteClassifier, self).get_params(deep=False) else: out = self.named_classifiers.copy() for name, step in self.named_classifiers.items(): for key, value in step.get_params( deep=True).items(): out['%s__%s' % (name, key)] = value return out

[ 233 ]

Combining Different Models for Ensemble Learning

Also, note that we defined our own modified version of the get_params method to use the _name_estimators function to access the parameters of individual classifiers in the ensemble; this may look a little bit complicated at first, but it will make perfect sense when we use grid search for hyperparameter tuning in later sections. VotingClassifier in scikit-learn Although the MajorityVoteClassifier implementation is very useful for demonstration purposes, we implemented a more sophisticated version of this majority vote classifier in scikit-learn based on the implementation in the first edition of this book. The ensemble classifier is available as sklearn. ensemble.VotingClassifier in scikit-learn version 0.17 and newer.

Using the majority voting principle to make predictions

Now it is time to put the MajorityVoteClassifier that we implemented in the previous section into action. But first, let's prepare a dataset that we can test it on. Since we are already familiar with techniques to load datasets from CSV files, we will take a shortcut and load the Iris dataset from scikit-learn's datasets module. Furthermore, we will only select two features, sepal width and petal length, to make the classification task more challenging for illustration purposes. Although our MajorityVoteClassifier generalizes to multiclass problems, we will only classify flower examples from the Iris-versicolor and Iris-virginica classes, with which we will compute the ROC AUC later. The code is as follows: >>> >>> >>> >>> >>> >>> >>> >>>

from sklearn import datasets from sklearn.model_selection import train_test_split from sklearn.preprocessing import StandardScaler from sklearn.preprocessing import LabelEncoder iris = datasets.load_iris() X, y = iris.data[50:, [1, 2]], iris.target[50:] le = LabelEncoder() y = le.fit_transform(y)

[ 234 ]

Chapter 7

Class membership probabilities from decision trees Note that scikit-learn uses the predict_proba method (if applicable) to compute the ROC AUC score. In Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, we saw how the class probabilities are computed in logistic regression models. In decision trees, the probabilities are calculated from a frequency vector that is created for each node at training time. The vector collects the frequency values of each class label computed from the class label distribution at that node. Then, the frequencies are normalized so that they sum up to 1. Similarly, the class labels of the k-nearest neighbors are aggregated to return the normalized class label frequencies in the k-nearest neighbors algorithm. Although the normalized probabilities returned by both the decision tree and k-nearest neighbors classifier may look similar to the probabilities obtained from a logistic regression model, we have to be aware that these are actually not derived from probability mass functions.

Next, we will split the Iris examples into 50 percent training and 50 percent test data: >>> X_train, X_test, y_train, y_test =\ ... train_test_split(X, y, ... test_size=0.5, ... random_state=1, ... stratify=y)

Using the training dataset, we now will train three different classifiers: • Logistic regression classifier • Decision tree classifier • k-nearest neighbors classifier We will then evaluate the model performance of each classifier via 10-fold crossvalidation on the training dataset before we combine them into an ensemble classifier: >>> >>> >>> >>> >>> >>>

from sklearn.model_selection import cross_val_score from sklearn.linear_model import LogisticRegression from sklearn.tree import DecisionTreeClassifier from sklearn.neighbors import KNeighborsClassifier from sklearn.pipeline import Pipeline import numpy as np

[ 235 ]

Combining Different Models for Ensemble Learning >>> ... ... ... >>> ... ... >>> ... ... >>> ... >>> ... >>> >>> >>> ... ... ... ... ... ... ...

clf1 = LogisticRegression(penalty='l2', C=0.001, solver='lbfgs', random_state=1) clf2 = DecisionTreeClassifier(max_depth=1, criterion='entropy', random_state=0) clf3 = KNeighborsClassifier(n_neighbors=1, p=2, metric='minkowski') pipe1 = Pipeline([['sc', StandardScaler()], ['clf', clf1]]) pipe3 = Pipeline([['sc', StandardScaler()], ['clf', clf3]]) clf_labels = ['Logistic regression', 'Decision tree', 'KNN'] print('10-fold cross validation:\n') for clf, label in zip([pipe1, clf2, pipe3], clf_labels): scores = cross_val_score(estimator=clf, X=X_train, y=y_train, cv=10, scoring='roc_auc') print("ROC AUC: %0.2f (+/- %0.2f) [%s]" % (scores.mean(), scores.std(), label))

The output that we receive, as shown in the following snippet, shows that the predictive performances of the individual classifiers are almost equal: 10-fold cross validation: ROC AUC: 0.92 (+/- 0.15) [Logistic regression] ROC AUC: 0.87 (+/- 0.18) [Decision tree] ROC AUC: 0.85 (+/- 0.13) [KNN]

You may be wondering why we trained the logistic regression and k-nearest neighbors classifier as part of a pipeline. The reason behind it is that, as discussed in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, both the logistic regression and k-nearest neighbors algorithms (using the Euclidean distance metric) are not scale-invariant, in contrast to decision trees. Although the Iris features are all measured on the same scale (cm), it is a good habit to work with standardized features.

[ 236 ]

Chapter 7

Now, let's move on to the more exciting part and combine the individual classifiers for majority rule voting in our MajorityVoteClassifier: >>> ... >>> >>> >>> ... ... ... ... ... ... ... ROC ROC ROC ROC

mv_clf = MajorityVoteClassifier( classifiers=[pipe1, clf2, pipe3]) clf_labels += ['Majority voting'] all_clf = [pipe1, clf2, pipe3, mv_clf] for clf, label in zip(all_clf, clf_labels): scores = cross_val_score(estimator=clf, X=X_train, y=y_train, cv=10, scoring='roc_auc') print("ROC AUC: %0.2f (+/- %0.2f) [%s]" % (scores.mean(), scores.std(), label)) AUC: 0.92 (+/- 0.15) [Logistic regression] AUC: 0.87 (+/- 0.18) [Decision tree] AUC: 0.85 (+/- 0.13) [KNN] AUC: 0.98 (+/- 0.05) [Majority voting]

As you can see, the performance of MajorityVotingClassifier has improved over the individual classifiers in the 10-fold cross-validation evaluation.

Evaluating and tuning the ensemble classifier

In this section, we are going to compute the ROC curves from the test dataset to check that MajorityVoteClassifier generalizes well with unseen data. We must remember that the test dataset is not to be used for model selection; its purpose is merely to report an unbiased estimate of the generalization performance of a classifier system: >>> >>> >>> >>> >>> ... ... ... ... ... ... ... ... ...

from sklearn.metrics import roc_curve from sklearn.metrics import auc colors = ['black', 'orange', 'blue', 'green'] linestyles = [':', '--', '-.', '-'] for clf, label, clr, ls \ in zip(all_clf, clf_labels, colors, linestyles): # assuming the label of the positive class is 1 y_pred = clf.fit(X_train, y_train).predict_proba(X_test)[:, 1] fpr, tpr, thresholds = roc_curve(y_true=y_test, y_score=y_pred) roc_auc = auc(x=fpr, y=tpr) plt.plot(fpr, tpr, color=clr, [ 237 ]

Combining Different Models for Ensemble Learning ... ... >>> >>> ... ... ... >>> >>> >>> >>> >>> >>>

linestyle=ls, label='%s (auc = %0.2f)' % (label, roc_auc)) plt.legend(loc='lower right') plt.plot([0, 1], [0, 1], linestyle='--', color='gray', linewidth=2) plt.xlim([-0.1, 1.1]) plt.ylim([-0.1, 1.1]) plt.grid(alpha=0.5) plt.xlabel('False positive rate (FPR)') plt.ylabel('True positive rate (TPR)') plt.show()

As you can see in the resulting ROC, the ensemble classifier also performs well on the test dataset (ROC AUC = 0.95). However, you can see that the logistic regression classifier performs similarly well on the same dataset, which is probably due to the high variance (in this case, sensitivity of how we split the dataset) given the small size of the dataset:

Since we only selected two features for the classification examples, it would be interesting to see what the decision region of the ensemble classifier actually looks like.

[ 238 ]

Chapter 7

Although it is not necessary to standardize the training features prior to model fitting, because our logistic regression and k-nearest neighbors pipelines will automatically take care of it, we will standardize the training dataset so that the decision regions of the decision tree will be on the same scale for visual purposes. The code is as follows: >>> >>> >>> >>> >>> >>> >>> >>> >>> ... >>> ... ... ... >>> ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... >>> ... ... >>> ... ... ... >>>

sc = StandardScaler() X_train_std = sc.fit_transform(X_train) from itertools import product x_min = X_train_std[:, 0].min() - 1 x_max = X_train_std[:, 0].max() + 1 y_min = X_train_std[:, 1].min() - 1 y_max = X_train_std[:, 1].max() + 1 xx, yy = np.meshgrid(np.arange(x_min, x_max, 0.1), np.arange(y_min, y_max, 0.1)) f, axarr = plt.subplots(nrows=2, ncols=2, sharex='col', sharey='row', figsize=(7, 5)) for idx, clf, tt in zip(product([0, 1], [0, 1]), all_clf, clf_labels): clf.fit(X_train_std, y_train) Z = clf.predict(np.c_[xx.ravel(), yy.ravel()]) Z = Z.reshape(xx.shape) axarr[idx[0], idx[1]].contourf(xx, yy, Z, alpha=0.3) axarr[idx[0], idx[1]].scatter(X_train_std[y_train==0, X_train_std[y_train==0, c='blue', marker='^', s=50) axarr[idx[0], idx[1]].scatter(X_train_std[y_train==1, X_train_std[y_train==1, c='green', marker='o', s=50) axarr[idx[0], idx[1]].set_title(tt) plt.text(-3.5, -5., s='Sepal width [standardized]', ha='center', va='center', fontsize=12) plt.text(-12.5, 4.5, s='Petal length [standardized]', ha='center', va='center', fontsize=12, rotation=90) plt.show()

[ 239 ]

0], 1],

0], 1],

Combining Different Models for Ensemble Learning

Interestingly, but also as expected, the decision regions of the ensemble classifier seem to be a hybrid of the decision regions from the individual classifiers. At first glance, the majority vote decision boundary looks a lot like the decision of the decision tree stump, which is orthogonal to the y axis for sepal width ≥ 1. However, you can also notice the non-linearity from the k-nearest neighbor classifier mixed in:

Before we tune the individual classifier's parameters for ensemble classification, let's call the get_params method to get a basic idea of how we can access the individual parameters inside a GridSearchCV object: >>> mv_clf.get_params() {'decisiontreeclassifier': DecisionTreeClassifier(class_weight=None, criterion='entropy', max_depth=1, max_features=None, max_leaf_nodes=None, min_samples_leaf=1, min_samples_split=2, min_weight_fraction_leaf=0.0, random_state=0, splitter='best'), 'decisiontreeclassifier__class_weight': None, 'decisiontreeclassifier__criterion': 'entropy', [...] 'decisiontreeclassifier__random_state': 0, 'decisiontreeclassifier__splitter': 'best', 'pipeline-1':

[ 240 ]

Chapter 7 Pipeline(steps=[('sc', StandardScaler(copy=True, with_mean=True, with_std=True)), ('clf', LogisticRegression(C=0.001, class_weight=None, dual=False, fit_intercept=True, intercept_scaling=1, max_iter=100, multi_class='ovr', penalty='l2', random_state=0, solver='liblinear', tol=0.0001, verbose=0))]), 'pipeline-1__clf': LogisticRegression(C=0.001, class_weight=None, dual=False, fit_intercept=True, intercept_scaling=1, max_iter=100, multi_class='ovr', penalty='l2', random_state=0, solver='liblinear', tol=0.0001, verbose=0), 'pipeline-1__clf__C': 0.001, 'pipeline-1__clf__class_weight': None, 'pipeline-1__clf__dual': False, [...] 'pipeline-1__sc__with_std': True, 'pipeline-2': Pipeline(steps=[('sc', StandardScaler(copy=True, with_mean=True, with_std=True)), ('clf', KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='minkowski', metric_params=None, n_neighbors=1, p=2, weights='uniform'))]), 'pipeline-2__clf': KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='minkowski', metric_params=None, n_neighbors=1, p=2, weights='uniform'), 'pipeline-2__clf__algorithm': 'auto', [...] 'pipeline-2__sc__with_std': True}

[ 241 ]

Combining Different Models for Ensemble Learning

Based on the values returned by the get_params method, we now know how to access the individual classifier's attributes. Let's now tune the inverse regularization parameter, C, of the logistic regression classifier and the decision tree depth via a grid search for demonstration purposes: >>> >>> ... >>> ... ... ... ... >>>

from sklearn.model_selection import GridSearchCV params = {'decisiontreeclassifier__max_depth': [1, 2], 'pipeline-1__clf__C': [0.001, 0.1, 100.0]} grid = GridSearchCV(estimator=mv_clf, param_grid=params, cv=10, iid=False, scoring='roc_auc') grid.fit(X_train, y_train)

After the grid search has completed, we can print the different hyperparameter value combinations and the average ROC AUC scores computed via 10-fold crossvalidation as follows: >>> for r, _ in enumerate(grid.cv_results_['mean_test_score']): ... print("%0.3f +/- %0.2f %r" ... % (grid.cv_results_['mean_test_score'][r], ... grid.cv_results_['std_test_score'][r] / 2.0, ... grid.cv_results_['params'][r])) 0.944 +/- 0.07 {'decisiontreeclassifier__max_depth': 1, 'pipeline-1__clf__C': 0.001} 0.956 +/- 0.07 {'decisiontreeclassifier__max_depth': 1, 'pipeline-1__clf__C': 0.1} 0.978 +/- 0.03 {'decisiontreeclassifier__max_depth': 1, 'pipeline-1__clf__C': 100.0} 0.956 +/- 0.07 {'decisiontreeclassifier__max_depth': 2, 'pipeline-1__clf__C': 0.001} 0.956 +/- 0.07 {'decisiontreeclassifier__max_depth': 2, 'pipeline-1__clf__C': 0.1} 0.978 +/- 0.03 {'decisiontreeclassifier__max_depth': 2, 'pipeline-1__clf__C': 100.0} >>> print('Best parameters: %s' % grid.best_params_) Best parameters: {'decisiontreeclassifier__max_depth': 1, 'pipeline-1__clf__C': 0.001} >>> print('Accuracy: %.2f' % grid.best_score_) Accuracy: 0.98

[ 242 ]

Chapter 7

As you can see, we get the best cross-validation results when we choose a lower regularization strength (C=0.001), whereas the tree depth does not seem to affect the performance at all, suggesting that a decision stump is sufficient to separate the data. To remind ourselves that it is a bad practice to use the test dataset more than once for model evaluation, we are not going to estimate the generalization performance of the tuned hyperparameters in this section. We will move on swiftly to an alternative approach for ensemble learning: bagging. Building ensembles using stacking The majority vote approach we implemented in this section is not to be confused with stacking. The stacking algorithm can be understood as a two-level ensemble, where the first level consists of individual classifiers that feed their predictions to the second level, where another classifier (typically logistic regression) is fit to the level-one classifier predictions to make the final predictions. The stacking algorithm has been described in more detail by David H. Wolpert in Stacked generalization, Neural Networks, 5(2):241–259, 1992. Unfortunately, an implementation of this algorithm has not been implemented in scikit-learn at the time of writing; however, this feature is underway. In the meantime, you can find scikit-learn-compatible implementations of stacking at http://rasbt.github.io/mlxtend/user_guide/ classifier/StackingClassifier/ and http:// rasbt.github.io/mlxtend/user_guide/classifier/ StackingCVClassifier/.

Bagging – building an ensemble of classifiers from bootstrap samples

Bagging is an ensemble learning technique that is closely related to the MajorityVoteClassifier that we implemented in the previous section. However, instead of using the same training dataset to fit the individual classifiers in the ensemble, we draw bootstrap samples (random samples with replacement) from the initial training dataset, which is why bagging is also known as bootstrap aggregating.

[ 243 ]

Combining Different Models for Ensemble Learning

The concept of bagging is summarized in the following diagram:

In the following subsections, we will work through a simple example of bagging by hand and use scikit-learn for classifying wine examples.

Bagging in a nutshell

To provide a more concrete example of how the bootstrap aggregating of a bagging classifier works, let's consider the example shown in the following figure. Here, we have seven different training instances (denoted as indices 1-7) that are sampled randomly with replacement in each round of bagging. Each bootstrap sample is then used to fit a classifier, 𝐶𝐶𝑗𝑗, which is most typically an unpruned decision tree:

[ 244 ]

Chapter 7

As you can see from the previous illustration, each classifier receives a random subset of examples from the training dataset. We denote these random samples obtained via bagging as Bagging round 1, Bagging round 2, and so on. Each subset contains a certain portion of duplicates and some of the original examples don't appear in a resampled dataset at all due to sampling with replacement. Once the individual classifiers are fit to the bootstrap samples, the predictions are combined using majority voting. Note that bagging is also related to the random forest classifier that we introduced in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn. In fact, random forests are a special case of bagging where we also use random feature subsets when fitting the individual decision trees. Model ensembles using bagging Bagging was first proposed by Leo Breiman in a technical report in 1994; he also showed that bagging can improve the accuracy of unstable models and decrease the degree of overfitting. I highly recommend that you read about his research in Bagging predictors, L. Breiman, Machine Learning, 24(2):123–140, 1996, which is freely available online, to learn more details about bagging.

Applying bagging to classify examples in the Wine dataset

To see bagging in action, let's create a more complex classification problem using the Wine dataset that was introduced in Chapter 4, Building Good Training Datasets – Data Preprocessing. Here, we will only consider the Wine classes 2 and 3, and we will select two features – Alcohol and OD280/OD315 of diluted wines: >>> import pandas as pd >>> df_wine = pd.read_csv('https://archive.ics.uci.edu/ml/' ... 'machine-learning-databases/wine/wine.data', ... header=None) >>> df_wine.columns = ['Class label', 'Alcohol', ... 'Malic acid', 'Ash', ... 'Alcalinity of ash', ... 'Magnesium', 'Total phenols', ... 'Flavanoids', 'Nonflavanoid phenols', ... 'Proanthocyanins', ... 'Color intensity', 'Hue', ... 'OD280/OD315 of diluted wines', [ 245 ]

Combining Different Models for Ensemble Learning ... >>> >>> >>> >>> ...

'Proline'] # drop 1 class df_wine = df_wine[df_wine['Class label'] != 1] y = df_wine['Class label'].values X = df_wine[['Alcohol', 'OD280/OD315 of diluted wines']].values

Next, we will encode the class labels into binary format and split the dataset into 80 percent training and 20 percent test datasets, respectively: >>> >>> >>> >>> >>> ... ... ... ...

from sklearn.preprocessing import LabelEncoder from sklearn.model_selection import train_test_split le = LabelEncoder() y = le.fit_transform(y) X_train, X_test, y_train, y_test =\ train_test_split(X, y, test_size=0.2, random_state=1, stratify=y)

Obtaining the Wine dataset You can find a copy of the Wine dataset (and all other datasets used in this book) in the code bundle of this book, which you can use if you are working offline or the UCI server at https:// archive.ics.uci.edu/ml/machine-learning-databases/ wine/wine.data is temporarily unavailable. For instance, to

load the Wine dataset from a local directory, take the following lines: df = pd.read_csv( 'https://archive.ics.uci.edu/ml/' 'machine-learning-databases' '/wine/wine.data', header=None)

and replace them with these: df = pd.read_csv( 'your/local/path/to/wine.data', header=None)

[ 246 ]

Chapter 7

A BaggingClassifier algorithm is already implemented in scikit-learn, which we can import from the ensemble submodule. Here, we will use an unpruned decision tree as the base classifier and create an ensemble of 500 decision trees fit on different bootstrap samples of the training dataset: >>> from sklearn.ensemble import BaggingClassifier >>> tree = DecisionTreeClassifier(criterion='entropy', ... random_state=1, ... max_depth=None) >>> bag = BaggingClassifier(base_estimator=tree, ... n_estimators=500, ... max_samples=1.0, ... max_features=1.0, ... bootstrap=True, ... bootstrap_features=False, ... n_jobs=1, ... random_state=1)

Next, we will calculate the accuracy score of the prediction on the training and test datasets to compare the performance of the bagging classifier to the performance of a single unpruned decision tree: >>> from sklearn.metrics import accuracy_score >>> tree = tree.fit(X_train, y_train) >>> y_train_pred = tree.predict(X_train) >>> y_test_pred = tree.predict(X_test) >>> tree_train = accuracy_score(y_train, y_train_pred) >>> tree_test = accuracy_score(y_test, y_test_pred) >>> print('Decision tree train/test accuracies %.3f/%.3f' ... % (tree_train, tree_test)) Decision tree train/test accuracies 1.000/0.833

Based on the accuracy values that we printed here, the unpruned decision tree predicts all the class labels of the training examples correctly; however, the substantially lower test accuracy indicates high variance (overfitting) of the model: >>> bag = bag.fit(X_train, y_train) >>> y_train_pred = bag.predict(X_train) >>> y_test_pred = bag.predict(X_test) >>> bag_train = accuracy_score(y_train, y_train_pred) >>> bag_test = accuracy_score(y_test, y_test_pred) >>> print('Bagging train/test accuracies %.3f/%.3f' ... % (bag_train, bag_test)) Bagging train/test accuracies 1.000/0.917

[ 247 ]

Combining Different Models for Ensemble Learning

Although the training accuracies of the decision tree and bagging classifier are similar on the training dataset (both 100 percent), we can see that the bagging classifier has a slightly better generalization performance, as estimated on the test dataset. Next, let's compare the decision regions between the decision tree and the bagging classifier: >>> >>> >>> >>> >>> ... >>> ... ... ... >>> ... ... ... ... ... ... ... ... ... ... ... ... ... ... >>> >>> >>> ... ... ... ... ... >>>

x_min = X_train[:, 0].min() - 1 x_max = X_train[:, 0].max() + 1 y_min = X_train[:, 1].min() - 1 y_max = X_train[:, 1].max() + 1 xx, yy = np.meshgrid(np.arange(x_min, x_max, 0.1), np.arange(y_min, y_max, 0.1)) f, axarr = plt.subplots(nrows=1, ncols=2, sharex='col', sharey='row', figsize=(8, 3)) for idx, clf, tt in zip([0, 1], [tree, bag], ['Decision tree', 'Bagging']): clf.fit(X_train, y_train) Z = clf.predict(np.c_[xx.ravel(), yy.ravel()]) Z = Z.reshape(xx.shape) axarr[idx].contourf(xx, yy, Z, alpha=0.3) axarr[idx].scatter(X_train[y_train==0, 0], X_train[y_train==0, 1], c='blue', marker='^') axarr[idx].scatter(X_train[y_train==1, 0], X_train[y_train==1, 1], c='green', marker='o') axarr[idx].set_title(tt) axarr[0].set_ylabel('Alcohol', fontsize=12) plt.tight_layout() plt.text(0, -0.2, s='OD280/OD315 of diluted wines', ha='center', va='center', fontsize=12, transform=axarr[1].transAxes) plt.show()

As we can see in the resulting plot, the piece-wise linear decision boundary of the three-node deep decision tree looks smoother in the bagging ensemble:

[ 248 ]

Chapter 7

We only looked at a very simple bagging example in this section. In practice, more complex classification tasks and a dataset's high dimensionality can easily lead to overfitting in single decision trees, and this is where the bagging algorithm can really play to its strengths. Finally, we must note that the bagging algorithm can be an effective approach to reducing the variance of a model. However, bagging is ineffective in reducing model bias, that is, models that are too simple to capture the trend in the data well. This is why we want to perform bagging on an ensemble of classifiers with low bias, for example, unpruned decision trees.

Leveraging weak learners via adaptive boosting

In this last section about ensemble methods, we will discuss boosting, with a special focus on its most common implementation: Adaptive Boosting (AdaBoost). AdaBoost recognition The original idea behind AdaBoost was formulated by Robert E. Schapire in 1990. The Strength of Weak Learnability, R. E. Schapire, Machine Learning, 5(2): 197-227, 1990. After Robert Schapire and Yoav Freund presented the AdaBoost algorithm in the Proceedings of the Thirteenth International Conference (ICML 1996), AdaBoost became one of the most widely used ensemble methods in the years that followed (Experiments with a New Boosting Algorithm by Y. Freund, R. E. Schapire, and others, ICML, volume 96, 148-156, 1996). In 2003, Freund and Schapire received the Goedel Prize for their groundbreaking work, which is a prestigious prize for the most outstanding publications in the field of computer science.

[ 249 ]

Combining Different Models for Ensemble Learning

In boosting, the ensemble consists of very simple base classifiers, also often referred to as weak learners, which often only have a slight performance advantage over random guessing—a typical example of a weak learner is a decision tree stump. The key concept behind boosting is to focus on training examples that are hard to classify, that is, to let the weak learners subsequently learn from misclassified training examples to improve the performance of the ensemble. The following subsections will introduce the algorithmic procedure behind the general concept of boosting and AdaBoost. Lastly, we will use scikit-learn for a practical classification example.

How boosting works

In contrast to bagging, the initial formulation of the boosting algorithm uses random subsets of training examples drawn from the training dataset without replacement; the original boosting procedure can be summarized in the following four key steps: 1. Draw a random subset (sample) of training examples, 𝑑𝑑1, without replacement from the training dataset, D, to train a weak learner, 𝐶𝐶1.

2. Draw a second random training subset, 𝑑𝑑2, without replacement from the training dataset and add 50 percent of the examples that were previously misclassified to train a weak learner, 𝐶𝐶2.

3. Find the training examples, 𝑑𝑑3, in the training dataset, D, which 𝐶𝐶1 and 𝐶𝐶2 disagree upon, to train a third weak learner, 𝐶𝐶3. 4. Combine the weak learners 𝐶𝐶1, 𝐶𝐶2, and 𝐶𝐶3 via majority voting.

As discussed by Leo Breiman (Bias, variance, and arcing classifiers, L. Breiman, 1996), boosting can lead to a decrease in bias as well as variance compared to bagging models. In practice, however, boosting algorithms such as AdaBoost are also known for their high variance, that is, the tendency to overfit the training data (An improvement of AdaBoost to avoid overfitting, G. Raetsch, T. Onoda, and K. R. Mueller. Proceedings of the International Conference on Neural Information Processing, CiteSeer, 1998). In contrast to the original boosting procedure described here, AdaBoost uses the complete training dataset to train the weak learners, where the training examples are reweighted in each iteration to build a strong classifier that learns from the mistakes of the previous weak learners in the ensemble.

[ 250 ]

Chapter 7

Before we dive deeper into the specific details of the AdaBoost algorithm, let's take a look at the following figure to get a better grasp of the basic concept behind AdaBoost:

To walk through the AdaBoost illustration step by step, we will start with subfigure 1, which represents a training dataset for binary classification where all training examples are assigned equal weights. Based on this training dataset, we train a decision stump (shown as a dashed line) that tries to classify the examples of the two classes (triangles and circles), as well as possibly minimizing the cost function (or the impurity score in the special case of decision tree ensembles). For the next round (subfigure 2), we assign a larger weight to the two previously misclassified examples (circles). Furthermore, we lower the weight of the correctly classified examples. The next decision stump will now be more focused on the training examples that have the largest weights—the training examples that are supposedly hard to classify. The weak learner shown in subfigure 2 misclassifies three different examples from the circle class, which are then assigned a larger weight, as shown in subfigure 3. Assuming that our AdaBoost ensemble only consists of three rounds of boosting, we then combine the three weak learners trained on different reweighted training subsets by a weighted majority vote, as shown in subfigure 4.

[ 251 ]

Combining Different Models for Ensemble Learning

Now that we have a better understanding of the basic concept of AdaBoost, let's take a more detailed look at the algorithm using pseudo code. For clarity, we will denote element-wise multiplication by the cross symbol (×) and the dot-product between two vectors by a dot symbol ( ∙): 1. Set the weight vector, w, to uniform weights, where ∑ 𝑤𝑤𝑖𝑖 = 1. 𝒊𝒊

2. For j in m boosting rounds, do the following:

a. Train a weighted weak learner: 𝐶𝐶𝑗𝑗 = train(𝑿𝑿, 𝒚𝒚, 𝒘𝒘). b. Predict class labels: 𝒚𝒚 ̂ = predict(𝐶𝐶𝑗𝑗 , 𝑿𝑿).

̂ ≠ 𝒚𝒚). c. Compute weighted error rate: 𝜀𝜀 = 𝒘𝒘 ∙ (𝒚𝒚 1 − 𝜺𝜺 d. Compute coefficient: 𝛼𝛼𝑗𝑗 = 0.5 log 𝜺𝜺 e. Update weights: 𝒘𝒘 ∶= 𝒘𝒘 × exp(−𝛼𝛼𝑗𝑗 × 𝒚𝒚 ̂ × 𝒚𝒚).

f. Normalize weights to sum to 1: 𝒘𝒘 ∶= 𝒘𝒘 / ∑ 𝑤𝑤𝑖𝑖 .

3. Compute the final prediction:

𝑚𝑚

𝑖𝑖

̂ = (∑ (𝜶𝜶𝑗𝑗 × predict(𝐶𝐶𝑗𝑗 , 𝑿𝑿)) > 0). 𝒚𝒚 𝑗𝑗=1

Note that the expression (𝒚𝒚̂ ≠ 𝒚𝒚) in step 2c refers to a binary vector consisting of 1s and 0s, where a 1 is assigned if the prediction is incorrect and 0 is assigned otherwise. Although the AdaBoost algorithm seems to be pretty straightforward, let's walk through a more concrete example using a training dataset consisting of 10 training examples, as illustrated in the following table:

[ 252 ]

Chapter 7

The first column of the table depicts the indices of training examples 1 to 10. In the second column, you can see the feature values of the individual samples, assuming this is a one-dimensional dataset. The third column shows the true class label, 𝑦𝑦𝑖𝑖, for each training sample, 𝑥𝑥𝑖𝑖, where 𝑦𝑦𝑖𝑖 ∈ {1, −1}. The initial weights are shown in the fourth column; we initialize the weights uniformly (assigning the same constant value) and normalize them to sum to 1. In the case of the 10-sample training dataset, we therefore assign 0.1 to each weight, 𝑤𝑤𝑖𝑖 , in the weight vector, w. The predicted class labels, 𝒚𝒚 ̂ , are shown in the fifth column, assuming that our splitting criterion is 𝑥𝑥 ≤ 3.0. The last column of the table then shows the updated weights based on the update rules that we defined in the pseudo code.

Since the computation of the weight updates may look a little bit complicated at first, we will now follow the calculation step by step. We will start by computing the weighted error rate, 𝜀𝜀, as described in step 2c: 𝜀𝜀

= 0.1 × 0 + 0.1 × 0 + 0.1 × 0 + 0.1 × 0 + 0.1 × 0 + 0.1 × 0 +0.1 × 1 + 0.1 × 1 + 0.1 × 1 + 0.1 × 0 =

3 = 0.3 10

Next, we will compute the coefficient, 𝛼𝛼𝑗𝑗—shown in step 2d—which will later be used in step 2e to update the weights, as well as for the weights in the majority vote prediction (step 3):

1 − 𝜀𝜀 𝛼𝛼𝑗𝑗 = 0.5 log ( ) ≈ 0.424 𝜀𝜀

After we have computed the coefficient, 𝛼𝛼𝑗𝑗, we can now update the weight vector using the following equation:

̂ × 𝒚𝒚) 𝒘𝒘 ∶= 𝒘𝒘 × exp(−𝛼𝛼𝑗𝑗 × 𝒚𝒚

̂ × 𝒚𝒚 is an element-wise multiplication between the vectors of the predicted Here, 𝒚𝒚 ̂𝒊𝒊 × 𝒚𝒚𝒊𝒊 will have and true class labels, respectively. Thus, if a prediction, 𝑦𝑦̂𝑖𝑖, is correct, 𝒚𝒚 a positive sign so that we decrease the ith weight, since 𝛼𝛼𝑗𝑗 is a positive number as well: 0.1 × exp(−0.424 × 1 × 1) ≈ 0.065

Similarly, we will increase the ith weight if 𝑦𝑦̂𝑖𝑖 predicted the label incorrectly, like this: 0.1 × exp(−0.424 × 1 × (−1)) ≈ 0.153

Alternatively, it's like this:

0.1 × exp(−0.424 × (−1) × (1)) ≈ 0.153

[ 253 ]

Combining Different Models for Ensemble Learning

After we have updated each weight in the weight vector, we normalize the weights so that they sum up to 1 (step 2f): 𝒘𝒘 ∶=

𝒘𝒘 ∑𝑖𝑖 𝑤𝑤𝑖𝑖

Here, ∑ 𝑤𝑤𝑖𝑖 = 7 × 0.065 + 3 × 0.153 = 0.914 . 𝑖𝑖

Thus, each weight that corresponds to a correctly classified example will be reduced from the initial value of 0.1 to 0.065/0.914 ≈ 0.071 for the next round of boosting. Similarly, the weights of the incorrectly classified examples will increase from 0.1 to 0.153/0.914 ≈ 0.167.

Applying AdaBoost using scikit-learn

The previous subsection introduced AdaBoost in a nutshell. Skipping to the more practical part, let's now train an AdaBoost ensemble classifier via scikit-learn. We will use the same Wine subset that we used in the previous section to train the bagging meta-classifier. Via the base_estimator attribute, we will train the AdaBoostClassifier on 500 decision tree stumps: >>> from sklearn.ensemble import AdaBoostClassifier >>> tree = DecisionTreeClassifier(criterion='entropy', ... random_state=1, ... max_depth=1) >>> ada = AdaBoostClassifier(base_estimator=tree, ... n_estimators=500, ... learning_rate=0.1, ... random_state=1) >>> tree = tree.fit(X_train, y_train) >>> y_train_pred = tree.predict(X_train) >>> y_test_pred = tree.predict(X_test) >>> tree_train = accuracy_score(y_train, y_train_pred) >>> tree_test = accuracy_score(y_test, y_test_pred) >>> print('Decision tree train/test accuracies %.3f/%.3f' ... % (tree_train, tree_test)) Decision tree train/test accuracies 0.916/0.875

As you can see, the decision tree stump seems to underfit the training data in contrast to the unpruned decision tree that we saw in the previous section: >>> >>> >>> >>>

ada = ada.fit(X_train, y_train) y_train_pred = ada.predict(X_train) y_test_pred = ada.predict(X_test) ada_train = accuracy_score(y_train, y_train_pred)

[ 254 ]

Chapter 7 >>> ada_test = accuracy_score(y_test, y_test_pred) >>> print('AdaBoost train/test accuracies %.3f/%.3f' ... % (ada_train, ada_test)) AdaBoost train/test accuracies 1.000/0.917

Here, you can see that the AdaBoost model predicts all class labels of the training dataset correctly and also shows a slightly improved test dataset performance compared to the decision tree stump. However, you can also see that we introduced additional variance by our attempt to reduce the model bias—a greater gap between training and test performance. Although we used another simple example for demonstration purposes, we can see that the performance of the AdaBoost classifier is slightly improved compared to the decision stump and achieved very similar accuracy scores as the bagging classifier that we trained in the previous section. However, we must note that it is considered bad practice to select a model based on the repeated usage of the test dataset. The estimate of the generalization performance may be over-optimistic, which we discussed in more detail in Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning. Lastly, let's check what the decision regions look like: >>> >>> >>> >>> >>> ... >>> ... ... ... >>> ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

x_min = X_train[:, 0].min() - 1 x_max = X_train[:, 0].max() + 1 y_min = X_train[:, 1].min() - 1 y_max = X_train[:, 1].max() + 1 xx, yy = np.meshgrid(np.arange(x_min, x_max, 0.1), np.arange(y_min, y_max, 0.1)) f, axarr = plt.subplots(1, 2, sharex='col', sharey='row', figsize=(8, 3)) for idx, clf, tt in zip([0, 1], [tree, ada], ['Decision Tree', 'AdaBoost']): clf.fit(X_train, y_train) Z = clf.predict(np.c_[xx.ravel(), yy.ravel()]) Z = Z.reshape(xx.shape) axarr[idx].contourf(xx, yy, Z, alpha=0.3) axarr[idx].scatter(X_train[y_train==0, 0], X_train[y_train==0, 1], c='blue', marker='^') axarr[idx].scatter(X_train[y_train==1, 0], X_train[y_train==1, 1], c='green', marker='o') axarr[idx].set_title(tt) axarr[0].set_ylabel('Alcohol', fontsize=12) [ 255 ]

Combining Different Models for Ensemble Learning >>> plt.tight_layout() >>> plt.text(0, -0.2, ... s='OD280/OD315 of diluted wines', ... ha='center', ... va='center', ... fontsize=12, ... transform=axarr[1].transAxes) >>> plt.show()

By looking at the decision regions, you can see that the decision boundary of the AdaBoost model is substantially more complex than the decision boundary of the decision stump. In addition, note that the AdaBoost model separates the feature space very similarly to the bagging classifier that we trained in the previous section:

As concluding remarks about ensemble techniques, it is worth noting that ensemble learning increases the computational complexity compared to individual classifiers. In practice, we need to think carefully about whether we want to pay the price of increased computational costs for an often relatively modest improvement in predictive performance. An often-cited example of this tradeoff is the famous $1 million Netflix Prize, which was won using ensemble techniques. The details about the algorithm were published in The BigChaos Solution to the Netflix Grand Prize by A. Toescher, M. Jahrer, and R. M. Bell, Netflix Prize documentation, 2009, which is available at http://www.stat.osu. edu/~dmsl/GrandPrize2009_BPC_BigChaos.pdf. The winning team received the $1 million grand prize money; however, Netflix never implemented their model due to its complexity, which made it infeasible for a real-world application: "We evaluated some of the new methods offline but the additional accuracy gains that we measured did not seem to justify the engineering effort needed to bring them into a production environment." http://techblog.netflix.com/2012/04/netflix-recommendations-beyond-5stars.html [ 256 ]

Chapter 7

Gradient boosting Another popular variant of boosting is gradient boosting. AdaBoost and gradient boosting share the main overall concept: boosting weak learners (such as decision tree stumps) to strong learners. The two approaches, adaptive and gradient boosting, differ mainly with regard to how the weights are updated and how the (weak) classifiers are combined. If you are familiar with gradient-based optimization and interested in gradient boosting, I recommend reading Jerome Friedman's work (Greedy function approximation: a gradient boosting machine. Jerome Friedman. Annals of Statistics 2001, pp.1189-1232) and the more recent paper on XGBoost, which is essentially a computationally efficient implementation of the original gradient boost algorithm (XGBoost: A scalable tree boosting system. Tianqi Chen and Carlos Guestrin. Proceeding of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM 2016, pp. 785794). Note that next to the GradientBoostingClassifier implementation in scikit-learn, scikit-learn now also includes a substantially faster version of gradient boosting in version 0.21, HistGradientBoostingClassifier, which is even faster than XGBoost. For more information about the GradientBoostingClassifier and HistGradientBoostingClassifier in scikit-learn, you can read the documentation at https://scikit-learn. org/stable/modules/ensemble.html#gradienttree-boosting. Also, a short, general explanation of

gradient boosting can be found in the lecture notes at

https://sebastianraschka.com/pdf/lecture-notes/ stat479fs19/07-ensembles__notes.pdf

Summary

In this chapter, we looked at some of the most popular and widely used techniques for ensemble learning. Ensemble methods combine different classification models to cancel out their individual weaknesses, which often results in stable and wellperforming models that are very attractive for industrial applications as well as machine learning competitions.

[ 257 ]

Combining Different Models for Ensemble Learning

At the beginning of this chapter, we implemented MajorityVoteClassifier in Python, which allows us to combine different algorithms for classification. We then looked at bagging, a useful technique for reducing the variance of a model by drawing random bootstrap samples from the training dataset and combining the individually trained classifiers via majority vote. Lastly, we learned about AdaBoost, which is an algorithm that is based on weak learners that subsequently learn from mistakes. Throughout the previous chapters, we learned a lot about different learning algorithms, tuning, and evaluation techniques. In the next chapter, we will look at a particular application of machine learning, sentiment analysis, which has become an interesting topic in the internet and social media era.

[ 258 ]

Applying Machine Learning to Sentiment Analysis In the modern internet and social media age, people's opinions, reviews, and recommendations have become a valuable resource for political science and businesses. Thanks to modern technologies, we are now able to collect and analyze such data most efficiently. In this chapter, we will delve into a subfield of natural language processing (NLP) called sentiment analysis and learn how to use machine learning algorithms to classify documents based on their polarity: the attitude of the writer. In particular, we are going to work with a dataset of 50,000 movie reviews from the Internet Movie Database (IMDb) and build a predictor that can distinguish between positive and negative reviews. The topics that we will cover in the following sections include the following: • Cleaning and preparing text data • Building feature vectors from text documents • Training a machine learning model to classify positive and negative movie reviews • Working with large text datasets using out-of-core learning • Inferring topics from document collections for categorization

Preparing the IMDb movie review data for text processing

As mentioned, sentiment analysis, sometimes also called opinion mining, is a popular subdiscipline of the broader field of NLP; it is concerned with analyzing the polarity of documents. A popular task in sentiment analysis is the classification of documents based on the expressed opinions or emotions of the authors with regard to a particular topic. [ 259 ]

Applying Machine Learning to Sentiment Analysis

In this chapter, we will be working with a large dataset of movie reviews from the Internet Movie Database (IMDb) that has been collected by Andrew Maas and others (Learning Word Vectors for Sentiment Analysis, A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, and C. Potts, Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150, Portland, Oregon, USA, Association for Computational Linguistics, June 2011). The movie review dataset consists of 50,000 polar movie reviews that are labeled as either positive or negative; here, positive means that a movie was rated with more than six stars on IMDb, and negative means that a movie was rated with fewer than five stars on IMDb. In the following sections, we will download the dataset, preprocess it into a useable format for machine learning tools, and extract meaningful information from a subset of these movie reviews to build a machine learning model that can predict whether a certain reviewer liked or disliked a movie.

Obtaining the movie review dataset

A compressed archive of the movie review dataset (84.1 MB) can be downloaded from http://ai.stanford.edu/~amaas/data/sentiment/ as a gzip-compressed tarball archive: • If you are working with Linux or macOS, you can open a new terminal window, cd into the download directory and execute tar -zxf aclImdb_v1.tar.gz to decompress the dataset. • If you are working with Windows, you can download a free archiver, such as 7-Zip (http://www.7-zip.org), to extract the files from the download archive. • Alternatively, you can directly unpack the gzip-compressed tarball archive directly in Python as follows: >>> import tarfile >>> with tarfile.open('aclImdb_v1.tar.gz', 'r:gz') as tar: ... tar.extractall()

Preprocessing the movie dataset into a more convenient format

Having successfully extracted the dataset, we will now assemble the individual text documents from the decompressed download archive into a single CSV file. In the following code section, we will be reading the movie reviews into a pandas DataFrame object, which can take up to 10 minutes on a standard desktop computer.

[ 260 ]

Chapter 8

To visualize the progress and estimated time until completion, we will use the Python Progress Indicator (PyPrind, https://pypi.python.org/pypi/PyPrind/) package, which was developed several years ago for such purposes. PyPrind can be installed by executing the pip install pyprind command: >>> import pyprind >>> import pandas as pd >>> import os >>> # change the 'basepath' to the directory of the >>> # unzipped movie dataset >>> basepath = 'aclImdb' >>> >>> labels = {'pos': 1, 'neg': 0} >>> pbar = pyprind.ProgBar(50000) >>> df = pd.DataFrame() >>> for s in ('test', 'train'): ... for l in ('pos', 'neg'): ... path = os.path.join(basepath, s, l) ... for file in sorted(os.listdir(path)): ... with open(os.path.join(path, file), ... 'r', encoding='utf-8') as infile: ... txt = infile.read() ... df = df.append([[txt, labels[l]]], ... ignore_index=True) ... pbar.update() >>> df.columns = ['review', 'sentiment'] 0% 100% [##############################] | ETA: 00:00:00 Total time elapsed: 00:02:05

In the preceding code, we first initialized a new progress bar object, pbar, with 50,000 iterations, which was the number of documents we were going to read in. Using the nested for loops, we iterated over the train and test subdirectories in the main aclImdb directory and read the individual text files from the pos and neg subdirectories that we eventually appended to the df pandas DataFrame, together with an integer class label (1 = positive and 0 = negative). Since the class labels in the assembled dataset are sorted, we will now shuffle DataFrame using the permutation function from the np.random submodule— this will be useful to split the dataset into training and test datasets in later sections, when we will stream the data from our local drive directly.

[ 261 ]

Applying Machine Learning to Sentiment Analysis

For our own convenience, we will also store the assembled and shuffled movie review dataset as a CSV file: >>> import numpy as np >>> np.random.seed(0) >>> df = df.reindex(np.random.permutation(df.index)) >>> df.to_csv('movie_data.csv', index=False, encoding='utf-8')

Since we are going to use this dataset later in this chapter, let's quickly confirm that we have successfully saved the data in the right format by reading in the CSV and printing an excerpt of the first three examples: >>> df = pd.read_csv('movie_data.csv', encoding='utf-8') >>> df.head(3)

If you are running the code examples in a Jupyter Notebook, you should now see the first three examples of the dataset, as shown in the following table:

As a sanity check, before we proceed to the next section, let's make sure that the DataFrame contains all 50,000 rows: >>> df.shape (50000, 2)

Introducing the bag-of-words model

You may remember from Chapter 4, Building Good Training Datasets – Data Preprocessing, that we have to convert categorical data, such as text or words, into a numerical form before we can pass it on to a machine learning algorithm. In this section, we will introduce the bag-of-words model, which allows us to represent text as numerical feature vectors. The idea behind bag-of-words is quite simple and can be summarized as follows: 1. We create a vocabulary of unique tokens—for example, words—from the entire set of documents. [ 262 ]

Chapter 8

2. We construct a feature vector from each document that contains the counts of how often each word occurs in the particular document. Since the unique words in each document represent only a small subset of all the words in the bag-of-words vocabulary, the feature vectors will mostly consist of zeros, which is why we call them sparse. Do not worry if this sounds too abstract; in the following subsections, we will walk through the process of creating a simple bagof-words model step by step.

Transforming words into feature vectors

To construct a bag-of-words model based on the word counts in the respective documents, we can use the CountVectorizer class implemented in scikit-learn. As you will see in the following code section, CountVectorizer takes an array of text data, which can be documents or sentences, and constructs the bag-of-words model for us: >>> >>> >>> >>> ... ... ... >>>

import numpy as np from sklearn.feature_extraction.text import CountVectorizer count = CountVectorizer() docs = np.array(['The sun is shining', 'The weather is sweet', 'The sun is shining, the weather is sweet,' 'and one and one is two']) bag = count.fit_transform(docs)

By calling the fit_transform method on CountVectorizer, we constructed the vocabulary of the bag-of-words model and transformed the following three sentences into sparse feature vectors: • 'The sun is shining' • 'The weather is sweet' • 'The sun is shining, the weather is sweet, and one and one is two'

Now, let's print the contents of the vocabulary to get a better understanding of the underlying concepts: >>> print(count.vocabulary_) {'and': 0, 'two': 7, 'shining': 3, 'one': 2, 'sun': 4,

[ 263 ]

Applying Machine Learning to Sentiment Analysis 'weather': 8, 'the': 6, 'sweet': 5, 'is': 1}

As you can see from executing the preceding command, the vocabulary is stored in a Python dictionary that maps the unique words to integer indices. Next, let's print the feature vectors that we just created: >>> [[0 [0 [2

print(bag.toarray()) 1 0 1 1 0 1 0 0] 1 0 0 0 1 1 0 1] 3 2 1 1 1 2 1 1]]

Each index position in the feature vectors shown here corresponds to the integer values that are stored as dictionary items in the CountVectorizer vocabulary. For example, the first feature at index position 0 resembles the count of the word 'and', which only occurs in the last document, and the word 'is', at index position 1 (the second feature in the document vectors), occurs in all three sentences. These values in the feature vectors are also called the raw term frequencies: tf(t, d)—the number of times a term, t, occurs in a document, d. It should be noted that, in the bag-ofwords model, the word or term order in a sentence or document does not matter. The order in which the term frequencies appear in the feature vector is derived from the vocabulary indices, which are usually assigned alphabetically. N-gram models The sequence of items in the bag-of-words model that we just created is also called the 1-gram or unigram model—each item or token in the vocabulary represents a single word. More generally, the contiguous sequences of items in NLP—words, letters, or symbols—are also called n-grams. The choice of the number, n, in the n-gram model depends on the particular application; for example, a study by Ioannis Kanaris and others revealed that n-grams of size 3 and 4 yield good performances in the anti-spam filtering of email messages (Words versus character n-grams for anti-spam filtering, Ioannis Kanaris, Konstantinos Kanaris, Ioannis Houvardas, and Efstathios Stamatatos, International Journal on Artificial Intelligence Tools, World Scientific Publishing Company, 16(06): 1047-1067, 2007).

[ 264 ]

Chapter 8

To summarize the concept of the n-gram representation, the 1-gram and 2-gram representations of our first document "the sun is shining" would be constructed as follows: • 1-gram: "the", "sun", "is", "shining" • 2-gram: "the sun", "sun is", "is shining" The CountVectorizer class in scikit-learn allows us to use different n-gram models via its ngram_range parameter. While a 1-gram representation is used by default, we could switch to a 2-gram representation by initializing a new CountVectorizer instance with ngram_range=(2,2).

Assessing word relevancy via term frequency-inverse document frequency

When we are analyzing text data, we often encounter words that occur across multiple documents from both classes. These frequently occurring words typically don't contain useful or discriminatory information. In this subsection, you will learn about a useful technique called the term frequency-inverse document frequency (tf-idf), which can be used to downweight these frequently occurring words in the feature vectors. The tf-idf can be defined as the product of the term frequency and the inverse document frequency: 𝑡𝑡𝑡𝑡-𝑖𝑖𝑖𝑖𝑖𝑖(𝑡𝑡, 𝑑𝑑) = 𝑡𝑡𝑡𝑡(𝑡𝑡, 𝑑𝑑) × 𝑖𝑖𝑖𝑖𝑖𝑖(𝑡𝑡, 𝑑𝑑)

Here, tf(t, d) is the term frequency that we introduced in the previous section, and idf(t, d) is the inverse document frequency, which can be calculated as follows: 𝑖𝑖𝑖𝑖𝑖𝑖(𝑡𝑡, 𝑑𝑑) = log

nd 1 + 𝑑𝑑𝑑𝑑(𝑑𝑑, 𝑡𝑡)

Here, nd is the total number of documents, and df(d, t) is the number of documents, d, that contain the term t. Note that adding the constant 1 to the denominator is optional and serves the purpose of assigning a non-zero value to terms that occur in none of the training examples; the log is used to ensure that low document frequencies are not given too much weight.

[ 265 ]

Applying Machine Learning to Sentiment Analysis

The scikit-learn library implements yet another transformer, the TfidfTransformer class, which takes the raw term frequencies from the CountVectorizer class as input and transforms them into tf-idfs: >>> from sklearn.feature_extraction.text import TfidfTransformer >>> tfidf = TfidfTransformer(use_idf=True, ... norm='l2', ... smooth_idf=True) >>> np.set_printoptions(precision=2) >>> print(tfidf.fit_transform(count.fit_transform(docs)) ... .toarray()) [[ 0. 0.43 0. 0.56 0.56 0. 0.43 0. 0. ] [ 0. 0.43 0. 0. 0. 0.56 0.43 0. 0.56] [ 0.5 0.45 0.5 0.19 0.19 0.19 0.3 0.25 0.19]]

As you saw in the previous subsection, the word 'is' had the largest term frequency in the third document, being the most frequently occurring word. However, after transforming the same feature vector into tf-idfs, the word 'is' is now associated with a relatively small tf-idf (0.45) in the third document, since it is also present in the first and second document and thus is unlikely to contain any useful discriminatory information. However, if we'd manually calculated the tf-idfs of the individual terms in our feature vectors, we would have noticed that TfidfTransformer calculates the tf-idfs slightly differently compared to the standard textbook equations that we defined previously. The equation for the inverse document frequency implemented in scikitlearn is computed as follows: 𝑖𝑖𝑖𝑖𝑖𝑖(𝑡𝑡, 𝑑𝑑) = log

1 + nd 1 + 𝑑𝑑𝑑𝑑(𝑑𝑑, 𝑡𝑡)

Similarly, the tf-idf computed in scikit-learn deviates slightly from the default equation we defined earlier: 𝑡𝑡𝑡𝑡-𝑖𝑖𝑖𝑖𝑖𝑖(𝑡𝑡, 𝑑𝑑) = 𝑡𝑡𝑡𝑡(𝑡𝑡, 𝑑𝑑) × (𝑖𝑖𝑖𝑖𝑖𝑖(𝑡𝑡, 𝑑𝑑) + 1)

Note that the "+1" in the previous equations is due to setting smooth_idf=True in the previous code example, which is helpful for assigning zero-weight (that is, idf(t, d) = log(1) = 0) to terms that occur in all documents. While it is also more typical to normalize the raw term frequencies before calculating the tf-idfs, the TfidfTransformer class normalizes the tf-idfs directly. By default (norm='l2'), scikit-learn's TfidfTransformer applies the L2-normalization, which returns a vector of length 1 by dividing an unnormalized feature vector, v, by its L2-norm: [ 266 ]

Chapter 8

𝑣𝑣𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 =

𝑣𝑣 𝑣𝑣 𝑣𝑣 = = 𝑛𝑛 2 1/2 ‖𝑣𝑣‖2 √𝑣𝑣 2 + 𝑣𝑣 2 + ⋯ + 𝑣𝑣 2 (∑𝑖𝑖=1 𝑣𝑣𝑖𝑖 ) 1 2 𝑛𝑛

To make sure that we understand how TfidfTransformer works, let's walk through an example and calculate the tf-idf of the word 'is' in the third document. The word 'is' has a term frequency of 3 (tf = 3) in the third document, and the document frequency of this term is 3 since the term 'is' occurs in all three documents (df = 3). Thus, we can calculate the inverse document frequency as follows:

𝑖𝑖𝑖𝑖𝑖𝑖("is", 𝑑𝑑3 ) = log

1+3 =0 1+3

Now, in order to calculate the tf-idf, we simply need to add 1 to the inverse document frequency and multiply it by the term frequency:

𝑡𝑡𝑡𝑡-𝑖𝑖𝑖𝑖𝑖𝑖("is", 𝑑𝑑3 ) = 3 × (0 + 1) = 3

If we repeated this calculation for all terms in the third document, we'd obtain the following tf-idf vectors: [3.39, 3.0, 3.39, 1.29, 1.29, 1.29, 2.0, 1.69, 1.29]. However, notice that the values in this feature vector are different from the values that we obtained from TfidfTransformer that we used previously. The final step that we are missing in this tf-idf calculation is the L2-normalization, which can be applied as follows: 𝑡𝑡𝑡𝑡-𝑖𝑖𝑖𝑖𝑖𝑖(𝑑𝑑3)𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛

=

3.392

[3.39, 3.0, 3.39, 1.29, 1.29, 1.29, 2.0, 1.69, 1.29]

+ 3.02 + 3.392 + 1.292 + 1.292 + 1.292 + 2.02 + 1.692 + 1.292 √ = [0.5, 0.45, 0.5, 0.19, 0.19, 0.19, 0.3, 0.25, 0.19]

𝑡𝑡𝑡𝑡-𝑖𝑖𝑖𝑖𝑖𝑖("is", 𝑑𝑑3 ) = 0.45

As you can see, the results now match the results returned by scikit-learn's TfidfTransformer, and since you now understand how tf-idfs are calculated, let's proceed to the next section and apply those concepts to the movie review dataset.

Cleaning text data

In the previous subsections, we learned about the bag-of-words model, term frequencies, and tf-idfs. However, the first important step—before we build our bagof-words model—is to clean the text data by stripping it of all unwanted characters.

[ 267 ]

Applying Machine Learning to Sentiment Analysis

To illustrate why this is important, let's display the last 50 characters from the first document in the reshuffled movie review dataset: >>> df.loc[0, 'review'][-50:] 'is seven.

Title (Brazil): Not Available'

As you can see here, the text contains HTML markup as well as punctuation and other non-letter characters. While HTML markup does not contain many useful semantics, punctuation marks can represent useful, additional information in certain NLP contexts. However, for simplicity, we will now remove all punctuation marks except for emoticon characters, such as :), since those are certainly useful for sentiment analysis. To accomplish this task, we will use Python's regular expression (regex) library, re, as shown here: >>> import re >>> def preprocessor(text): ... text = re.sub(']*>', '', text) ... emoticons = re.findall('(?::|;|=)(?:-)?(?:\)|\(|D|P)', ... text) ... text = (re.sub('[\W]+', ' ', text.lower()) + ... ' '.join(emoticons).replace('-', '')) ... return text

Via the first regex, ]*>, in the preceding code section, we tried to remove all of the HTML markup from the movie reviews. Although many programmers generally advise against the use of regex to parse HTML, this regex should be sufficient to clean this particular dataset. Since we are only interested in removing HTML markup and do not plan to use the HTML markup further, using regex to do the job should be acceptable. However, if you prefer using sophisticated tools for removing HTML markup from text, you can take a look at Python's HTML parser module, which is described at https://docs.python.org/3/library/html.parser.html. After we removed the HTML markup, we used a slightly more complex regex to find emoticons, which we temporarily stored as emoticons. Next, we removed all non-word characters from the text via the regex [\W]+ and converted the text into lowercase characters. Dealing with word capitalization In the context of this analysis, we assume that the capitalization of a word—for example, whether it appears at the beginning of a sentence—does not contain semantically relevant information. However, note that there are exceptions; for instance, we remove the notation of proper names. But again, in the context of this analysis, it is a simplifying assumption that the letter case does not contain information that is relevant for sentiment analysis. [ 268 ]

Chapter 8

Eventually, we added the temporarily stored emoticons to the end of the processed document string. Additionally, we removed the nose character (- in :-)) from the emoticons for consistency. Regular expressions Although regular expressions offer an efficient and convenient approach to searching for characters in a string, they also come with a steep learning curve. Unfortunately, an in-depth discussion of regular expressions is beyond the scope of this book. However, you can find a great tutorial on the Google Developers portal at https://developers.google.com/ edu/python/regular-expressions or you can check out the official documentation of Python's re module at https://docs.python.org/3.7/library/re.html.

Although the addition of the emoticon characters to the end of the cleaned document strings may not look like the most elegant approach, we must note that the order of the words doesn't matter in our bag-of-words model if our vocabulary consists of only one-word tokens. But before we talk more about the splitting of documents into individual terms, words, or tokens, let's confirm that our preprocessor function works correctly: >>> preprocessor(df.loc[0, 'review'][-50:]) 'is seven title brazil not available' >>> preprocessor("This :) is :( a test :-)!") 'this is a test :) :( :)'

Lastly, since we will make use of the cleaned text data over and over again during the next sections, let's now apply our preprocessor function to all the movie reviews in our DataFrame: >>> df['review'] = df['review'].apply(preprocessor)

Processing documents into tokens

After successfully preparing the movie review dataset, we now need to think about how to split the text corpora into individual elements. One way to tokenize documents is to split them into individual words by splitting the cleaned documents at their whitespace characters: >>> def tokenizer(text): ... return text.split() >>> tokenizer('runners like running and thus they run') ['runners', 'like', 'running', 'and', 'thus', 'they', 'run'] [ 269 ]

Applying Machine Learning to Sentiment Analysis

In the context of tokenization, another useful technique is word stemming, which is the process of transforming a word into its root form. It allows us to map related words to the same stem. The original stemming algorithm was developed by Martin F. Porter in 1979 and is hence known as the Porter stemmer algorithm (An algorithm for suffix stripping, Martin F. Porter, Program: Electronic Library and Information Systems, 14(3): 130–137, 1980). The Natural Language Toolkit (NLTK, http://www.nltk.org) for Python implements the Porter stemming algorithm, which we will use in the following code section. In order to install the NLTK, you can simply execute conda install nltk or pip install nltk. NLTK online book Although the NLTK is not the focus of this chapter, I highly recommend that you visit the NLTK website as well as read the official NLTK book, which is freely available at http://www.nltk.org/book/, if you are interested in more advanced applications in NLP.

The following code shows how to use the Porter stemming algorithm: >>> from nltk.stem.porter import PorterStemmer >>> porter = PorterStemmer() >>> def tokenizer_porter(text): ... return [porter.stem(word) for word in text.split()] >>> tokenizer_porter('runners like running and thus they run') ['runner', 'like', 'run', 'and', 'thu', 'they', 'run']

Using the PorterStemmer from the nltk package, we modified our tokenizer function to reduce words to their root form, which was illustrated by the simple preceding example where the word 'running' was stemmed to its root form 'run'. Stemming algorithms The Porter stemming algorithm is probably the oldest and simplest stemming algorithm. Other popular stemming algorithms include the newer Snowball stemmer (Porter2 or English stemmer) and the Lancaster stemmer (Paice/Husk stemmer). While both the Snowball and Lancaster stemmers are faster than the original Porter stemmer, the Lancaster stemmer is also notorious for being more aggressive than the Porter stemmer. These alternative stemming algorithms are also available through the NLTK package (http://www.nltk.org/ api/nltk.stem.html).

[ 270 ]

Chapter 8

While stemming can create non-real words, such as 'thu' (from 'thus'), as shown in the previous example, a technique called lemmatization aims to obtain the canonical (grammatically correct) forms of individual words—the socalled lemmas. However, lemmatization is computationally more difficult and expensive compared to stemming and, in practice, it has been observed that stemming and lemmatization have little impact on the performance of text classification (Influence of Word Normalization on Text Classification, Michal Toman, Roman Tesar, and Karel Jezek, Proceedings of InSciT, pages 354–358, 2006).

Before we jump into the next section, where we will train a machine learning model using the bag-of-words model, let's briefly talk about another useful topic called stop-word removal. Stop-words are simply those words that are extremely common in all sorts of texts and probably bear no (or only a little) useful information that can be used to distinguish between different classes of documents. Examples of stopwords are is, and, has, and like. Removing stop-words can be useful if we are working with raw or normalized term frequencies rather than tf-idfs, which are already downweighting frequently occurring words. In order to remove stop-words from the movie reviews, we will use the set of 127 English stop-words that is available from the NLTK library, which can be obtained by calling the nltk.download function: >>> import nltk >>> nltk.download('stopwords') After we download the stop-words set, we can load and apply the English stop-word set as follows: >>> from nltk.corpus import stopwords >>> stop = stopwords.words('english') >>> [w for w in tokenizer_porter('a runner likes' ... ' running and runs a lot')[-10:] ... if w not in stop] ['runner', 'like', 'run', 'run', 'lot']

[ 271 ]

Applying Machine Learning to Sentiment Analysis

Training a logistic regression model for document classification

In this section, we will train a logistic regression model to classify the movie reviews into positive and negative reviews based on the bag-of-words model. First, we will divide the DataFrame of cleaned text documents into 25,000 documents for training and 25,000 documents for testing: >>> >>> >>> >>>

X_train = df.loc[:25000, 'review'].values y_train = df.loc[:25000, 'sentiment'].values X_test = df.loc[25000:, 'review'].values y_test = df.loc[25000:, 'sentiment'].values

Next, we will use a GridSearchCV object to find the optimal set of parameters for our logistic regression model using 5-fold stratified cross-validation: >>> >>> >>> >>>

from from from from

sklearn.model_selection import GridSearchCV sklearn.pipeline import Pipeline sklearn.linear_model import LogisticRegression sklearn.feature_extraction.text import TfidfVectorizer

>>> ... ... >>> ... ... ... ... ... ... ... ... ... ... ... ... ... ... >>> ... ... ... >>> ... ... ...

tfidf = TfidfVectorizer(strip_accents=None, lowercase=False, preprocessor=None) param_grid = [{'vect__ngram_range': [(1,1)], 'vect__stop_words': [stop, None], 'vect__tokenizer': [tokenizer, tokenizer_porter], 'clf__penalty': ['l1', 'l2'], 'clf__C': [1.0, 10.0, 100.0]}, {'vect__ngram_range': [(1,1)], 'vect__stop_words': [stop, None], 'vect__tokenizer': [tokenizer, tokenizer_porter], 'vect__use_idf':[False], 'vect__norm':[None], 'clf__penalty': ['l1', 'l2'], 'clf__C': [1.0, 10.0, 100.0]} ] lr_tfidf = Pipeline([('vect', tfidf), ('clf', LogisticRegression(random_state=0, solver='liblinear'))]) gs_lr_tfidf = GridSearchCV(lr_tfidf, param_grid, scoring='accuracy', cv=5, verbose=2, n_jobs=1) [ 272 ]

Chapter 8 >>> gs_lr_tfidf.fit(X_train, y_train)

Multiprocessing via the n_jobs parameter Please note that it is highly recommended to set n_jobs=-1 (instead of n_jobs=1) in the previous code example to utilize all available cores on your machine and speed up the grid search. However, some Windows users reported issues when running the previous code with the n_jobs=-1 setting related to pickling the tokenizer and tokenizer_porter functions for multiprocessing on Windows. Another workaround would be to replace those two functions, [tokenizer, tokenizer_porter], with [str.split]. However, note that replacement by the simple str.split would not support stemming.

When we initialized the GridSearchCV object and its parameter grid using the preceding code, we restricted ourselves to a limited number of parameter combinations, since the number of feature vectors, as well as the large vocabulary, can make the grid search computationally quite expensive. Using a standard desktop computer, our grid search may take up to 40 minutes to complete. In the previous code example, we replaced CountVectorizer and TfidfTransformer from the previous subsection with TfidfVectorizer, which combines CountVectorizer with the TfidfTransformer. Our param_grid consisted of two parameter dictionaries. In the first dictionary, we used TfidfVectorizer with its default settings (use_idf=True, smooth_idf=True, and norm='l2') to calculate the tf-idfs; in the second dictionary, we set those parameters to use_idf=False, smooth_idf=False, and norm=None in order to train a model based on raw term frequencies. Furthermore, for the logistic regression classifier itself, we trained models using L2 and L1 regularization via the penalty parameter and compared different regularization strengths by defining a range of values for the inverse-regularization parameter C. After the grid search has finished, we can print the best parameter set: >>> print('Best parameter set: %s ' % gs_lr_tfidf.best_params_) Best parameter set: {'clf__C': 10.0, 'vect__stop_words': None, 'clf__penalty': 'l2', 'vect__tokenizer': , 'vect__ngram_range': (1, 1)}

As you can see in the preceding output, we obtained the best grid search results using the regular tokenizer without Porter stemming, no stop-word library, and tf-idfs in combination with a logistic regression classifier that uses L2-regularization with the regularization strength C of 10.0. [ 273 ]

Applying Machine Learning to Sentiment Analysis

Using the best model from this grid search, let's print the average 5-fold crossvalidation accuracy scores on the training dataset and the classification accuracy on the test dataset: >>> print('CV Accuracy: %.3f' ... % gs_lr_tfidf.best_score_) CV Accuracy: 0.897 >>> clf = gs_lr_tfidf.best_estimator_ >>> print('Test Accuracy: %.3f' ... % clf.score(X_test, y_test)) Test Accuracy: 0.899

The results reveal that our machine learning model can predict whether a movie review is positive or negative with 90 percent accuracy. The naïve Bayes classifier A still very popular classifier for text classification is the naïve Bayes classifier, which gained popularity in applications of email spam filtering. Naïve Bayes classifiers are easy to implement, computationally efficient, and tend to perform particularly well on relatively small datasets compared to other algorithms. Although we don't discuss naïve Bayes classifiers in this book, the interested reader can find an article about naïve Bayes text classification that is freely available on arXiv (Naive Bayes and Text Classification I – Introduction and Theory, S. Raschka, Computing Research Repository (CoRR), abs/1410.5329, 2014, http://arxiv.org/pdf/1410.5329v3.pdf).

Working with bigger data – online algorithms and out-of-core learning

If you executed the code examples in the previous section, you may have noticed that it could be computationally quite expensive to construct the feature vectors for the 50,000-movie review dataset during grid search. In many real-world applications, it is not uncommon to work with even larger datasets that can exceed our computer's memory. Since not everyone has access to supercomputer facilities, we will now apply a technique called out-of-core learning, which allows us to work with such large datasets by fitting the classifier incrementally on smaller batches of a dataset.

[ 274 ]

Chapter 8

Text classification with recurrent neural networks In Chapter 16, Modeling Sequential Data Using Recurrent Neural Networks, we will revisit this dataset and train a deep learningbased classifier (a recurrent neural network) to classify the reviews in the IMDb movie review dataset. This neural networkbased classifier follows the same out-of-core principle using the stochastic gradient descent optimization algorithm but does not require the construction of a bag-of-words model.

Back in Chapter 2, Training Simple Machine Learning Algorithms for Classification, the concept of stochastic gradient descent was introduced; it is an optimization algorithm that updates the model's weights using one example at a time. In this section, we will make use of the partial_fit function of SGDClassifier in scikitlearn to stream the documents directly from our local drive and train a logistic regression model using small mini-batches of documents. First, we will define a tokenizer function that cleans the unprocessed text data from the movie_data.csv file that we constructed at the beginning of this chapter and separate it into word tokens while removing stop-words: >>> >>> >>> >>> >>> ... ... ... ... ... ... ...

import numpy as np import re from nltk.corpus import stopwords stop = stopwords.words('english') def tokenizer(text): text = re.sub(']*>', '', text) emoticons = re.findall('(?::|;|=)(?:-)?(?:\)|\(|D|P)', text.lower()) text = re.sub('[\W]+', ' ', text.lower()) \ + ' '.join(emoticons).replace('-', '') tokenized = [w for w in text.split() if w not in stop] return tokenized

Next, we will define a generator function, stream_docs, that reads in and returns one document at a time: >>> def stream_docs(path): ... with open(path, 'r', encoding='utf-8') as csv: ... next(csv) # skip header ... for line in csv: ... text, label = line[:-3], int(line[-2]) ... yield text, label

[ 275 ]

Applying Machine Learning to Sentiment Analysis

To verify that our stream_docs function works correctly, let's read in the first document from the movie_data.csv file, which should return a tuple consisting of the review text as well as the corresponding class label: >>> next(stream_docs(path='movie_data.csv')) ('"In 1974, the teenager Martha Moxley ... ',1)

We will now define a function, get_minibatch, that will take a document stream from the stream_docs function and return a particular number of documents specified by the size parameter: >>> def get_minibatch(doc_stream, size): ... docs, y = [], [] ... try: ... for _ in range(size): ... text, label = next(doc_stream) ... docs.append(text) ... y.append(label) ... except StopIteration: ... return None, None ... return docs, y

Unfortunately, we can't use CountVectorizer for out-of-core learning since it requires holding the complete vocabulary in memory. Also, TfidfVectorizer needs to keep all the feature vectors of the training dataset in memory to calculate the inverse document frequencies. However, another useful vectorizer for text processing implemented in scikit-learn is HashingVectorizer. HashingVectorizer is data-independent and makes use of the hashing trick via the 32-bit MurmurHash3 function by Austin Appleby (https://sites.google.com/site/murmurhash/): >>> >>> >>> ... ... ... >>> >>>

from sklearn.feature_extraction.text import HashingVectorizer from sklearn.linear_model import SGDClassifier vect = HashingVectorizer(decode_error='ignore', n_features=2**21, preprocessor=None, tokenizer=tokenizer) clf = SGDClassifier(loss='log', random_state=1) doc_stream = stream_docs(path='movie_data.csv')

Using the preceding code, we initialized HashingVectorizer with our tokenizer function and set the number of features to 2**21. Furthermore, we reinitialized a logistic regression classifier by setting the loss parameter of SGDClassifier to 'log'. Note that by choosing a large number of features in HashingVectorizer, we reduce the chance of causing hash collisions, but we also increase the number of coefficients in our logistic regression model.

[ 276 ]

Chapter 8

Now comes the really interesting part – having set up all the complementary functions, we can start the out-of-core learning using the following code: >>> import pyprind >>> pbar = pyprind.ProgBar(45) >>> classes = np.array([0, 1]) >>> for _ in range(45): ... X_train, y_train = get_minibatch(doc_stream, size=1000) ... if not X_train: ... break ... X_train = vect.transform(X_train) ... clf.partial_fit(X_train, y_train, classes=classes) ... pbar.update() 0% 100% [##############################] | ETA: 00:00:00 Total time elapsed: 00:00:21

Again, we made use of the PyPrind package in order to estimate the progress of our learning algorithm. We initialized the progress bar object with 45 iterations and, in the following for loop, we iterated over 45 mini-batches of documents where each mini-batch consists of 1,000 documents. Having completed the incremental learning process, we will use the last 5,000 documents to evaluate the performance of our model: >>> X_test, y_test = get_minibatch(doc_stream, size=5000) >>> X_test = vect.transform(X_test) >>> print('Accuracy: %.3f' % clf.score(X_test, y_test)) Accuracy: 0.868

As you can see, the accuracy of the model is approximately 87 percent, slightly below the accuracy that we achieved in the previous section using the grid search for hyperparameter tuning. However, out-of-core learning is very memory efficient and it took less than a minute to complete. Finally, we can use the last 5,000 documents to update our model: >>> clf = clf.partial_fit(X_test, y_test)

The word2vec model A more modern alternative to the bag-of-words model is word2vec, an algorithm that Google released in 2013 (Efficient Estimation of Word Representations in Vector Space, T. Mikolov, K. Chen, G. Corrado, and J. Dean, arXiv preprint arXiv:1301.3781, 2013).

[ 277 ]

Applying Machine Learning to Sentiment Analysis

The word2vec algorithm is an unsupervised learning algorithm based on neural networks that attempts to automatically learn the relationship between words. The idea behind word2vec is to put words that have similar meanings into similar clusters, and via clever vector-spacing, the model can reproduce certain words using simple vector math, for example, king – man + woman = queen. The original C-implementation with useful links to the relevant papers and alternative implementations can be found at https://code.google.com/p/word2vec/.

Topic modeling with Latent Dirichlet Allocation

Topic modeling describes the broad task of assigning topics to unlabeled text documents. For example, a typical application would be the categorization of documents in a large text corpus of newspaper articles. In applications of topic modeling, we then aim to assign category labels to those articles, for example, sports, finance, world news, politics, local news, and so forth. Thus, in the context of the broad categories of machine learning that we discussed in Chapter 1, Giving Computers the Ability to Learn from Data, we can consider topic modeling as a clustering task, a subcategory of unsupervised learning. In this section, we will discuss a popular technique for topic modeling called Latent Dirichlet Allocation (LDA). However, note that while Latent Dirichlet Allocation is often abbreviated as LDA, it is not to be confused with linear discriminant analysis, a supervised dimensionality reduction technique that was introduced in Chapter 5, Compressing Data via Dimensionality Reduction. Embedding the movie review classifier into a web application LDA is different from the supervised learning approach that we took in this chapter to classify movie reviews as positive and negative. Thus, if you are interested in embedding scikit-learn models into a web application via the Flask framework using the movie reviewer as an example, please feel free to jump to the next chapter and revisit this standalone section on topic modeling later on.

[ 278 ]

Chapter 8

Decomposing text documents with LDA

Since the mathematics behind LDA is quite involved and requires knowledge about Bayesian inference, we will approach this topic from a practitioner's perspective and interpret LDA using layman's terms. However, the interested reader can read more about LDA in the following research paper: Latent Dirichlet Allocation, David M. Blei, Andrew Y. Ng, and Michael I. Jordan, Journal of Machine Learning Research 3, pages: 9931022, Jan 2003. LDA is a generative probabilistic model that tries to find groups of words that appear frequently together across different documents. These frequently appearing words represent our topics, assuming that each document is a mixture of different words. The input to an LDA is the bag-of-words model that we discussed earlier in this chapter. Given a bag-of-words matrix as input, LDA decomposes it into two new matrices: • A document-to-topic matrix • A word-to-topic matrix LDA decomposes the bag-of-words matrix in such a way that if we multiply those two matrices together, we will be able to reproduce the input, the bag-of-words matrix, with the lowest possible error. In practice, we are interested in those topics that LDA found in the bag-of-words matrix. The only downside may be that we must define the number of topics beforehand—the number of topics is a hyperparameter of LDA that has to be specified manually.

LDA with scikit-learn

In this subsection, we will use the LatentDirichletAllocation class implemented in scikit-learn to decompose the movie review dataset and categorize it into different topics. In the following example, we will restrict the analysis to 10 different topics, but readers are encouraged to experiment with the hyperparameters of the algorithm to further explore the topics that can be found in this dataset. First, we are going to load the dataset into a pandas DataFrame using the local movie_data.csv file of the movie reviews that we created at the beginning of this

chapter:

>>> import pandas as pd >>> df = pd.read_csv('movie_data.csv', encoding='utf-8')

Next, we are going to use the already familiar CountVectorizer to create the bag-ofwords matrix as input to the LDA.

[ 279 ]

Applying Machine Learning to Sentiment Analysis

For convenience, we will use scikit-learn's built-in English stop-word library via stop_words='english': >>> from sklearn.feature_extraction.text import CountVectorizer >>> count = CountVectorizer(stop_words='english', ... max_df=.1, ... max_features=5000) >>> X = count.fit_transform(df['review'].values)

Notice that we set the maximum document frequency of words to be considered to 10 percent (max_df=.1) to exclude words that occur too frequently across documents. The rationale behind the removal of frequently occurring words is that these might be common words appearing across all documents that are, therefore, less likely to be associated with a specific topic category of a given document. Also, we limited the number of words to be considered to the most frequently occurring 5,000 words (max_features=5000), to limit the dimensionality of this dataset to improve the inference performed by LDA. However, both max_df=.1 and max_features=5000 are hyperparameter values chosen arbitrarily, and readers are encouraged to tune them while comparing the results. The following code example demonstrates how to fit a LatentDirichletAllocation estimator to the bag-of-words matrix and infer the 10 different topics from the documents (note that the model fitting can take up to 5 minutes or more on a laptop or standard desktop computer): >>> from sklearn.decomposition import LatentDirichletAllocation >>> lda = LatentDirichletAllocation(n_components=10, ... random_state=123, ... learning_method='batch') >>> X_topics = lda.fit_transform(X)

By setting learning_method='batch', we let the lda estimator do its estimation based on all available training data (the bag-of-words matrix) in one iteration, which is slower than the alternative 'online' learning method but can lead to more accurate results (setting learning_method='online' is analogous to online or minibatch learning, which we discussed in Chapter 2, Training Simple Machine Learning Algorithms for Classification, and in this chapter).

[ 280 ]

Chapter 8

Expectation-Maximization The scikit-learn library's implementation of LDA uses the expectation-maximization (EM) algorithm to update its parameter estimates iteratively. We haven't discussed the EM algorithm in this chapter, but if you are curious to learn more, please see the excellent overview on Wikipedia (https:// en.wikipedia.org/wiki/Expectation–maximization_ algorithm) and the detailed tutorial on how it is used in

LDA in Colorado Reed's tutorial, Latent Dirichlet Allocation: Towards a Deeper Understanding, which is freely available at http://obphio.us/pdfs/lda_tutorial.pdf.

After fitting the LDA, we now have access to the components_ attribute of the lda instance, which stores a matrix containing the word importance (here, 5000) for each of the 10 topics in increasing order: >>> lda.components_.shape (10, 5000)

To analyze the results, let's print the five most important words for each of the 10 topics. Note that the word importance values are ranked in increasing order. Thus, to print the top five words, we need to sort the topic array in reverse order: >>> n_top_words = 5 >>> feature_names = count.get_feature_names() >>> for topic_idx, topic in enumerate(lda.components_): ... print("Topic %d:" % (topic_idx + 1)) ... print(" ".join([feature_names[i] ... for i in topic.argsort()\ ... [:-n_top_words - 1:-1]])) Topic 1: worst minutes awful script stupid Topic 2: family mother father children girl Topic 3: american war dvd music tv Topic 4: human audience cinema art sense Topic 5: police guy car dead murder Topic 6: horror house sex girl woman [ 281 ]

Applying Machine Learning to Sentiment Analysis Topic 7: role performance comedy actor performances Topic 8: series episode war episodes tv Topic 9: book version original read novel Topic 10: action fight guy guys cool

Based on reading the five most important words for each topic, you may guess that the LDA identified the following topics: 1. Generally bad movies (not really a topic category) 2. Movies about families 3. War movies 4. Art movies 5. Crime movies 6. Horror movies 7. Comedy movies reviews 8. Movies somehow related to TV shows 9. Movies based on books 10. Action movies To confirm that the categories make sense based on the reviews, let's plot three movies from the horror movie category (horror movies belong to category 6 at index position 5): >>> horror = X_topics[:, 5].argsort()[::-1] >>> for iter_idx, movie_idx in enumerate(horror[:3]): ... print('\nHorror movie #%d:' % (iter_idx + 1)) ... print(df['review'][movie_idx][:300], '...') Horror movie #1: House of Dracula works from the same basic premise as House of Frankenstein from the year before; namely that Universal's three most famous monsters; Dracula, Frankenstein's Monster and The Wolf Man are appearing in the movie together. Naturally, the film is rather messy therefore, but the fact that ... Horror movie #2: Okay, what the hell kind of TRASH have I been watching now? Witches' Mountain" has got to be one of the most incoherent Spanish exploitation flicks ever and yet, at the same time, strangely compelling. There's absolutely nothing that makes and I even doubt there ... [ 282 ]

"The and insane it's also sense here

Chapter 8

Horror movie #3:

Horror movie time, Japanese style. Uzumaki/Spiral was a total freakfest from start to finish. A fun freakfest at that, but at times it was a tad too reliant on kitsch rather than the horror. The story is difficult to summarize succinctly: a carefree, normal teenage girl starts coming fac ...

Using the preceding code example, we printed the first 300 characters from the top three horror movies. The reviews—even though we don't know which exact movie they belong to—sound like reviews of horror movies (however, one might argue that Horror movie #2 could also be a good fit for topic category 1: Generally bad movies).

Summary

In this chapter, you learned how to use machine learning algorithms to classify text documents based on their polarity, which is a basic task in sentiment analysis in the field of NLP. Not only did you learn how to encode a document as a feature vector using the bag-of-words model, but you also learned how to weight the term frequency by relevance using tf-idf. Working with text data can be computationally quite expensive due to the large feature vectors that are created during this process; in the last section, we covered how to utilize out-of-core or incremental learning to train a machine learning algorithm without loading the whole dataset into a computer's memory. Lastly, you were introduced to the concept of topic modeling using LDA to categorize the movie reviews into different categories in an unsupervised fashion. In the next chapter, we will use our document classifier and learn how to embed it into a web application.

[ 283 ]

Embedding a Machine Learning Model into a Web Application In the previous chapters, you learned about the many different machine learning concepts and algorithms that can help us with better and more efficient decisionmaking. However, machine learning techniques are not limited to offline applications and analyses, and they have become the predictive engine of various web services. For example, popular and useful applications of machine learning models in web applications include spam detection in submission forms, search engines, recommendation systems for media or shopping portals, and many more. In this chapter, you will learn how to embed a machine learning model into a web application that can not only classify, but also learn from data in real time. The topics that we will cover are as follows: • Saving the current state of a trained machine learning model • Using SQLite databases for data storage • Developing a web application using the popular Flask web framework • Deploying a machine learning application to a public web server

Serializing fitted scikit-learn estimators

Training a machine learning model can be computationally expensive, as you saw in Chapter 8, Applying Machine Learning to Sentiment Analysis. Surely, we don't want to retrain our model every time we close our Python interpreter and want to make a new prediction or reload our web application?

[ 285 ]

Embedding a Machine Learning Model into a Web Application

One option for model persistence is Python's in-built pickle module (https:// docs.python.org/3.7/library/pickle.html), which allows us to serialize and

deserialize Python object structures to compact bytecode so that we can save our classifier in its current state and reload it if we want to classify new, unlabeled examples, without needing the model to learn from the training data all over again. Before you execute the following code, please make sure that you have trained the out-of-core logistic regression model from the last section of Chapter 8 and have it ready in your current Python session: >>> >>> >>> >>> ... >>> ... ... >>> ... ...

import pickle import os dest = os.path.join('movieclassifier', 'pkl_objects') if not os.path.exists(dest): os.makedirs(dest) pickle.dump(stop, open(os.path.join(dest, 'stopwords.pkl'), 'wb'), protocol=4) pickle.dump(clf, open(os.path.join(dest, 'classifier.pkl'), 'wb'), protocol=4)

Using the preceding code, we created a movieclassifier directory where we will later store the files and data for our web application. Within this movieclassifier directory, we created a pkl_objects subdirectory to save the serialized Python objects to our local hard drive or solid-state drive. Via the dump method of the pickle module, we then serialized the trained logistic regression model as well as the stop-word set from the Natural Language Toolkit (NLTK) library, so that we don't have to install the NLTK vocabulary on our server. The dump method takes as its first argument the object that we want to pickle. For the second argument, we provided an open file object that the Python object will be written to. Via the wb argument inside the open function, we opened the file in binary mode for pickle, and we set protocol=4 to choose the latest and most efficient pickle protocol that was added to Python 3.4, which is compatible with Python 3.4 or newer. If you have problems using protocol=4, please check whether you are using the latest Python 3 version—Python 3.7 is recommended for this book. Alternatively, you may consider choosing a lower protocol number. Also note that if you are using a custom web server, you have to ensure that the Python installation on that server is compatible with this protocol version as well.

[ 286 ]

Chapter 9

Serializing NumPy arrays with joblib Our logistic regression model contains several NumPy arrays, such as the weight vector, and a more efficient way to serialize NumPy arrays is to use the alternative joblib library. To ensure compatibility with the server environment that we will use in later sections, we will use the standard pickle approach. If you are interested, you can find more information about joblib at https://joblib.readthedocs.io.

We don't need to pickle HashingVectorizer, since it does not need to be fitted. Instead, we can create a new Python script file from which we can import the vectorizer into our current Python session. Now, copy the following code and save it as vectorizer.py in the movieclassifier directory: from sklearn.feature_extraction.text import HashingVectorizer import re import os import pickle cur_dir = os.path.dirname(__file__) stop = pickle.load(open(os.path.join( cur_dir, 'pkl_objects', 'stopwords.pkl'), 'rb')) def tokenizer(text): text = re.sub(']*>', '', text) emoticons = re.findall('(?::|;|=)(?:-)?(?:\)|\(|D|P)', text.lower()) text = re.sub('[\W]+', ' ', text.lower()) \ + ' '.join(emoticons).replace('-', '') tokenized = [w for w in text.split() if w not in stop] return tokenized vect = HashingVectorizer(decode_error='ignore', n_features=2**21, preprocessor=None, tokenizer=tokenizer)

After we have pickled the Python objects and created the vectorizer.py file, it would be a good idea to restart our Python interpreter or Jupyter Notebook kernel to test whether we can deserialize the objects without error.

[ 287 ]

Embedding a Machine Learning Model into a Web Application

Pickle can be a security risk Please note that unpickling data from an untrusted source can be a potential security risk, since the pickle module is not secured against malicious code. Since pickle was designed to serialize arbitrary objects, the unpickling process will execute code that has been stored in a pickle file. Thus, if you receive pickle files from an untrusted source (for example, by downloading them from the internet), please proceed with extra care and unpickle the items in a virtual environment and/or on a non-essential machine that does not store important data that no one except you should have access to.

From your terminal, navigate to the movieclassifier directory, start a new Python session, and execute the following code to verify that you can import the vectorizer and unpickle the classifier: >>> >>> >>> >>> >>> ... ...

import pickle import re import os from vectorizer import vect clf = pickle.load(open(os.path.join( 'pkl_objects', 'classifier.pkl'), 'rb'))

After we have successfully loaded the vectorizer and unpickled the classifier, we can use these objects to preprocess document examples and make predictions about their sentiments: >>> import numpy as np >>> label = {0:'negative', 1:'positive'} >>> example = ["I love this movie. It's amazing."] >>> X = vect.transform(example) >>> print('Prediction: %s\nProbability: %.2f%%' %\ ... (label[clf.predict(X)[0]], ... np.max(clf.predict_proba(X))*100)) Prediction: positive Probability: 95.55%

Since our classifier returns the class label predictions as integers, we defined a simple Python dictionary to map these integers to their sentiment ("positive" or "negative"). While this is a simple application with two classes only, it should be noted that this dictionary-mapping approach also generalizes to multiclass settings. Furthermore, this mapping dictionary should also be archived alongside the model. [ 288 ]

Chapter 9

In this case, since the dictionary definition only consists of one line of code, we will not go to the trouble of serializing it using pickle. However, in realworld applications with more extensive mapping dictionaries, you can utilize the same pickle.dump and pickle.load commands that we used in the previous code example. Continuing with the discussion of the previous code example, we then used

HashingVectorizer to transform the simple example document into a word vector, X. Finally, we used the predict method of the logistic regression classifier to predict the class label, as well as the predict_proba method to return the corresponding probability of our prediction. Note that the predict_proba method call returns an

array with a probability value for each unique class label. Since the class label with the largest probability corresponds to the class label that is returned by the predict call, we used the np.max function to return the probability of the predicted class.

Setting up an SQLite database for data storage

In this section, we will set up a simple SQLite database to collect optional feedback about the predictions from users of the web application. We can use this feedback to update our classification model. SQLite is an open source SQL database engine that doesn't require a separate server to operate, which makes it ideal for smaller projects and simple web applications. Essentially, an SQLite database can be understood as a single, self-contained database file that allows us to directly access storage files. Furthermore, SQLite doesn't require any system-specific configuration and is supported by all common operating systems. It has gained a reputation for being very reliable and is used by popular companies such as Google, Mozilla, Adobe, Apple, Microsoft, and many more. If you want to learn more about SQLite, visit the official website at http://www.sqlite.org. Fortunately, following Python's batteries included philosophy, there is already an API in the Python standard library, sqlite3, which allows us to work with SQLite databases. (For more information about sqlite3, please visit https://docs.python.org/3.7/library/sqlite3.html.) By executing the following code, we will create a new SQLite database inside the movieclassifier directory and store two example movie reviews: >>> import sqlite3 >>> import os >>> conn = sqlite3.connect('reviews.sqlite') [ 289 ]

Embedding a Machine Learning Model into a Web Application >>> c = conn.cursor() >>> c.execute('DROP TABLE IF EXISTS review_db') >>> c.execute('CREATE TABLE review_db'\ ... ' (review TEXT, sentiment INTEGER, date TEXT)') >>> example1 = 'I love this movie' >>> c.execute("INSERT INTO review_db"\ ... " (review, sentiment, date) VALUES"\ ... " (?, ?, DATETIME('now'))", (example1, 1)) >>> >>> ... ... >>> >>>

example2 = 'I disliked this movie' c.execute("INSERT INTO review_db"\ " (review, sentiment, date) VALUES"\ " (?, ?, DATETIME('now'))", (example2, 0)) conn.commit() conn.close()

Following the preceding code example, we created a connection (conn) to an SQLite database file by calling the connect method of the sqlite3 library, which created the new database file reviews.sqlite in the movieclassifier directory if it didn't already exist. Next, we created a cursor via the cursor method, which allows us to traverse over the database records using the versatile SQL syntax. Via the first execute call, we then created a new database table, review_db. We used this to store and access database entries. Along with review_db, we also created three columns in this database table: review, sentiment, and date. We used these to store two example movie reviews and respective class labels (sentiments). Using the DATETIME('now') SQL command, we also added date and timestamps to our entries. In addition to the timestamps, we used the question mark symbols (?) to pass the movie review texts (example1 and example2) and the corresponding class labels (1 and 0) as positional arguments to the execute method, as members of a tuple. Lastly, we called the commit method to save the changes that we made to the database and closed the connection via the close method. To check if the entries have been stored in the database table correctly, we will now reopen the connection to the database and use the SQL SELECT command to fetch all rows in the database table that have been committed between the beginning of the year 2017 and today: >>> >>> >>> ... >>>

conn = sqlite3.connect('reviews.sqlite') c = conn.cursor() c.execute("SELECT * FROM review_db WHERE date"\ " BETWEEN '2017-01-01 00:00:00' AND DATETIME('now')") results = c.fetchall() [ 290 ]

Chapter 9 >>> conn.close() >>> print(results) [('I love this movie', 1, '2019-06-15 17:53:46'), ('I disliked this movie', 0, '2019-06-15 17:53:46')]

Alternatively, we could also use the free DB browser for SQLite app (available at https://sqlitebrowser.org/dl/), which offers a nice graphical user interface for working with SQLite databases, as shown in the following figure:

Developing a web application with Flask

Having prepared the code for classifying movie reviews in the previous subsection, let's discuss the basics of the Flask web framework to develop our web application. Since Armin Ronacher's initial release of Flask in 2010, the framework has gained huge popularity, and examples of popular applications that use Flask include LinkedIn and Pinterest. Since Flask is written in Python, it provides us Python programmers with a convenient interface for embedding existing Python code, such as our movie classifier. The Flask microframework Flask is also known as a microframework, which means that its core is kept lean and simple but it can be easily extended with other libraries. Although the learning curve of the lightweight Flask API is not nearly as steep as those of other popular Python web frameworks, such as Django, you are encouraged to take a look at the official Flask documentation at https://flask.palletsprojects.com/en/1.0.x/ to learn more about its functionality.

[ 291 ]

Embedding a Machine Learning Model into a Web Application

If the Flask library is not already installed in your current Python environment, you can simply install it via conda or pip from your terminal (at the time of writing, the latest stable release was version 1.0.2): conda install flask # or: pip install flask

Our first Flask web application

In this subsection, we will develop a very simple web application to become more familiar with the Flask API before we implement our movie classifier. This first application that we are going to build consists of a simple web page with a form field that lets us enter a name. After submitting the name to the web application, it will render it on a new page. While this is a very simple example of a web application, it helps with building an understanding of how to store and pass variables and values between the different parts of our code within the Flask framework. First, we create a directory tree: 1st_flask_app_1/ app.py templates/ first_app.html

The app.py file will contain the main code that will be executed by the Python interpreter to run the Flask web application. The templates directory is the directory in which Flask will look for static HTML files for rendering in the web browser. Let's now take a look at the contents of app.py: from flask import Flask, render_template app = Flask(__name__) @app.route('/') def index(): return render_template('first_app.html') if __name__ == '__main__': app.run()

After looking at the previous code example, let's discuss the individual pieces step by step: 1. We ran our application as a single module; thus, we initialized a new Flask instance with the argument __name__ to let Flask know that it can find the HTML template folder (templates) in the same directory where it is located. [ 292 ]

Chapter 9

2. Next, we used the route decorator (@app.route('/')) to specify the URL that should trigger the execution of the index function. 3. Here, our index function simply rendered the first_app.html HTML file, which is located in the templates folder. 4. Lastly, we used the run function to run the application on the server only when this script was directly executed by the Python interpreter, which we ensured using the if statement with __name__ == '__main__'. Now, let's take a look at the contents of the first_app.html file:

First app

Hi, this is my first Flask web app!

HTML basics If you are not familiar with the HTML syntax yet, visit https://developer.mozilla.org/en-US/docs/Web/HTML

for useful tutorials on learning the basics of HTML.

Here, we have simply filled an empty HTML template file with a element (a block-level element) that contains this sentence: Hi, this is my first Flask web app!. Conveniently, Flask allows us to run our applications locally, which is useful for developing and testing web applications before we deploy them on a public web server. Now, let's start our web application by executing the command from the terminal inside the 1st_flask_app_1 directory: python3 app.py

We should see a line such as the following displayed in the terminal: * Running on http://127.0.0.1:5000/

This line contains the address of our local server. We can enter this address in our web browser to see the web application in action.

[ 293 ]

Embedding a Machine Learning Model into a Web Application

If everything has executed correctly, we should see a simple website with the content Hi, this is my first Flask web app! as shown in the following figure:

Form validation and rendering

In this subsection, we will extend our simple Flask web application with HTML form elements to learn how to collect data from a user using the WTForms library (https://wtforms.readthedocs.org/en/latest/), which can be installed via conda or pip: conda install wtforms # or pip install wtforms

This web application will prompt a user to type in his or her name into a text field, as shown in the following screenshot:

After the submission button (Say Hello) has been clicked and the form has been validated, a new HTML page will be rendered to display the user's name:

[ 294 ]

Chapter 9

Setting up the directory structure

The new directory structure that we need to set up for this application looks like this: 1st_flask_app_2/ app.py static/ style.css templates/ _formhelpers.html first_app.html hello.html

The following are the contents of our modified app.py file: from flask import Flask, render_template, request from wtforms import Form, TextAreaField, validators app = Flask(__name__) class HelloForm(Form): sayhello = TextAreaField('',[validators.DataRequired()]) @app.route('/') def index(): form = HelloForm(request.form) return render_template('first_app.html', form=form) @app.route('/hello', methods=['POST']) def hello(): form = HelloForm(request.form) if request.method == 'POST' and form.validate(): name = request.form['sayhello'] return render_template('hello.html', name=name) return render_template('first_app.html', form=form) if __name__ == '__main__': app.run(debug=True)

Let's discuss what the previous code does step by step: 1. Using wtforms, we extended the index function with a text field that we will embed in our start page using the TextAreaField class, which automatically checks whether a user has provided valid input text or not. 2. Furthermore, we defined a new function, hello, which will render an HTML page, hello.html, after validating the HTML form. [ 295 ]

Embedding a Machine Learning Model into a Web Application

3. Here, we used the POST method to transport the form data to the server in the message body. Finally, by setting the debug=True argument inside the app.run method, we further activated Flask's debugger. This is a useful feature for developing new web applications.

Implementing a macro using the Jinja2 templating engine

Now, we will implement a generic macro in the _formhelpers.html file via the Jinja2 templating engine, which we will later import in our first_app.html file to render the text field: {% macro render_field(field) %} {{ field.label }} {{ field(**kwargs)|safe }} {% if field.errors %}

- {% for error in field.errors %}
- {{ error }} {% endfor %}

{% endmacro %}

An in-depth discussion about the Jinja2 templating language is beyond the scope of this book. However, you can find comprehensive documentation on the Jinja2 syntax at http://jinja.pocoo.org.

Adding style via CSS

Next, we will set up a simple Cascading Style Sheets (CSS) file, style.css, to demonstrate how the look and feel of HTML documents can be modified. We have to save the following CSS file, which will simply double the font size of our HTML body elements, in a subdirectory called static, which is the default directory where Flask looks for static files such as CSS. The file content is as follows: body { font-size: 2em; }

[ 296 ]

Chapter 9

The following are the contents of the modified first_app.html file that will now render a text form where a user can enter a name:

First app

{% from "_formhelpers.html" import render_field %} What's your name?

{{ render_field(form.sayhello) }}

In the header section of first_app.html, we loaded the CSS file. It should now alter the size of all text elements in the HTML body. In the HTML body section, we imported the form macro from _formhelpers.html, and we rendered the sayhello form that we specified in the app.py file. Furthermore, we added a button to the same form element so that a user can submit the text field entry. The changes between the original and modified first_app.html file are illustrated in the following figure:

[ 297 ]

Embedding a Machine Learning Model into a Web Application

Creating the result page

Lastly, we will create a hello.html file that will be rendered via the render_ template('hello.html', name=name) line return inside the hello function, which we defined in the app.py script to display the text that a user submitted via the text field. The file content is as follows:

First app

Hello {{ name }}

Since we have covered a lot of ground in the previous section, the following figure provides an overview of the files we have created:

[ 298 ]

Chapter 9

Please note that you do not need to copy any code from the previous figure, since all file contents are present in the previous sections. For your convenience, copies of all files can also be found online at https://github.com/rasbt/pythonmachine-learning-book-3rd-edition/tree/master/ch09/1st_flask_app_2. Having set up our modified Flask web application, we can run it locally by executing the following command from the application's main directory: python3 app.py

Then, to see the resulting web page, enter the IP address shown in your terminal, which is usually http://127.0.0.1:5000/, into your web browser to view the rendered web app as summarized in the following figure:

Flask documentation and examples If you are new to web development, some of those concepts may seem very complicated at first sight. In that case, simply set up the preceding files in a directory on your hard drive and examine them closely. You will see that the Flask web framework is relatively straightforward and much simpler than it might initially appear! Also, for more help, don't forget to consult the excellent Flask documentation and examples at http://flask.pocoo.org/docs/1.0/.

[ 299 ]

Embedding a Machine Learning Model into a Web Application

Turning the movie review classifier into a web application

Now that we are somewhat familiar with the basics of Flask web development, let's advance to the next step and implement our movie classifier into a web application. In this section, we will develop a web application that will first prompt a user to enter a movie review, as shown in the following screenshot:

After the review has been submitted, the user will see a new page that shows the predicted class label and the probability of the prediction. Furthermore, the user will be able to provide feedback about this prediction by clicking on the Correct or Incorrect button, as shown in the following screenshot:

[ 300 ]

Chapter 9

If a user clicked on either the Correct or Incorrect button, our classification model will be updated with respect to the user's feedback. Furthermore, we will also store the movie review text provided by the user, as well as the suggested class label, which can be inferred from the button click, in an SQLite database for future reference. (Alternatively, a user could skip the update step and click the Submit another review button to submit another review.) The third page that the user will see after clicking on one of the feedback buttons is a simple thank you screen with a Submit another review button that redirects the user back to the start page. This is shown in the following screenshot:

Live demo Before we take a closer look at the code implementation of this web application, take a look at this live demo at http://raschkas.pythonanywhere.com to get a better understanding of what we are trying to accomplish in this section.

Files and folders – looking at the directory tree

To start with the big picture, let's take a look at the directory tree that we are going to create for this movie classification application, which is shown here:

[ 301 ]

Embedding a Machine Learning Model into a Web Application

Earlier in this chapter, we created the vectorizer.py file, the SQLite database, reviews.sqlite, and the pkl_objects subdirectory with the pickled Python objects. The app.py file in the main directory is the Python script that contains our Flask code, and we will use the review.sqlite database file (which we created earlier in this chapter) to store the movie reviews that are being submitted to our web application. The templates subdirectory contains the HTML templates that will be rendered by Flask and displayed in the browser, and the static subdirectory will contain a simple CSS file to adjust the look of the rendered HTML code. Getting the movieclassifier code files A separate directory containing the movie review classifier application with the code discussed in this section is provided with the code examples for this book, which you can either obtain directly from Packt or download from GitHub at https://github.com/rasbt/python-machine-learningbook-3rd-edition/. The code in this section can be found in the .../code/ch09/movieclassifier subdirectory.

Implementing the main application as app.py

Since the app.py file is rather long, we will conquer it in two steps. The first section of app.py imports the Python modules and objects that we are going to need, as well as the code to unpickle and set up our classification model: [ 302 ]

Chapter 9 from flask import Flask, render_template, request from wtforms import Form, TextAreaField, validators import pickle import sqlite3 import os import numpy as np # import HashingVectorizer from local dir from vectorizer import vect app = Flask(__name__) ######## Preparing the Classifier cur_dir = os.path.dirname(__file__) clf = pickle.load(open(os.path.join(cur_dir, 'pkl_objects', 'classifier.pkl'), 'rb')) db = os.path.join(cur_dir, 'reviews.sqlite') def classify(document): label = {0: 'negative', 1: 'positive'} X = vect.transform([document]) y = clf.predict(X)[0] proba = np.max(clf.predict_proba(X)) return label[y], proba def train(document, y): X = vect.transform([document]) clf.partial_fit(X, [y]) def sqlite_entry(path, document, y): conn = sqlite3.connect(path) c = conn.cursor() c.execute("INSERT INTO review_db (review, sentiment, date)"\ " VALUES (?, ?, DATETIME('now'))", (document, y)) conn.commit() conn.close()

This first part of the app.py script should look very familiar by now. We simply imported HashingVectorizer and unpickled the logistic regression classifier. Next, we defined a classify function to return the predicted class label, as well as the corresponding probability prediction of a given text document. The train function can be used to update the classifier, given that a document and a class label are provided. [ 303 ]

Embedding a Machine Learning Model into a Web Application

Using the sqlite_entry function, we can store a submitted movie review in our SQLite database along with its class label and timestamp for our personal records. Note that the clf object will be reset to its original, pickled state if we restart the web application. At the end of this chapter, you will learn how to use the data that we collect in the SQLite database to update the classifier permanently. The concepts in the second part of the app.py script should also look quite familiar: ######## Flask class ReviewForm(Form): moviereview = TextAreaField('', [validators.DataRequired(), validators.length(min=15)]) @app.route('/') def index(): form = ReviewForm(request.form) return render_template('reviewform.html', form=form) @app.route('/results', methods=['POST']) def results(): form = ReviewForm(request.form) if request.method == 'POST' and form.validate(): review = request.form['moviereview'] y, proba = classify(review) return render_template('results.html', content=review, prediction=y, probability=round(proba*100, 2)) return render_template('reviewform.html', form=form) @app.route('/thanks', methods=['POST']) def feedback(): feedback = request.form['feedback_button'] review = request.form['review'] prediction = request.form['prediction'] inv_label = {'negative': 0, 'positive': 1} y = inv_label[prediction] if feedback == 'Incorrect': y = int(not(y)) train(review, y) sqlite_entry(db, review, y) return render_template('thanks.html')

[ 304 ]

Chapter 9 if __name__ == '__main__': app.run(debug=True)

We defined a ReviewForm class that instantiates a TextAreaField, which will be rendered in the reviewform.html template file (the landing page of our web application). This, in turn, will be rendered by the index function. With the validators.length(min=15) parameter, we require the user to enter a review that contains at least 15 characters. Inside the results function, we fetch the contents of the submitted web form and pass it on to our classifier to predict the sentiment of the movie classifier, which will then be displayed in the rendered results. html template. The feedback function, which we implemented in app.py in the previous subsection, may look a little bit complicated at first glance. It essentially fetches the predicted class label from the results.html template if a user clicked on the Correct or Incorrect feedback button, and it transforms the predicted sentiment back into an integer class label that will be used to update the classifier via the train function, which we implemented in the first section of the app.py script. Also, a new entry to the SQLite database will be made via the sqlite_entry function if feedback was provided, and eventually, the thanks.html template will be rendered to thank the user for the feedback.

Setting up the review form

Next, let's take a look at the reviewform.html template, which constitutes the starting page of our application:

Movie Classification

Please enter your movie review: {% from "_formhelpers.html" import render_field %}

{{ render_field(form.moviereview, cols='30', rows='10') }}

[ 305 ]

Embedding a Machine Learning Model into a Web Application

Here, we simply imported the same _formhelpers.html template that we defined in the Form validation and rendering section earlier in this chapter. The render_field function of this macro is used to render a TextAreaField where a user can provide a movie review and submit it via the Submit review button displayed at the bottom of the page. This TextAreaField is 30 columns wide and 10 rows tall, and will look like this:

Creating a results page template

Our next template, results.html, looks a little bit more interesting:

Movie Classification

Your movie review: [ 306 ]

Chapter 9 {{ content }} Prediction: This movie review is

First, we inserted the submitted review, as well as the results of the prediction, in the corresponding fields {{ content }}, {{ prediction }}, and {{ probability }}. You may notice that we used the {{ content }} and {{ prediction }} placeholder variables (in this context, also known as hidden fields) a second time in the form that contains the Correct and Incorrect buttons. This is a workaround to POST those values back to the server to update the classifier and store the review in case the user clicks on one of those two buttons. Furthermore, we imported a CSS file (style.css) at the beginning of the results.html file. The setup of this file is quite simple: it limits the width of the

contents of this web application to 600 pixels and moves the Incorrect and Correct buttons labeled with the div id button down by 20 pixels: body{ width:600px; } .button{ padding-top: 20px; }

[ 307 ]

Embedding a Machine Learning Model into a Web Application

This CSS file is merely a placeholder, so please feel free to modify it to adjust the look and feel of the web application to your liking. The last HTML file we will implement for our web application is the thanks.html template. As the name suggests, it simply provides a nice thank you message to the user after providing feedback via the Correct or Incorrect button. Furthermore, we will put a Submit another review button at the bottom of this page, which will redirect the user to the starting page. The contents of the thanks.html file are as follows:

Movie Classification

Thank you for your feedback!

Now, it would be a good idea to start the web application locally from our command-line terminal via the following command before we advance to the next subsection and deploy it on a public web server: python3 app.py

After we have finished testing our application, we also shouldn't forget to remove the debug=True argument in the app.run() command of our app.py script (or set debug=False ) as illustrated in the following figure:

[ 308 ]

Chapter 9

Deploying the web application to a public server After we have tested the web application locally, we are now ready to deploy our web application onto a public web server. For this tutorial, we will be using the PythonAnywhere web hosting service, which specializes in the hosting of Python web applications and makes it extremely simple and hassle-free. Furthermore, PythonAnywhere offers a beginner account option that lets us run a single web application free of charge.

Creating a PythonAnywhere account

To create a new PythonAnywhere account, we visit the website at https://www.pythonanywhere.com/ and click on the Pricing & signup link that is located in the top-right corner. Next, we click on the Create a Beginner account button where we need to provide a username, password, and valid email address. After we have read and agreed to the terms and conditions, we should have a new account. Unfortunately, the free beginner account doesn't allow us to access the remote server via the Secure Socket Shell (SSH) protocol from our terminal. Thus, we need to use the PythonAnywhere web interface to manage our web application. But before we can upload our local application files to the server, we need to create a new web application for our PythonAnywhere account. After we click on the Dashboard button in the top-right corner, we have access to the control panel shown at the top of the page. Next, we click on the Web tab that is now visible at the top of the page. We proceed by clicking on the +Add a new web app button on the left, which lets us create a new Python 3.7 Flask web application that we name movieclassifier.

[ 309 ]

Embedding a Machine Learning Model into a Web Application

Uploading the movie classifier application

After creating a new application for our PythonAnywhere account, we head over to the Files tab to upload the files from our local movieclassifier directory using the PythonAnywhere web interface. After uploading the web application files that we created locally on our computer, we should have a movieclassifier directory in our PythonAnywhere account. It will contain the same directories and files as our local movieclassifier directory, as shown in the following screenshot:

Then, we head over to the Web tab one more time and click on the Reload .pythonanywhere.com button to propagate the changes and refresh our web application. Finally, our web application should now be up and running and publicly available via .pythonanywhere.com.

[ 310 ]

Chapter 9

Troubleshooting Unfortunately, web servers can be quite sensitive to the tiniest problems in our web application. If you are experiencing problems with running the web application on PythonAnywhere and are receiving error messages in your browser, you can check the server and error logs, which can be accessed from the Web tab in your PythonAnywhere account, to better diagnose the problem.

Updating the movie classifier

While our predictive model is updated on the fly whenever a user provides feedback about the classification, the updates to the clf object will be reset if the web server crashes or restarts. If we reload the web application, the clf object will be reinitialized from the classifier.pkl pickle file. One option to apply the updates permanently would be to pickle the clf object once again after each update. However, this would become computationally very inefficient with a growing number of users and could corrupt the pickle file if users provide feedback simultaneously. An alternative solution is to update the predictive model from the feedback data that is being collected in the SQLite database. One option would be to download the SQLite database from the PythonAnywhere server, update the clf object locally on our computer, and upload the new pickle file to PythonAnywhere. To update the classifier locally on our computer, we create an update.py script file in the movieclassifier directory with the following contents: import import import import

pickle sqlite3 numpy as np os

# import HashingVectorizer from local dir from vectorizer import vect def update_model(db_path, model, batch_size=10000): conn = sqlite3.connect(db_path) c = conn.cursor() c.execute('SELECT * from review_db') results = c.fetchmany(batch_size)

[ 311 ]

Embedding a Machine Learning Model into a Web Application while results: data = np.array(results) X = data[:, 0] y = data[:, 1].astype(int) classes = np.array([0, 1]) X_train = vect.transform(X) model.partial_fit(X_train, y, classes=classes) results = c.fetchmany(batch_size) conn.close() return model cur_dir = os.path.dirname(__file__) clf = pickle.load(open(os.path.join(cur_dir, 'pkl_objects', 'classifier.pkl'), 'rb')) db = os.path.join(cur_dir, 'reviews.sqlite') clf = update_model(db_path=db, model=clf, batch_size=10000) # Uncomment the following lines if you are sure that # you want to update your classifier.pkl file # permanently. # pickle.dump(clf, open(os.path.join(cur_dir, # 'pkl_objects', 'classifier.pkl'), 'wb'), # protocol=4)

Getting the movieclassifier code files with the update functionality A separate directory containing the movie review classifier application with the update functionality discussed in this chapter comes with the code examples for this book, which you can either obtain directly from Packt or download from GitHub at https://github.com/rasbt/python-machinelearning-book-3rd-edition. The code in this section is located in the .../code/ch09/movieclassifier_with_ update subdirectory.

[ 312 ]

Chapter 9

The update_model function will fetch entries from the SQLite database in batches of 10,000 entries at a time, unless the database contains fewer entries. Alternatively, we could also fetch one entry at a time by using fetchone instead of fetchmany, which would be computationally very inefficient. However, keep in mind that using the alternative fetchall method could be a problem if we are working with large datasets that exceed the computer or server's memory capacity. Now that we have created the update.py script, we could also upload it to the movieclassifier directory on PythonAnywhere and import the update_model function in the main application script, app.py, to update the classifier from the SQLite database every time we restart the web application. In order to do so, we just need to add a line of code to import the update_model function from the update.py script at the top of app.py: # import update function from local dir from update import update_model

We then need to call the update_model function in the main application body: ... if __name__ == '__main__': clf = update_model(db_path=db, model=clf, batch_size=10000) ...

As discussed, the modification in the previous code snippet will update the pickle file on PythonAnywhere. However, in practice, we do not often have to restart our web application, and it would make sense to validate the user feedback in the SQLite database prior to the update to make sure that the feedback is valuable information for the classifier. Creating backups In a real-world application, you may also want to back up the classifier.pkl pickle file periodically to have a safeguard against file corruption, for instance, by creating a timestamped version prior to each update. In order to create backups of the pickled classifier, you can import the following: from shutil import copyfile import time

[ 313 ]

Embedding a Machine Learning Model into a Web Application

Then, above the code that updates the pickled classifier, which is as follows, pickle.dump( clf, open( os.path.join( cur_dir, 'pkl_objects', 'classifier.pkl'), 'wb'), protocol=4)

insert the following lines of code: timestr = time.strftime("%Y%m%d-%H%M%S") orig_path = os.path.join( cur_dir, 'pkl_objects', 'classifier.pkl') backup_path = os.path.join( cur_dir, 'pkl_objects', 'classifier_%s.pkl' % timestr) copyfile(orig_path, backup_path)

As a consequence, backup files of the pickled classifier will be created following the format YearMonthDay-HourMinuteSecond, for example, classifier_20190822-092148.pkl.

Summary

In this chapter, you learned about many useful and practical topics that will extend your knowledge of machine learning theory. You learned how to serialize a model after training and how to load it for later use cases. Furthermore, we created an SQLite database for efficient data storage and created a web application that lets us make our movie classifier available to the outside world. So far, in this book, we have covered many machine learning concepts, best practices, and supervised models for classification. In the next chapter, we will take a look at another subcategory of supervised learning, regression analysis, which lets us predict outcome variables on a continuous scale, in contrast to the categorical class labels of the classification models that we have been working with so far.

[ 314 ]

Predicting Continuous Target Variables with Regression Analysis Throughout the previous chapters, you learned a lot about the main concepts behind supervised learning and trained many different models for classification tasks to predict group memberships or categorical variables. In this chapter, we will dive into another subcategory of supervised learning: regression analysis. Regression models are used to predict target variables on a continuous scale, which makes them attractive for addressing many questions in science. They also have applications in industry, such as understanding relationships between variables, evaluating trends, or making forecasts. One example is predicting the sales of a company in future months. In this chapter, we will discuss the main concepts of regression models and cover the following topics: • • • • •

Exploring and visualizing datasets Looking at different approaches to implement linear regression models Training regression models that are robust to outliers Evaluating regression models and diagnosing common problems Fitting regression models to nonlinear data

Introducing linear regression

The goal of linear regression is to model the relationship between one or multiple features and a continuous target variable. In contrast to classification—a different subcategory of supervised learning—regression analysis aims to predict outputs on a continuous scale rather than categorical class labels. [ 315 ]

Predicting Continuous Target Variables with Regression Analysis

In the following subsections, you will be introduced to the most basic type of linear regression, simple linear regression, and understand how to relate it to the more general, multivariate case (linear regression with multiple features).

Simple linear regression

The goal of simple (univariate) linear regression is to model the relationship between a single feature (explanatory variable, x) and a continuous-valued target (response variable, y). The equation of a linear model with one explanatory variable is defined as follows:

𝑦𝑦 = 𝑤𝑤0 + 𝑤𝑤1 𝑥𝑥

Here, the weight, 𝑤𝑤0 , represents the y axis intercept and 𝑤𝑤1 is the weight coefficient of the explanatory variable. Our goal is to learn the weights of the linear equation to describe the relationship between the explanatory variable and the target variable, which can then be used to predict the responses of new explanatory variables that were not part of the training dataset. Based on the linear equation that we defined previously, linear regression can be understood as finding the best-fitting straight line through the training examples, as shown in the following figure:

[ 316 ]

Chapter 10

This best-fitting line is also called the regression line, and the vertical lines from the regression line to the training examples are the so-called offsets or residuals—the errors of our prediction.

Multiple linear regression

The previous section introduced simple linear regression, a special case of linear regression with one explanatory variable. Of course, we can also generalize the linear regression model to multiple explanatory variables; this process is called multiple linear regression: 𝑛𝑛

𝑦𝑦 = 𝑤𝑤0 𝑥𝑥0 + 𝑤𝑤1 𝑥𝑥1 + … + 𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚 = ∑ 𝑤𝑤𝑖𝑖 𝑥𝑥𝑖𝑖 = 𝑤𝑤 𝑇𝑇 𝑥𝑥 Here, 𝑤𝑤0 is the y axis intercept with 𝑥𝑥0 = 1.

𝑖𝑖=0

The following figure shows how the two-dimensional, fitted hyperplane of a multiple linear regression model with two features could look:

[ 317 ]

Predicting Continuous Target Variables with Regression Analysis

As you can see, visualizations of multiple linear regression hyperplanes in a three-dimensional scatterplot are already challenging to interpret when looking at static figures. Since we have no good means of visualizing hyperplanes with two dimensions in a scatterplot (multiple linear regression models fit to datasets with three or more features), the examples and visualizations in this chapter will mainly focus on the univariate case, using simple linear regression. However, simple and multiple linear regression are based on the same concepts and the same evaluation techniques; the code implementations that we will discuss in this chapter are also compatible with both types of regression model.

Exploring the Housing dataset

Before we implement the first linear regression model, we will discuss a new dataset, the Housing dataset, which contains information about houses in the suburbs of Boston collected by D. Harrison and D.L. Rubinfeld in 1978. The Housing dataset has been made freely available and is included in the code bundle of this book. The dataset has recently been removed from the UCI Machine Learning Repository but is available online at https://raw.githubusercontent.com/rasbt/pythonmachine-learning-book-3rd-edition/master/ch10/housing.data.txt or scikit-learn (https://github.com/scikit-learn/scikit-learn/blob/master/ sklearn/datasets/data/boston_house_prices.csv). As with each new dataset, it is always helpful to explore the data through a simple visualization, to get a better feeling of what we are working with.

Loading the Housing dataset into a data frame

In this section, we will load the Housing dataset using the pandas read_csv function, which is fast and versatile and a recommended tool for working with tabular data stored in a plaintext format.

The features of the 506 examples in the Housing dataset have been taken from the original source that was previously shared on https://archive.ics.uci.edu/ml/ datasets/Housing and summarized here: • CRIM: Per capita crime rate by town • ZN: Proportion of residential land zoned for lots over 25,000 sq. ft. • INDUS: Proportion of non-retail business acres per town • CHAS: Charles River dummy variable (= 1 if tract bounds river and 0 otherwise) • NOX: Nitric oxide concentration (parts per 10 million)

[ 318 ]

Chapter 10

• RM: Average number of rooms per dwelling • AGE: Proportion of owner-occupied units built prior to 1940 • DIS: Weighted distances to five Boston employment centers • RAD: Index of accessibility to radial highways • TAX: Full-value property tax rate per $10,000 • PTRATIO: Pupil-teacher ratio by town • B: 1000(Bk – 0.63)2, where Bk is the proportion of [people of African American descent] by town • LSTAT: Percentage of lower status of the population • MEDV: Median value of owner-occupied homes in $1000s For the rest of this chapter, we will regard the house prices (MEDV) as our target variable—the variable that we want to predict using one or more of the 13 explanatory variables. Before we explore this dataset further, let's load it into a pandas DataFrame: >>> >>> ... ... ... ... >>> ... ... >>>

import pandas as pd df = pd.read_csv('https://raw.githubusercontent.com/rasbt/' 'python-machine-learning-book-3rd-edition' '/master/ch10/housing.data.txt', header=None, sep='\s+') df.columns = ['CRIM', 'ZN', 'INDUS', 'CHAS', 'NOX', 'RM', 'AGE', 'DIS', 'RAD', 'TAX', 'PTRATIO', 'B', 'LSTAT', 'MEDV'] df.head()

To confirm that the dataset was loaded successfully, we can display the first five lines of the dataset, as shown in the following figure:

[ 319 ]

Predicting Continuous Target Variables with Regression Analysis

Obtaining the Housing dataset You can find a copy of the Housing dataset (and all other datasets used in this book) in the code bundle of this book, which you can use if you are working offline or the web link https://raw.githubusercontent.com/rasbt/pythonmachine-learning-book-3rd-edition/master/code/ ch10/housing.data.txt is temporarily unavailable. For

instance, to load the Housing dataset from a local directory, you can replace these lines: df = pd.read_csv( 'https://raw.githubusercontent.com/rasbt/' 'python-machine-learning-book-3rd-edition/' 'master/ch10/housing.data.txt', header=None, sep='\s+')

in the preceding code example with this line: df = pd.read_csv('./housing.data.txt', sep='\s+')

Visualizing the important characteristics of a dataset

Exploratory data analysis (EDA) is an important and recommended first step prior to the training of a machine learning model. In the rest of this section, we will use some simple yet useful techniques from the graphical EDA toolbox that may help us to visually detect the presence of outliers, the distribution of the data, and the relationships between features. First, we will create a scatterplot matrix that allows us to visualize the pair-wise correlations between the different features in this dataset in one place. To plot the scatterplot matrix, we will use the scatterplotmatrix function from the MLxtend library (http://rasbt.github.io/mlxtend/), which is a Python library that contains various convenience functions for machine learning and data science applications in Python. You can install the mlxtend package via conda install mlxtend or pip install mlxtend. After the installation is complete, you can import the package and create the scatterplot matrix as follows: >>> import matplotlib.pyplot as plt >>> from mlxtend.plotting import scatterplotmatrix >>> cols = ['LSTAT', 'INDUS', 'NOX', 'RM', 'MEDV'] [ 320 ]

Chapter 10 >>> scatterplotmatrix(df[cols].values, figsize=(10, 8), ... names=cols, alpha=0.5) >>> plt.tight_layout() >>> plt.show()

As you can see in the following figure, the scatterplot matrix provides us with a useful graphical summary of the relationships in a dataset:

Due to space constraints and in the interest of readability, we have only plotted five columns from the dataset: LSTAT, INDUS, NOX, RM, and MEDV. However, you are encouraged to create a scatterplot matrix of the whole DataFrame to explore the dataset further by choosing different column names in the previous scatterplotmatrix function call, or including all variables in the scatterplot matrix by omitting the column selector.

[ 321 ]

Predicting Continuous Target Variables with Regression Analysis

Using this scatterplot matrix, we can now quickly eyeball how the data is distributed and whether it contains outliers. For example, we can see that there is a linear relationship between RM and house prices, MEDV (the fifth column of the fourth row). Furthermore, we can see in the histogram—the lower-right subplot in the scatterplot matrix—that the MEDV variable seems to be normally distributed but contains several outliers. Normality assumption of linear regression Note that in contrast to common belief, training a linear regression model does not require that the explanatory or target variables are normally distributed. The normality assumption is only a requirement for certain statistics and hypothesis tests that are beyond the scope of this book (for more information on this topic, please refer to Introduction to Linear Regression Analysis, Montgomery, Douglas C. Montgomery, Elizabeth A. Peck, and G. Geoffrey Vining, Wiley, 2012, pages: 318-319).

Looking at relationships using a correlation matrix

In the previous section, we visualized the data distributions of the Housing dataset variables in the form of histograms and scatterplots. Next, we will create a correlation matrix to quantify and summarize linear relationships between variables. A correlation matrix is closely related to the covariance matrix that we covered in the section Unsupervised dimensionality reduction via principal component analysis in Chapter 5, Compressing Data via Dimensionality Reduction. We can interpret the correlation matrix as being a rescaled version of the covariance matrix. In fact, the correlation matrix is identical to a covariance matrix computed from standardized features. The correlation matrix is a square matrix that contains the Pearson product-moment correlation coefficient (often abbreviated as Pearson's r), which measures the linear dependence between pairs of features. The correlation coefficients are in the range –1 to 1. Two features have a perfect positive correlation if r = 1, no correlation if r = 0, and a perfect negative correlation if r = –1. As mentioned previously, Pearson's correlation coefficient can simply be calculated as the covariance between two features, x and y (numerator), divided by the product of their standard deviations (denominator): 𝑟𝑟 =

∑𝑛𝑛𝑖𝑖=1[(𝑥𝑥 (𝑖𝑖) − 𝜇𝜇𝑥𝑥 )(𝑦𝑦 (𝑖𝑖) − 𝜇𝜇𝑦𝑦 )]

√∑𝑛𝑛𝑖𝑖=1(𝑥𝑥 (𝑖𝑖) − 𝜇𝜇𝑥𝑥 )2 √∑𝑛𝑛𝑖𝑖=1(𝑦𝑦 (𝑖𝑖) − 𝜇𝜇𝑦𝑦 ) [ 322 ]

2

=

𝜎𝜎𝑥𝑥𝑥𝑥 𝜎𝜎𝑥𝑥 𝜎𝜎𝑦𝑦

Chapter 10

Here, 𝜇𝜇 denotes the mean of the corresponding feature, 𝜎𝜎𝑥𝑥𝑥𝑥 is the covariance between the features x and y, and 𝜎𝜎𝑥𝑥 and 𝜎𝜎𝑦𝑦 are the features' standard deviations. Covariance versus correlation for standardized features

We can show that the covariance between a pair of standardized features is, in fact, equal to their linear correlation coefficient. To show this, let's first standardize the features x and y to obtain their z-scores, which we will denote as 𝑥𝑥 ′ and 𝑦𝑦 ′, respectively:

𝑥𝑥 ′ =

𝑥𝑥 − 𝜇𝜇𝑥𝑥 ′ 𝑦𝑦 − 𝜇𝜇𝑦𝑦 , 𝑦𝑦 = 𝜎𝜎𝑥𝑥 𝜎𝜎𝑦𝑦

Remember that we compute the (population) covariance between two features as follows: 𝜎𝜎𝑥𝑥𝑥𝑥

𝑛𝑛

1 = ∑(𝑥𝑥 (𝑖𝑖) − 𝜇𝜇𝑥𝑥 ) (𝑦𝑦 (𝑖𝑖) − 𝜇𝜇𝑦𝑦 ) 𝑛𝑛 𝑖𝑖

Since standardization centers a feature variable at mean zero, we can now calculate the covariance between the scaled features as follows:

𝜎𝜎

′

𝑥𝑥𝑥𝑥

𝑛𝑛

1 = ∑(𝑥𝑥′(𝑖𝑖) − 0) (𝑦𝑦′(𝑖𝑖) − 0) 𝑛𝑛 𝑖𝑖

Through resubstitution, we then get the following result:

𝜎𝜎

′

𝜎𝜎

′

𝑥𝑥𝑥𝑥

𝑥𝑥𝑥𝑥

𝑛𝑛

1 𝑥𝑥 − 𝜇𝜇𝑥𝑥 𝑦𝑦 − 𝜇𝜇𝑦𝑦 = ∑( )( ) 𝜎𝜎𝑥𝑥 𝜎𝜎𝑦𝑦 𝑛𝑛 𝑖𝑖

𝑛𝑛

1 = ∑(𝑥𝑥 (𝑖𝑖) − 𝜇𝜇𝑥𝑥 ) (𝑦𝑦 (𝑖𝑖) − 𝜇𝜇𝑦𝑦 ) 𝑛𝑛 ∙ 𝜎𝜎𝑥𝑥 𝜎𝜎𝑦𝑦 𝑖𝑖

Finally, we can simplify this equation as follows:

𝜎𝜎′𝑥𝑥𝑥𝑥 =

[ 323 ]

𝜎𝜎𝑥𝑥𝑥𝑥 𝜎𝜎𝑥𝑥 𝜎𝜎𝑦𝑦

Predicting Continuous Target Variables with Regression Analysis

In the following code example, we will use NumPy's corrcoef function on the five feature columns that we previously visualized in the scatterplot matrix, and we will use MLxtend's heatmap function to plot the correlation matrix array as a heat map: >>> >>> >>> >>> ... ... >>>

from mlxtend.plotting import heatmap import numpy as np cm = np.corrcoef(df[cols].values.T) hm = heatmap(cm, row_names=cols, column_names=cols) plt.show()

As you can see in the resulting figure, the correlation matrix provides us with another useful summary graphic that can help us to select features based on their respective linear correlations:

[ 324 ]

Chapter 10

To fit a linear regression model, we are interested in those features that have a high correlation with our target variable, MEDV. Looking at the previous correlation matrix, we can see that our target variable, MEDV, shows the largest correlation with the LSTAT variable (-0.74); however, as you might remember from inspecting the scatterplot matrix, there is a clear nonlinear relationship between LSTAT and MEDV. On the other hand, the correlation between RM and MEDV is also relatively high (0.70). Given the linear relationship between these two variables that we observed in the scatterplot, RM seems to be a good choice for an exploratory variable to introduce the concepts of a simple linear regression model in the following section.

Implementing an ordinary least squares linear regression model

At the beginning of this chapter, it was mentioned that linear regression can be understood as obtaining the best-fitting straight line through the examples of our training data. However, we have neither defined the term best-fitting nor have we discussed the different techniques of fitting such a model. In the following subsections, we will fill in the missing pieces of this puzzle using the ordinary least squares (OLS) method (sometimes also called linear least squares) to estimate the parameters of the linear regression line that minimizes the sum of the squared vertical distances (residuals or errors) to the training examples.

Solving regression for regression parameters with gradient descent

Consider our implementation of the Adaptive Linear Neuron (Adaline) from Chapter 2, Training Simple Machine Learning Algorithms for Classification. You will remember that the artificial neuron uses a linear activation function. Also, we defined a cost function, J(w), which we minimized to learn the weights via optimization algorithms, such as gradient descent (GD) and stochastic gradient descent (SGD). This cost function in Adaline is the sum of squared errors (SSE), which is identical to the cost function that we use for OLS: 𝑛𝑛

1 2 𝐽𝐽(𝒘𝒘) = ∑(𝑦𝑦 (𝑖𝑖) − 𝑦𝑦̂ (𝑖𝑖) ) 2 𝑖𝑖=1

[ 325 ]

Predicting Continuous Target Variables with Regression Analysis

1 is just used for 2 convenience to derive the update rule of GD). Essentially, OLS regression can be understood as Adaline without the unit step function so that we obtain continuous target values instead of the class labels -1 and 1. To demonstrate this, let's take the GD implementation of Adaline from Chapter 2, Training Simple Machine Learning Algorithms for Classification, and remove the unit step function to implement our first linear regression model: Here, 𝑦𝑦̂ is the predicted value 𝑦𝑦̂ = 𝒘𝒘𝑇𝑇 𝒙𝒙 (note that the term

class LinearRegressionGD(object): def __init__(self, eta=0.001, n_iter=20): self.eta = eta self.n_iter = n_iter def fit(self, X, y): self.w_ = np.zeros(1 + X.shape[1]) self.cost_ = [] for i in range(self.n_iter): output = self.net_input(X) errors = (y - output) self.w_[1:] += self.eta * X.T.dot(errors) self.w_[0] += self.eta * errors.sum() cost = (errors**2).sum() / 2.0 self.cost_.append(cost) return self def net_input(self, X): return np.dot(X, self.w_[1:]) + self.w_[0] def predict(self, X): return self.net_input(X)

Weight updates with gradient descent If you need a refresher about how the weights are updated— taking a step in the opposite direction of the gradient—please revisit the Adaptive linear neurons and the convergence of learning section in Chapter 2, Training Simple Machine Learning Algorithms for Classification.

[ 326 ]

Chapter 10

To see our LinearRegressionGD regressor in action, let's use the RM (number of rooms) variable from the Housing dataset as the explanatory variable and train a model that can predict MEDV (house prices). Furthermore, we will standardize the variables for better convergence of the GD algorithm. The code is as follows: >>> >>> >>> >>> >>> >>> >>> >>> >>>

X = df[['RM']].values y = df['MEDV'].values from sklearn.preprocessing import StandardScaler sc_x = StandardScaler() sc_y = StandardScaler() X_std = sc_x.fit_transform(X) y_std = sc_y.fit_transform(y[:, np.newaxis]).flatten() lr = LinearRegressionGD() lr.fit(X_std, y_std)

Notice the workaround regarding y_std, using np.newaxis and flatten. Most transformers in scikit-learn expect data to be stored in two-dimensional arrays. In the previous code example, the use of np.newaxis in y[:, np.newaxis] added a new dimension to the array. Then, after StandardScaler returned the scaled variable, we converted it back to the original one-dimensional array representation using the flatten() method for our convenience. We discussed in Chapter 2, Training Simple Machine Learning Algorithms for Classification, that it is always a good idea to plot the cost as a function of the number of epochs (complete iterations) over the training dataset when we are using optimization algorithms, such as GD, to check that the algorithm converged to a cost minimum (here, a global cost minimum): >>> >>> >>> >>>

plt.plot(range(1, lr.n_iter+1), lr.cost_) plt.ylabel('SSE') plt.xlabel('Epoch') plt.show()

[ 327 ]

Predicting Continuous Target Variables with Regression Analysis

As you can see in the following plot, the GD algorithm converged after the fifth epoch:

Next, let's visualize how well the linear regression line fits the training data. To do so, we will define a simple helper function that will plot a scatterplot of the training examples and add the regression line: >>> def lin_regplot(X, y, model): ... plt.scatter(X, y, c='steelblue', edgecolor='white', s=70) ... plt.plot(X, model.predict(X), color='black', lw=2) ... return None

Now, we will use this lin_regplot function to plot the number of rooms against the house price: >>> >>> >>> >>>

lin_regplot(X_std, y_std, lr) plt.xlabel('Average number of rooms [RM] (standardized)') plt.ylabel('Price in $1000s [MEDV] (standardized)') plt.show()

As you can see in the following plot, the linear regression line reflects the general trend that house prices tend to increase with the number of rooms:

[ 328 ]

Chapter 10

Although this observation makes sense, the data also tells us that the number of rooms does not explain house prices very well in many cases. Later in this chapter, we will discuss how to quantify the performance of a regression model. Interestingly, we can also observe that several data points lined up at y = 3, which suggests that the prices may have been clipped. In certain applications, it may also be important to report the predicted outcome variables on their original scale. To scale the predicted price outcome back onto the Price in $1000s axis, we can simply apply the inverse_transform method of the StandardScaler: >>> num_rooms_std = sc_x.transform(np.array([[5.0]])) >>> price_std = lr.predict(num_rooms_std) >>> print("Price in $1000s: %.3f" % \ ... sc_y.inverse_transform(price_std)) Price in $1000s: 10.840

In this code example, we used the previously trained linear regression model to predict the price of a house with five rooms. According to our model, such a house will be worth $10,840. As a side note, it is also worth mentioning that we technically don't have to update the weights of the intercept if we are working with standardized variables, since the y axis intercept is always 0 in those cases. We can quickly confirm this by printing the weights: >>> print('Slope: %.3f' % lr.w_[1]) Slope: 0.695 >>> print('Intercept: %.3f' % lr.w_[0]) Intercept: -0.000 [ 329 ]

Predicting Continuous Target Variables with Regression Analysis

Estimating the coefficient of a regression model via scikit-learn

In the previous section, we implemented a working model for regression analysis; however, in a real-world application, we may be interested in more efficient implementations. For example, many of scikit-learn's estimators for regression make use of the least squares implementation in SciPy (scipy.linalg.lstsq), which in turn uses highly optimized code optimizations based on the Linear Algebra Package (LAPACK). The linear regression implementation in scikit-learn also works (better) with unstandardized variables, since it does not use (S)GD-based optimization, so we can skip the standardization step: >>> from sklearn.linear_model import LinearRegression >>> slr = LinearRegression() >>> slr.fit(X, y) >>> y_pred = slr.predict(X) >>> print('Slope: %.3f' % slr.coef_[0]) Slope: 9.102 >>> print('Intercept: %.3f' % slr.intercept_) Intercept: -34.671

As you can see from executing this code, scikit-learn's LinearRegression model, fitted with the unstandardized RM and MEDV variables, yielded different model coefficients, since the features have not been standardized. However, when we compare it to our GD implementation by plotting MEDV against RM, we can qualitatively see that it fits the data similarly well: >>> >>> >>> >>>

lin_regplot(X, y, slr) plt.xlabel('Average number of rooms [RM]') plt.ylabel('Price in $1000s [MEDV]') plt.show()

For instance, we can see that the overall result looks identical to our GD implementation:

[ 330 ]

Chapter 10

Analytical solutions of linear regression As an alternative to using machine learning libraries, there is also a closed-form solution for solving OLS involving a system of linear equations that can be found in most introductory statistics textbooks: 𝒘𝒘 = (𝑿𝑿𝑇𝑇 𝑿𝑿)−1 𝑿𝑿𝑇𝑇 𝒚𝒚 We can implement it in Python as follows:

# adding a column vector of "ones" >>> Xb = np.hstack((np.ones((X.shape[0], 1)), X)) >>> w = np.zeros(X.shape[1]) >>> z = np.linalg.inv(np.dot(Xb.T, Xb)) >>> w = np.dot(z, np.dot(Xb.T, y)) >>> print('Slope: %.3f' % w[1]) Slope: 9.102 >>> print('Intercept: %.3f' % w[0]) Intercept: -34.671

The advantage of this method is that it is guaranteed to find the optimal solution analytically. However, if we are working with very large datasets, it can be computationally too expensive to invert the matrix in this formula (sometimes also called the normal equation), or the matrix containing the training examples may be singular (non-invertible), which is why we may prefer iterative methods in certain cases. If you are interested in more information on how to obtain normal equations, take a look at Dr. Stephen Pollock's chapter The Classical Linear Regression Model from his lectures at the University of Leicester, which is available for free at: http://www.le.ac.uk/users/dsgp1/COURSES/MESOMET/ ECMETXT/06mesmet.pdf.

Also, if you want to compare linear regression solutions obtained via GD, SGD, the closed-form solution, QR factorization, and singular vector decomposition, you can use the LinearRegression class implemented in MLxtend (http://rasbt.github.io/mlxtend/user_guide/ regressor/LinearRegression/), which lets users toggle between these options. Another great library to recommend for regression modeling in Python is Statsmodels, which implements more advanced linear regression models, as illustrated at https://www.statsmodels.org/stable/ examples/index.html#regression. [ 331 ]

Predicting Continuous Target Variables with Regression Analysis

Fitting a robust regression model using RANSAC

Linear regression models can be heavily impacted by the presence of outliers. In certain situations, a very small subset of our data can have a big effect on the estimated model coefficients. There are many statistical tests that can be used to detect outliers, which are beyond the scope of the book. However, removing outliers always requires our own judgment as data scientists as well as our domain knowledge. As an alternative to throwing out outliers, we will look at a robust method of regression using the RANdom SAmple Consensus (RANSAC) algorithm, which fits a regression model to a subset of the data, the so-called inliers. We can summarize the iterative RANSAC algorithm as follows: 1. Select a random number of examples to be inliers and fit the model. 2. Test all other data points against the fitted model and add those points that fall within a user-given tolerance to the inliers. 3. Refit the model using all inliers. 4. Estimate the error of the fitted model versus the inliers. 5. Terminate the algorithm if the performance meets a certain user-defined threshold or if a fixed number of iterations were reached; go back to step 1 otherwise. Let's now use a linear model in combination with the RANSAC algorithm as implemented in scikit-learn's RANSACRegressor class: >>> from sklearn.linear_model import RANSACRegressor >>> ransac = RANSACRegressor(LinearRegression(), ... max_trials=100, ... min_samples=50, ... loss='absolute_loss', ... residual_threshold=5.0, ... random_state=0) >>> ransac.fit(X, y)

We set the maximum number of iterations of the RANSACRegressor to 100, and using min_samples=50, we set the minimum number of the randomly chosen training examples to be at least 50. Using 'absolute_loss' as an argument for the loss parameter, the algorithm computes absolute vertical distances between the fitted line and the training examples. By setting the residual_threshold parameter to 5.0, we only allow training examples to be included in the inlier set if their vertical distance to the fitted line is within 5 distance units, which works well on this particular dataset. [ 332 ]

Chapter 10

By default, scikit-learn uses the MAD estimate to select the inlier threshold, where MAD stands for the median absolute deviation of the target values, y. However, the choice of an appropriate value for the inlier threshold is problem-specific, which is one disadvantage of RANSAC. Many different approaches have been developed in recent years to select a good inlier threshold automatically. You can find a detailed discussion in Automatic Estimation of the Inlier Threshold in Robust Multiple Structures Fitting, R. Toldo, A. Fusiello's, Springer, 2009 (in Image Analysis and Processing–ICIAP 2009, pages: 123-131). Once we have fitted the RANSAC model, let's obtain the inliers and outliers from the fitted RANSAC-linear regression model and plot them together with the linear fit: >>> >>> >>> >>> >>> ... ... >>> ... ... >>> >>> >>> >>> >>>

inlier_mask = ransac.inlier_mask_ outlier_mask = np.logical_not(inlier_mask) line_X = np.arange(3, 10, 1) line_y_ransac = ransac.predict(line_X[:, np.newaxis]) plt.scatter(X[inlier_mask], y[inlier_mask], c='steelblue', edgecolor='white', marker='o', label='Inliers') plt.scatter(X[outlier_mask], y[outlier_mask], c='limegreen', edgecolor='white', marker='s', label='Outliers') plt.plot(line_X, line_y_ransac, color='black', lw=2) plt.xlabel('Average number of rooms [RM]') plt.ylabel('Price in $1000s [MEDV]') plt.legend(loc='upper left') plt.show()

As you can see in the following scatterplot, the linear regression model was fitted on the detected set of inliers, which are shown as circles:

[ 333 ]

Predicting Continuous Target Variables with Regression Analysis

When we print the slope and intercept of the model by executing the following code, the linear regression line will be slightly different from the fit that we obtained in the previous section without using RANSAC: >>> print('Slope: %.3f' % ransac.estimator_.coef_[0]) Slope: 10.735 >>> print('Intercept: %.3f' % ransac.estimator_.intercept_) Intercept: -44.089

Using RANSAC, we reduced the potential effect of the outliers in this dataset, but we don't know whether this approach will have a positive effect on the predictive performance for unseen data or not. Thus, in the next section, we will look at different approaches for evaluating a regression model, which is a crucial part of building systems for predictive modeling.

Evaluating the performance of linear regression models

In the previous section, you learned how to fit a regression model on training data. However, you discovered in previous chapters that it is crucial to test the model on data that it hasn't seen during training to obtain a more unbiased estimate of its generalization performance. As you will remember from Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, we want to split our dataset into separate training and test datasets, where we will use the former to fit the model and the latter to evaluate its performance on unseen data to estimate the generalization performance. Instead of proceeding with the simple regression model, we will now use all variables in the dataset and train a multiple regression model: >>> >>> >>> >>> ... >>> >>> >>> >>>

from sklearn.model_selection import train_test_split X = df.iloc[:, :-1].values y = df['MEDV'].values X_train, X_test, y_train, y_test = train_test_split( X, y, test_size=0.3, random_state=0) slr = LinearRegression() slr.fit(X_train, y_train) y_train_pred = slr.predict(X_train) y_test_pred = slr.predict(X_test)

[ 334 ]

Chapter 10

Since our model uses multiple explanatory variables, we can't visualize the linear regression line (or hyperplane, to be precise) in a two-dimensional plot, but we can plot the residuals (the differences or vertical distances between the actual and predicted values) versus the predicted values to diagnose our regression model. Residual plots are a commonly used graphical tool for diagnosing regression models. They can help to detect nonlinearity and outliers, and check whether the errors are randomly distributed. Using the following code, we will now plot a residual plot where we simply subtract the true target variables from our predicted responses: >>> ... ... >>> ... ... >>> >>> >>> >>> >>> >>>

plt.scatter(y_train_pred, y_train_pred - y_train, c='steelblue', marker='o', edgecolor='white', label='Training data') plt.scatter(y_test_pred, y_test_pred - y_test, c='limegreen', marker='s', edgecolor='white', label='Test data') plt.xlabel('Predicted values') plt.ylabel('Residuals') plt.legend(loc='upper left') plt.hlines(y=0, xmin=-10, xmax=50, color='black', lw=2) plt.xlim([-10, 50]) plt.show()

After executing the code, we should see a residual plot with a line passing through the x axis origin, as shown here:

[ 335 ]

Predicting Continuous Target Variables with Regression Analysis

In the case of a perfect prediction, the residuals would be exactly zero, which we will probably never encounter in realistic and practical applications. However, for a good regression model, we would expect the errors to be randomly distributed and the residuals to be randomly scattered around the centerline. If we see patterns in a residual plot, it means that our model is unable to capture some explanatory information, which has leaked into the residuals, as you can slightly see in our previous residual plot. Furthermore, we can also use residual plots to detect outliers, which are represented by the points with a large deviation from the centerline. Another useful quantitative measure of a model's performance is the so-called mean squared error (MSE), which is simply the averaged value of the SSE cost that we minimized to fit the linear regression model. The MSE is useful for comparing different regression models or for tuning their parameters via grid search and crossvalidation, as it normalizes the SSE by the sample size: 𝑛𝑛

1 2 𝑀𝑀𝑀𝑀𝑀𝑀 = ∑(𝑦𝑦 (𝑖𝑖) − 𝑦𝑦̂ (𝑖𝑖) ) 𝑛𝑛 𝑖𝑖=1

Let's compute the MSE of our training and test predictions: >>> from sklearn.metrics import mean_squared_error >>> print('MSE train: %.3f, test: %.3f' % ( ... mean_squared_error(y_train, y_train_pred), ... mean_squared_error(y_test, y_test_pred))) MSE train: 19.958, test: 27.196

You can see that the MSE on the training dataset is 19.96, and the MSE on the test dataset is much larger, with a value of 27.20, which is an indicator that our model is overfitting the training data in this case. However, please be aware that the MSE is unbounded in contrast to the classification accuracy, for example. In other words, the interpretation of the MSE depends on the dataset and feature scaling. For example, if the house prices were presented as multiples of 1,000 (with the K suffix), the same model would yield a lower MSE compared to a model that worked with unscaled features. To further illustrate this point, ($10K − 15K)2 < ($10,000 − $15,000)2.

Thus, it may sometimes be more useful to report the coefficient of determination (𝑅𝑅2), which can be understood as a standardized version of the MSE, for better interpretability of the model's performance. Or, in other words, 𝑅𝑅 2 is the fraction of response variance that is captured by the model. The 𝑅𝑅 2 value is defined as:

𝑅𝑅 2 = 1 −

[ 336 ]

𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆

Chapter 10

Here, SSE is the sum of squared errors and SST is the total sum of squares: 𝑛𝑛

2

𝑆𝑆𝑆𝑆𝑆𝑆 = ∑(𝑦𝑦 (𝑖𝑖) − 𝜇𝜇𝑦𝑦 ) 𝑖𝑖=1

In other words, SST is simply the variance of the response. Let's quickly show that 𝑅𝑅 2 is indeed just a rescaled version of the MSE:

𝑅𝑅 2

=

= =

1−

𝑆𝑆𝑆𝑆𝑆𝑆 𝑆𝑆𝑆𝑆𝑆𝑆

1−

𝑀𝑀𝑀𝑀𝑀𝑀 𝑉𝑉𝑉𝑉𝑉𝑉(𝑦𝑦)

1 𝑛𝑛 (𝑖𝑖) (𝑖𝑖) 2 ∑ 𝑛𝑛 𝑖𝑖=1(𝑦𝑦 − 𝑦𝑦̂ ) 2 1 𝑛𝑛 (𝑖𝑖) ∑ 𝑛𝑛 𝑖𝑖=1(𝑦𝑦 − 𝜇𝜇𝑦𝑦 )

For the training dataset, the 𝑅𝑅2 is bounded between 0 and 1, but it can become negative for the test dataset. If 𝑅𝑅2 = 1, the model fits the data perfectly with a corresponding MSE = 0.

Evaluated on the training data, the 𝑅𝑅2 of our model is 0.765, which doesn't sound too bad. However, the 𝑅𝑅2 on the test dataset is only 0.673, which we can compute by executing the following code: >>> from sklearn.metrics import r2_score >>> print('R^2 train: %.3f, test: %.3f' % ... (r2_score(y_train, y_train_pred), ... r2_score(y_test, y_test_pred))) R^2 train: 0.765, test: 0.673

Using regularized methods for regression As we discussed in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, regularization is one approach to tackling the problem of overfitting by adding additional information, and thereby shrinking the parameter values of the model to induce a penalty against complexity. The most popular approaches to regularized linear regression are the so-called Ridge Regression, least absolute shrinkage and selection operator (LASSO), and elastic Net.

[ 337 ]

Predicting Continuous Target Variables with Regression Analysis

Ridge Regression is an L2 penalized model where we simply add the squared sum of the weights to our least-squares cost function: 𝑛𝑛

2

𝐽𝐽(𝒘𝒘)𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = ∑(𝑦𝑦 (𝑖𝑖) − 𝑦𝑦̂ (𝑖𝑖) ) + 𝜆𝜆‖𝒘𝒘‖22 Here:

𝑖𝑖=1

𝐿𝐿2: 𝜆𝜆‖𝒘𝒘‖22

𝑚𝑚

= 𝜆𝜆 ∑ 𝑤𝑤𝑗𝑗 2 𝑗𝑗=1

By increasing the value of hyperparameter 𝜆𝜆 , we increase the regularization strength and thereby shrink the weights of our model. Please note that we don't regularize the intercept term, 𝑤𝑤0. An alternative approach that can lead to sparse models is LASSO. Depending on the regularization strength, certain weights can become zero, which also makes LASSO useful as a supervised feature selection technique: 𝑛𝑛

2

𝐽𝐽(𝒘𝒘)𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = ∑(𝑦𝑦 (𝑖𝑖) − 𝑦𝑦̂ (𝑖𝑖) ) + 𝜆𝜆‖𝒘𝒘‖1 𝑖𝑖=1

Here, the L1 penalty for LASSO is defined as the sum of the absolute magnitudes of the model weights, as follows: 𝑚𝑚

𝐿𝐿1: 𝜆𝜆‖𝒘𝒘‖1 = 𝜆𝜆 ∑|𝑤𝑤𝑗𝑗 | 𝑗𝑗=1

However, a limitation of LASSO is that it selects at most n features if m > n, where n is the number of training examples. This may be undesirable in certain applications of feature selection. In practice, however, this property of LASSO is often an advantage because it avoids saturated models. Saturation of a model occurs if the number of training examples is equal to the number of features, which is a form of overparameterization. As a consequence, a saturated model can always fit the training data perfectly but is merely a form of interpolation and thus is not expected to generalize well.

[ 338 ]

Chapter 10

A compromise between Ridge Regression and LASSO is elastic net, which has an L1 penalty to generate sparsity and an L2 penalty such that it can be used for selecting more than n features if m > n: 𝑛𝑛

𝐽𝐽(𝒘𝒘)𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸 = ∑(𝑦𝑦 𝑖𝑖=1

(𝑖𝑖)

(𝑖𝑖) 2

𝑚𝑚

2

𝑚𝑚

− 𝑦𝑦̂ ) + 𝜆𝜆1 ∑ 𝑤𝑤𝑗𝑗 + 𝜆𝜆2 ∑|𝑤𝑤𝑗𝑗 | 𝑗𝑗=1

𝑗𝑗=1

Those regularized regression models are all available via scikit-learn, and their usage is similar to the regular regression model except that we have to specify the regularization strength via the parameter 𝜆𝜆, for example, optimized via k-fold crossvalidation. A Ridge Regression model can be initialized via:

>>> from sklearn.linear_model import Ridge >>> ridge = Ridge(alpha=1.0)

Note that the regularization strength is regulated by the parameter alpha, which is similar to the parameter 𝜆𝜆. Likewise, we can initialize a LASSO regressor from the linear_model submodule: >>> from sklearn.linear_model import Lasso >>> lasso = Lasso(alpha=1.0)

Lastly, the ElasticNet implementation allows us to vary the L1 to L2 ratio: >>> from sklearn.linear_model import ElasticNet >>> elanet = ElasticNet(alpha=1.0, l1_ratio=0.5)

For example, if we set the l1_ratio to 1.0, the ElasticNet regressor would be equal to LASSO regression. For more detailed information about the different implementations of linear regression, please see the documentation at http://scikit-learn.org/stable/modules/linear_model.html.

Turning a linear regression model into a curve – polynomial regression

In the previous sections, we assumed a linear relationship between explanatory and response variables. One way to account for the violation of linearity assumption is to use a polynomial regression model by adding polynomial terms:

𝑦𝑦 = 𝑤𝑤0 + 𝑤𝑤1 𝑥𝑥 + 𝑤𝑤2 𝑥𝑥 2 +. . . +𝑤𝑤𝑑𝑑 𝑥𝑥 𝑑𝑑 [ 339 ]

Predicting Continuous Target Variables with Regression Analysis

Here, d denotes the degree of the polynomial. Although we can use polynomial regression to model a nonlinear relationship, it is still considered a multiple linear regression model because of the linear regression coefficients, w. In the following subsections, we will see how we can add such polynomial terms to an existing dataset conveniently and fit a polynomial regression model.

Adding polynomial terms using scikit-learn

We will now learn how to use the PolynomialFeatures transformer class from scikit-learn to add a quadratic term (d = 2) to a simple regression problem with one explanatory variable. Then, we will compare the polynomial to the linear fit by following these steps: 1. Add a second-degree polynomial term: >>> >>> ... ... >>> ... >>> >>> >>> >>>

from sklearn.preprocessing import PolynomialFeatures X = np.array([ 258.0, 270.0, 294.0, 320.0, 342.0, 368.0, 396.0, 446.0, 480.0, 586.0])\ [:, np.newaxis] y = np.array([ 236.4, 234.4, 252.8, 298.6, 314.2, 342.2, 360.8, 368.0, 391.2, 390.8]) lr = LinearRegression() pr = LinearRegression() quadratic = PolynomialFeatures(degree=2) X_quad = quadratic.fit_transform(X)

2. Fit a simple linear regression model for comparison: >>> lr.fit(X, y) >>> X_fit = np.arange(250, 600, 10)[:, np.newaxis] >>> y_lin_fit = lr.predict(X_fit)

3. Fit a multiple regression model on the transformed features for polynomial regression: >>> pr.fit(X_quad, y) >>> y_quad_fit = pr.predict(quadratic.fit_transform(X_fit))

4. Plot the results: >>> >>> ... >>> ... >>> >>> >>> >>> >>>

plt.scatter(X, y, label='Training points') plt.plot(X_fit, y_lin_fit, label='Linear fit', linestyle='--') plt.plot(X_fit, y_quad_fit, label='Quadratic fit') plt.xlabel('Explanatory variable') plt.ylabel('Predicted or known target values') plt.legend(loc='upper left') plt.tight_layout() plt.show() [ 340 ]

Chapter 10

In the resulting plot, you can see that the polynomial fit captures the relationship between the response and explanatory variables much better than the linear fit:

Next, we will compute the MSE and 𝑅𝑅2 evaluation metrics:

>>> y_lin_pred = lr.predict(X) >>> y_quad_pred = pr.predict(X_quad) >>> print('Training MSE linear: %.3f, quadratic: %.3f' % ( ... mean_squared_error(y, y_lin_pred), ... mean_squared_error(y, y_quad_pred))) Training MSE linear: 569.780, quadratic: 61.330 >>> print('Training R^2 linear: %.3f, quadratic: %.3f' % ( ... r2_score(y, y_lin_pred), ... r2_score(y, y_quad_pred))) Training R^2 linear: 0.832, quadratic: 0.982

As you can see after executing the code, the MSE decreased from 570 (linear fit) to 61 (quadratic fit); also, the coefficient of determination reflects a closer fit of the quadratic model (𝑅𝑅2 = 0.982) as opposed to the linear fit (𝑅𝑅2 = 0.832) in this particular toy problem.

[ 341 ]

Predicting Continuous Target Variables with Regression Analysis

Modeling nonlinear relationships in the Housing dataset

In the preceding subsection, you learned how to construct polynomial features to fit nonlinear relationships in a toy problem; let's now take a look at a more concrete example and apply those concepts to the data in the Housing dataset. By executing the following code, we will model the relationship between house prices and LSTAT (percentage of lower status of the population) using second-degree (quadratic) and third-degree (cubic) polynomials and compare that to a linear fit: >>> X = df[['LSTAT']].values >>> y = df['MEDV'].values >>> regr = LinearRegression() # create quadratic features >>> quadratic = PolynomialFeatures(degree=2) >>> cubic = PolynomialFeatures(degree=3) >>> X_quad = quadratic.fit_transform(X) >>> X_cubic = cubic.fit_transform(X) # fit features >>> X_fit = np.arange(X.min(), X.max(), 1)[:, np.newaxis] >>> regr = regr.fit(X, y) >>> y_lin_fit = regr.predict(X_fit) >>> linear_r2 = r2_score(y, regr.predict(X)) >>> regr = regr.fit(X_quad, y) >>> y_quad_fit = regr.predict(quadratic.fit_transform(X_fit)) >>> quadratic_r2 = r2_score(y, regr.predict(X_quad)) >>> regr = regr.fit(X_cubic, y) >>> y_cubic_fit = regr.predict(cubic.fit_transform(X_fit)) >>> cubic_r2 = r2_score(y, regr.predict(X_cubic)) # plot results >>> plt.scatter(X, y, label='Training points', color='lightgray')

[ 342 ]

Chapter 10 >>> plt.plot(X_fit, y_lin_fit, ... label='Linear (d=1), $R^2=%.2f$' % linear_r2, ... color='blue', ... lw=2, ... linestyle=':') >>> plt.plot(X_fit, y_quad_fit, ... label='Quadratic (d=2), $R^2=%.2f$' % quadratic_r2, ... color='red', ... lw=2, ... linestyle='-') >>> plt.plot(X_fit, y_cubic_fit, ... label='Cubic (d=3), $R^2=%.2f$' % cubic_r2, ... color='green', ... lw=2, ... linestyle='--') >>> >>> >>> >>>

plt.xlabel('% lower status of the population [LSTAT]') plt.ylabel('Price in $1000s [MEDV]') plt.legend(loc='upper right') plt.show()

The resulting plot is as follows:

[ 343 ]

Predicting Continuous Target Variables with Regression Analysis

As you can see, the cubic fit captures the relationship between house prices and LSTAT better than the linear and quadratic fit. However, you should be aware that adding more and more polynomial features increases the complexity of a model and therefore increases the chance of overfitting. Thus, in practice, it is always recommended to evaluate the performance of the model on a separate test dataset to estimate the generalization performance. In addition, polynomial features are not always the best choice for modeling nonlinear relationships. For example, with some experience or intuition, just looking at the MEDV-LSTAT scatterplot may lead to the hypothesis that a log-transformation of the LSTAT feature variable and the square root of MEDV may project the data onto a linear feature space suitable for a linear regression fit. For instance, my perception is that this relationship between the two variables looks quite similar to an exponential function:

𝑓𝑓(𝑥𝑥) = 𝑒𝑒 −𝑥𝑥

Since the natural logarithm of an exponential function is a straight line, I assume that such a log-transformation can be usefully applied here: log(𝑓𝑓(𝑥𝑥)) = −𝑥𝑥

Let's test this hypothesis by executing the following code: >>> >>> >>> >>> >>> >>> ... >>> >>> >>>

# transform features X_log = np.log(X) y_sqrt = np.sqrt(y) # fit features X_fit = np.arange(X_log.min()-1, X_log.max()+1, 1)[:, np.newaxis] regr = regr.fit(X_log, y_sqrt) y_lin_fit = regr.predict(X_fit) linear_r2 = r2_score(y_sqrt, regr.predict(X_log))

>>> # plot results >>> plt.scatter(X_log, y_sqrt, ... label='Training points', ... color='lightgray') >>> plt.plot(X_fit, y_lin_fit, ... label='Linear (d=1), $R^2=%.2f$' % linear_r2, ... color='blue', ... lw=2)

[ 344 ]

Chapter 10 >>> >>> >>> >>> >>>

plt.xlabel('log(% lower status of the population [LSTAT])') plt.ylabel('$\sqrt{Price \; in \; \$1000s \; [MEDV]}$') plt.legend(loc='lower left') plt.tight_layout() plt.show()

After transforming the explanatory onto the log space and taking the square root of the target variables, we were able to capture the relationship between the two variables with a linear regression line that seems to fit the data better (𝑅𝑅 2 = 0.69) than any of the previous polynomial feature transformations:

Dealing with nonlinear relationships using random forests

In this section, we are going to take a look at random forest regression, which is conceptually different from the previous regression models in this chapter. A random forest, which is an ensemble of multiple decision trees, can be understood as the sum of piecewise linear functions, in contrast to the global linear and polynomial regression models that we discussed previously. In other words, via the decision tree algorithm, we subdivide the input space into smaller regions that become more manageable.

[ 345 ]

Predicting Continuous Target Variables with Regression Analysis

Decision tree regression

An advantage of the decision tree algorithm is that it does not require any transformation of the features if we are dealing with nonlinear data, because decision trees analyze one feature at a time, rather than taking weighted combinations into account. (Likewise, normalizing or standardizing features is not required for decision trees.) You will remember from Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, that we grow a decision tree by iteratively splitting its nodes until the leaves are pure or a stopping criterion is satisfied. When we used decision trees for classification, we defined entropy as a measure of impurity to determine which feature split maximizes the information gain (IG), which can be defined as follows for a binary split:

𝐼𝐼𝐼𝐼(𝐷𝐷𝑝𝑝 , 𝑥𝑥𝑖𝑖 ) = 𝐼𝐼(𝐷𝐷𝑝𝑝 ) −

𝑁𝑁𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑁𝑁𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 𝐼𝐼(𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 ) − 𝐼𝐼(𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 ) 𝑁𝑁𝑝𝑝 𝑁𝑁𝑝𝑝

Here, 𝑥𝑥𝑖𝑖 is the feature to perform the split, 𝑁𝑁𝑝𝑝 is the number of training examples in the parent node, I is the impurity function, 𝐷𝐷𝑝𝑝 is the subset of training examples at the parent node, and 𝐷𝐷𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 and 𝐷𝐷𝑟𝑟𝑟𝑟𝑟𝑟ℎ𝑡𝑡 are the subsets of training examples at the left and right child nodes after the split. Remember that our goal is to find the feature split that maximizes the information gain; in other words, we want to find the feature split that reduces the impurities in the child nodes most. In Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, we discussed Gini impurity and entropy as measures of impurity, which are both useful criteria for classification. To use a decision tree for regression, however, we need an impurity metric that is suitable for continuous variables, so we define the impurity measure of a node, t, as the MSE instead:

𝐼𝐼(𝑡𝑡) = 𝑀𝑀𝑀𝑀𝑀𝑀(𝑡𝑡) =

1 2 ∑(𝑦𝑦 (𝑖𝑖) − 𝑦𝑦̂𝑡𝑡 ) 𝑁𝑁𝑡𝑡 𝑖𝑖∈𝐷𝐷𝑡𝑡

Here, 𝑁𝑁𝑡𝑡 is the number of training examples at node t, 𝐷𝐷𝑡𝑡 is the training subset at node t, 𝑦𝑦 (𝑖𝑖) is the true target value, and 𝑦𝑦̂𝑡𝑡 is the predicted target value (sample mean):

𝑦𝑦̂𝑡𝑡 =

1 ∑ 𝑦𝑦 (𝑖𝑖) 𝑁𝑁𝑡𝑡 𝑖𝑖∈𝐷𝐷𝑡𝑡

[ 346 ]

Chapter 10

In the context of decision tree regression, the MSE is often referred to as withinnode variance, which is why the splitting criterion is also better known as variance reduction. To see what the line fit of a decision tree looks like, let's use the DecisionTreeRegressor implemented in scikit-learn to model the nonlinear relationship between the MEDV and LSTAT variables: >>> >>> >>> >>> >>> >>> >>> >>> >>> >>>

from sklearn.tree import DecisionTreeRegressor X = df[['LSTAT']].values y = df['MEDV'].values tree = DecisionTreeRegressor(max_depth=3) tree.fit(X, y) sort_idx = X.flatten().argsort() lin_regplot(X[sort_idx], y[sort_idx], tree) plt.xlabel('% lower status of the population [LSTAT]') plt.ylabel('Price in $1000s [MEDV]') plt.show()

As you can see in the resulting plot, the decision tree captures the general trend in the data. However, a limitation of this model is that it does not capture the continuity and differentiability of the desired prediction. In addition, we need to be careful about choosing an appropriate value for the depth of the tree so as to not overfit or underfit the data; here, a depth of three seemed to be a good choice:

In the next section, we will take a look at a more robust way of fitting regression trees: random forests.

[ 347 ]

Predicting Continuous Target Variables with Regression Analysis

Random forest regression

As you learned in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, the random forest algorithm is an ensemble technique that combines multiple decision trees. A random forest usually has a better generalization performance than an individual decision tree due to randomness, which helps to decrease the model's variance. Other advantages of random forests are that they are less sensitive to outliers in the dataset and don't require much parameter tuning. The only parameter in random forests that we typically need to experiment with is the number of trees in the ensemble. The basic random forest algorithm for regression is almost identical to the random forest algorithm for classification that we discussed in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn. The only difference is that we use the MSE criterion to grow the individual decision trees, and the predicted target variable is calculated as the average prediction over all decision trees. Now, let's use all the features in the Housing dataset to fit a random forest regression model on 60 percent of the examples and evaluate its performance on the remaining 40 percent. The code is as follows: >>> >>> >>> ... ... ... >>> >>> >>> ... ... ... >>> >>> >>> >>> ... ... MSE >>> ... ... R^2

X = df.iloc[:, :-1].values y = df['MEDV'].values X_train, X_test, y_train, y_test =\ train_test_split(X, y, test_size=0.4, random_state=1) from sklearn.ensemble import RandomForestRegressor forest = RandomForestRegressor(n_estimators=1000, criterion='mse', random_state=1, n_jobs=-1) forest.fit(X_train, y_train) y_train_pred = forest.predict(X_train) y_test_pred = forest.predict(X_test) print('MSE train: %.3f, test: %.3f' % ( mean_squared_error(y_train, y_train_pred), mean_squared_error(y_test, y_test_pred))) train: 1.642, test: 11.052 print('R^2 train: %.3f, test: %.3f' % ( r2_score(y_train, y_train_pred), r2_score(y_test, y_test_pred))) train: 0.979, test: 0.878

Unfortunately, you can see that the random forest tends to overfit the training data. However, it's still able to explain the relationship between the target and explanatory variables relatively well (𝑅𝑅2 = 0.878 on the test dataset). [ 348 ]

Chapter 10

Lastly, let's also take a look at the residuals of the prediction: >>> ... ... ... ... ... ... ... >>> ... ... ... ... ... ... ... >>> >>> >>> >>> >>> >>> >>>

plt.scatter(y_train_pred, y_train_pred - y_train, c='steelblue', edgecolor='white', marker='o', s=35, alpha=0.9, label='Training data') plt.scatter(y_test_pred, y_test_pred - y_test, c='limegreen', edgecolor='white', marker='s', s=35, alpha=0.9, label='Test data') plt.xlabel('Predicted values') plt.ylabel('Residuals') plt.legend(loc='upper left') plt.hlines(y=0, xmin=-10, xmax=50, lw=2, color='black') plt.xlim([-10, 50]) plt.tight_layout() plt.show()

As it was already summarized by the 𝑅𝑅2 coefficient, you can see that the model fits the training data better than the test data, as indicated by the outliers in the y axis direction. Also, the distribution of the residuals does not seem to be completely random around the zero center point, indicating that the model is not able to capture all the exploratory information. However, the residual plot indicates a large improvement over the residual plot of the linear model that we plotted earlier in this chapter:

[ 349 ]

Predicting Continuous Target Variables with Regression Analysis

Ideally, our model error should be random or unpredictable. In other words, the error of the predictions should not be related to any of the information contained in the explanatory variables; rather, it should reflect the randomness of the real-world distributions or patterns. If we find patterns in the prediction errors, for example, by inspecting the residual plot, it means that the residual plots contain predictive information. A common reason for this could be that explanatory information is leaking into those residuals. Unfortunately, there is not a universal approach for dealing with non-randomness in residual plots, and it requires experimentation. Depending on the data that is available to us, we may be able to improve the model by transforming variables, tuning the hyperparameters of the learning algorithm, choosing simpler or more complex models, removing outliers, or including additional variables. Regression with support vector machines In Chapter 3, A Tour of Machine Learning Classifiers Using scikitlearn, we also learned about the kernel trick, which can be used in combination with a support vector machine (SVM) for classification, and is useful if we are dealing with nonlinear problems. Although a discussion is beyond the scope of this book, SVMs can also be used in nonlinear regression tasks. The interested reader can find more information about SVMs for regression in an excellent report: Support Vector Machines for Classification and Regression, S. R. Gunn and others, University of Southampton technical report, 14, 1998 (http://citeseerx. ist.psu.edu/viewdoc/download?doi=10.1.1.579.6867& rep=rep1&type=pdf). An SVM regressor is also implemented

in scikit-learn, and more information about its usage can be found at http://scikit-learn.org/stable/modules/ generated/sklearn.svm.SVR.html#sklearn.svm.SVR.

Summary

At the beginning of this chapter, you learned about simple linear regression analysis to model the relationship between a single explanatory variable and a continuous response variable. We then discussed a useful explanatory data analysis technique to look at patterns and anomalies in data, which is an important first step in predictive modeling tasks.

[ 350 ]

Chapter 10

We built our first model by implementing linear regression using a gradient-based optimization approach. You then saw how to utilize scikit-learn's linear models for regression and also implement a robust regression technique (RANSAC) as an approach for dealing with outliers. To assess the predictive performance of regression models, we computed the mean sum of squared errors and the related 𝑅𝑅2 metric. Furthermore, we also discussed a useful graphical approach to diagnose the problems of regression models: the residual plot. After we explored how regularization can be applied to regression models to reduce the model complexity and avoid overfitting, we also covered several approaches to model nonlinear relationships, including polynomial feature transformation and random forest regressors. We discussed supervised learning, classification, and regression analysis in great detail in the previous chapters. In the next chapter, we are going to learn about another interesting subfield of machine learning, unsupervised learning, and also how to use cluster analysis to find hidden structures in data in the absence of target variables.

[ 351 ]

Working with Unlabeled Data – Clustering Analysis In the previous chapters, we used supervised learning techniques to build machine learning models, using data where the answer was already known—the class labels were already available in our training data. In this chapter, we will switch gears and explore cluster analysis, a category of unsupervised learning techniques that allows us to discover hidden structures in data where we do not know the right answer upfront. The goal of clustering is to find a natural grouping in data so that items in the same cluster are more similar to each other than to those from different clusters. Given its exploratory nature, clustering is an exciting topic, and in this chapter, you will learn about the following concepts, which can help us to organize data into meaningful structures: • Finding centers of similarity using the popular k-means algorithm • Taking a bottom-up approach to building hierarchical clustering trees • Identifying arbitrary shapes of objects using a density-based clustering approach

Grouping objects by similarity using k-means

In this section, we will learn about one of the most popular clustering algorithms, k-means, which is widely used in academia as well as in industry. Clustering (or cluster analysis) is a technique that allows us to find groups of similar objects that are more related to each other than to objects in other groups. Examples of businessoriented applications of clustering include the grouping of documents, music, and movies by different topics, or finding customers that share similar interests based on common purchase behaviors as a basis for recommendation engines. [ 353 ]

Working with Unlabeled Data – Clustering Analysis

K-means clustering using scikit-learn

As you will see in a moment, the k-means algorithm is extremely easy to implement, but it is also computationally very efficient compared to other clustering algorithms, which might explain its popularity. The k-means algorithm belongs to the category of prototype-based clustering. We will discuss two other categories of clustering, hierarchical and density-based clustering, later in this chapter. Prototype-based clustering means that each cluster is represented by a prototype, which is usually either the centroid (average) of similar points with continuous features, or the medoid (the most representative or the point that minimizes the distance to all other points that belong to a particular cluster) in the case of categorical features. While k-means is very good at identifying clusters with a spherical shape, one of the drawbacks of this clustering algorithm is that we have to specify the number of clusters, k, a priori. An inappropriate choice for k can result in poor clustering performance. Later in this chapter, we will discuss the elbow method and silhouette plots, which are useful techniques to evaluate the quality of a clustering to help us determine the optimal number of clusters, k. Although k-means clustering can be applied to data in higher dimensions, we will walk through the following examples using a simple two-dimensional dataset for the purpose of visualization: >>> from sklearn.datasets import make_blobs >>> X, y = make_blobs(n_samples=150, ... n_features=2, ... centers=3, ... cluster_std=0.5, ... shuffle=True, ... random_state=0) >>> >>> ... ... ... ... ... >>> >>> >>>

import matplotlib.pyplot as plt plt.scatter(X[:, 0], X[:, 1], c='white', marker='o', edgecolor='black', s=50) plt.grid() plt.tight_layout() plt.show()

The dataset that we just created consists of 150 randomly generated points that are roughly grouped into three regions with higher density, which is visualized via a two-dimensional scatterplot:

[ 354 ]

Chapter 11

In real-world applications of clustering, we do not have any ground truth category information (information provided as empirical evidence as opposed to inference) about those examples; if we were given class labels, this task would fall into the category of supervised learning. Thus, our goal is to group the examples based on their feature similarities, which can be achieved using the k-means algorithm, as summarized by the following four steps: 1. Randomly pick k centroids from the examples as initial cluster centers. 2. Assign each example to the nearest centroid, 𝜇𝜇 (𝑗𝑗) , 𝑗𝑗 ∈ {1, … , 𝑘𝑘} .

3. Move the centroids to the center of the examples that were assigned to it. 4. Repeat steps 2 and 3 until the cluster assignments do not change or a userdefined tolerance or maximum number of iterations is reached.

Now, the next question is, how do we measure similarity between objects? We can define similarity as the opposite of distance, and a commonly used distance for clustering examples with continuous features is the squared Euclidean distance between two points, x and y, in m-dimensional space: 𝑚𝑚

𝑑𝑑(𝒙𝒙, 𝒚𝒚) = ∑(𝑥𝑥𝑗𝑗 − 𝑦𝑦𝑗𝑗 )2 = ‖𝒙𝒙 − 𝒚𝒚‖22 2

𝑗𝑗=1

Note that, in the preceding equation, the index j refers to the jth dimension (feature column) of the example inputs, x and y. In the rest of this section, we will use the superscripts i and j to refer to the index of the example (data record) and cluster index, respectively.

[ 355 ]

Working with Unlabeled Data – Clustering Analysis

Based on this Euclidean distance metric, we can describe the k-means algorithm as a simple optimization problem, an iterative approach for minimizing the withincluster sum of squared errors (SSE), which is sometimes also called cluster inertia: 𝑛𝑛

𝑘𝑘

2

𝑆𝑆𝑆𝑆𝑆𝑆 = ∑ ∑ 𝑤𝑤 (𝑖𝑖,𝑗𝑗) ‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(𝑗𝑗) ‖2 𝑖𝑖=1 𝑗𝑗=1

Here, 𝝁𝝁(𝑗𝑗) is the representative point (centroid) for cluster j. 𝑤𝑤 (𝑖𝑖,𝑗𝑗) = 1 if the example, 𝒙𝒙(𝑖𝑖) , is in cluster j, or 0 otherwise.

𝑤𝑤 (𝑖𝑖,𝑗𝑗) = {

1, if 𝑥𝑥 (𝑖𝑖) ∈ 𝑗𝑗 0, otherwise

Now that you have learned how the simple k-means algorithm works, let's apply it to our example dataset using the KMeans class from scikit-learn's cluster module: >>> from sklearn.cluster import KMeans >>> km = KMeans(n_clusters=3, ... init='random', ... n_init=10, ... max_iter=300, ... tol=1e-04, ... random_state=0) >>> y_km = km.fit_predict(X)

Using the preceding code, we set the number of desired clusters to 3; specifying the number of clusters a priori is one of the limitations of k-means. We set n_init=10 to run the k-means clustering algorithms 10 times independently, with different random centroids to choose the final model as the one with the lowest SSE. Via the max_iter parameter, we specify the maximum number of iterations for each single run (here, 300). Note that the k-means implementation in scikit-learn stops early if it converges before the maximum number of iterations is reached. However, it is possible that k-means does not reach convergence for a particular run, which can be problematic (computationally expensive) if we choose relatively large values for max_iter. One way to deal with convergence problems is to choose larger values for tol, which is a parameter that controls the tolerance with regard to the changes in the within-cluster SSE to declare convergence. In the preceding code, we chose a tolerance of 1e-04 (=0.0001). A problem with k-means is that one or more clusters can be empty. Note that this problem does not exist for k-medoids or fuzzy C-means, an algorithm that we will discuss later in this section.

[ 356 ]

Chapter 11

However, this problem is accounted for in the current k-means implementation in scikit-learn. If a cluster is empty, the algorithm will search for the example that is farthest away from the centroid of the empty cluster. Then it will reassign the centroid to be this farthest point. Feature scaling When we are applying k-means to real-world data using a Euclidean distance metric, we want to make sure that the features are measured on the same scale and apply z-score standardization or min-max scaling if necessary.

Having predicted the cluster labels, y_km, and discussed some of the challenges of the k-means algorithm, let's now visualize the clusters that k-means identified in the dataset together with the cluster centroids. These are stored under the cluster_centers_ attribute of the fitted KMeans object: >>> ... ... ... ... >>> ... ... ... ... >>> ... ... ... ... >>> ... ... ... ... >>> >>> >>> >>>

plt.scatter(X[y_km == 0, 0], X[y_km == 0, 1], s=50, c='lightgreen', marker='s', edgecolor='black', label='Cluster 1') plt.scatter(X[y_km == 1, 0], X[y_km == 1, 1], s=50, c='orange', marker='o', edgecolor='black', label='Cluster 2') plt.scatter(X[y_km == 2, 0], X[y_km == 2, 1], s=50, c='lightblue', marker='v', edgecolor='black', label='Cluster 3') plt.scatter(km.cluster_centers_[:, 0], km.cluster_centers_[:, 1], s=250, marker='*', c='red', edgecolor='black', label='Centroids') plt.legend(scatterpoints=1) plt.grid() plt.tight_layout() plt.show()

[ 357 ]

Working with Unlabeled Data – Clustering Analysis

In the following scatterplot, you can see that k-means placed the three centroids at the center of each sphere, which looks like a reasonable grouping given this dataset:

Although k-means worked well on this toy dataset, we will highlight another drawback of k-means: we have to specify the number of clusters, k, a priori. The number of clusters to choose may not always be so obvious in real-world applications, especially if we are working with a higher dimensional dataset that cannot be visualized. The other properties of k-means are that clusters do not overlap and are not hierarchical, and we also assume that there is at least one item in each cluster. Later in this chapter, we will encounter different types of clustering algorithms, hierarchical and density-based clustering. Neither type of algorithm requires us to specify the number of clusters upfront or assume spherical structures in our dataset. In the next subsection, we will cover a popular variant of the classic k-means algorithm called k-means++. While it doesn't address those assumptions and drawbacks of k-means that were discussed in the previous paragraph, it can greatly improve the clustering results through more clever seeding of the initial cluster centers.

A smarter way of placing the initial cluster centroids using k-means++

So far, we have discussed the classic k-means algorithm, which uses a random seed to place the initial centroids, which can sometimes result in bad clusterings or slow convergence if the initial centroids are chosen poorly. One way to address this issue is to run the k-means algorithm multiple times on a dataset and choose the best performing model in terms of the SSE. [ 358 ]

Chapter 11

Another strategy is to place the initial centroids far away from each other via the k-means++ algorithm, which leads to better and more consistent results than the classic k-means (k-means++: The Advantages of Careful Seeding, D. Arthur and S. Vassilvitskii in Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1027-1035. Society for Industrial and Applied Mathematics, 2007). The initialization in k-means++ can be summarized as follows: 1. Initialize an empty set, M, to store the k centroids being selected. 2. Randomly choose the first centroid, 𝝁𝝁(𝑗𝑗), from the input examples and assign it to M. 3. For each example, 𝒙𝒙(𝑖𝑖), that is not in M, find the minimum squared distance, 2 𝑑𝑑(𝒙𝒙(𝑖𝑖) , 𝚳𝚳) , to any of the centroids in M.

4. To randomly select the next centroid, 𝝁𝝁(𝑝𝑝), use a weighted probability 2

𝑑𝑑(𝝁𝝁(𝑝𝑝) , 𝚳𝚳) . distribution equal to ∑𝑖𝑖 𝑑𝑑(𝒙𝒙(𝑖𝑖) , 𝚳𝚳)2

5. Repeat steps 2 and 3 until k centroids are chosen. 6. Proceed with the classic k-means algorithm. To use k-means++ with scikit-learn's KMeans object, we just need to set the init parameter to 'k-means++'. In fact, 'k-means++' is the default argument to the init parameter, which is strongly recommended in practice. The only reason we didn't use it in the previous example was to not introduce too many concepts all at once. The rest of this section on k-means will use k-means++, but you are encouraged to experiment more with the two different approaches (classic k-means via init='random' versus k-means++ via init='k-means++') for placing the initial cluster centroids.

Hard versus soft clustering

Hard clustering describes a family of algorithms where each example in a dataset is assigned to exactly one cluster, as in the k-means and k-means++ algorithms that we discussed earlier in this chapter. In contrast, algorithms for soft clustering (sometimes also called fuzzy clustering) assign an example to one or more clusters. A popular example of soft clustering is the fuzzy C-means (FCM) algorithm (also called soft k-means or fuzzy k-means). The original idea goes back to the 1970s, when Joseph C. Dunn first proposed an early version of fuzzy clustering to improve k-means (A Fuzzy Relative of the ISODATA Process and Its Use in Detecting Compact Well-Separated Clusters, J. C. Dunn, 1973). Almost a decade later, James C. Bedzek published his work on the improvement of the fuzzy clustering algorithm, which is now known as the FCM algorithm (Pattern Recognition with Fuzzy Objective Function Algorithms, J. C. Bezdek, Springer Science+Business Media, 2013). [ 359 ]

Working with Unlabeled Data – Clustering Analysis

The FCM procedure is very similar to k-means. However, we replace the hard cluster assignment with probabilities for each point belonging to each cluster. In k-means, we could express the cluster membership of an example, x, with a sparse vector of binary values:

𝑥𝑥 ∈ 𝜇𝜇 (1) [𝑥𝑥 ∈ 𝜇𝜇 (2) 𝑥𝑥 ∈ 𝜇𝜇 (3)

→ 𝑤𝑤 (𝑖𝑖,𝑗𝑗) = 0 → 𝑤𝑤 (𝑖𝑖,𝑗𝑗) = 1] → 𝑤𝑤 (𝑖𝑖,𝑗𝑗) = 0

Here, the index position with value 1 indicates the cluster centroid, 𝝁𝝁(𝑗𝑗), that the {1,3}2,).3} 𝑘𝑘 = 𝑘𝑘3,= 3, 𝑗𝑗 ∈ 𝑗𝑗{1,∈ 2, In contrast, a membership example is assigned to (assuming vector in FCM could be represented as follows:

𝑥𝑥 ∈ 𝜇𝜇 (1) [𝑥𝑥 ∈ 𝜇𝜇 (2) 𝑥𝑥 ∈ 𝜇𝜇 (3)

→ 𝑤𝑤 (𝑖𝑖,𝑗𝑗) = 0.1 → 𝑤𝑤 (𝑖𝑖,𝑗𝑗) = 0.85] → 𝑤𝑤 (𝑖𝑖,𝑗𝑗) = 0.05

Here, each value falls in the range [0, 1] and represents a probability of membership of the respective cluster centroid. The sum of the memberships for a given example is equal to 1. As with the k-means algorithm, we can summarize the FCM algorithm in four key steps: 1. Specify the number of k centroids and randomly assign the cluster memberships for each point. 2. Compute the cluster centroids, 𝝁𝝁(𝑗𝑗) , 𝑗𝑗 ∈ {1, … , 𝑘𝑘}.

3. Update the cluster memberships for each point. 4. Repeat steps 2 and 3 until the membership coefficients do not change or a user-defined tolerance or maximum number of iterations is reached.

The objective function of FCM—we abbreviate it as 𝐽𝐽𝑚𝑚—looks very similar to the within-cluster SSE that we minimize in k-means: 𝑛𝑛

𝑘𝑘

𝑚𝑚

2

𝐽𝐽𝑚𝑚 = ∑ ∑ 𝑤𝑤 (𝑖𝑖,𝑗𝑗) ‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(𝑗𝑗) ‖2 𝑖𝑖=1 𝑗𝑗=1

However, note that the membership indicator, 𝑤𝑤 (𝑖𝑖,𝑗𝑗), is not a binary value as in k-means ( 𝑤𝑤 (𝑖𝑖,𝑗𝑗) ∈ {0, 1}), but a real value that denotes the cluster membership probability ( 𝑤𝑤 (𝑖𝑖,𝑗𝑗) ∈ [0, 1]). You also may have noticed that we added an additional exponent to 𝑤𝑤 (𝑖𝑖,𝑗𝑗); the exponent m, any number greater than or equal to one (typically m = 2), is the so-called fuzziness coefficient (or simply fuzzifier), which controls the degree of fuzziness. [ 360 ]

Chapter 11

The larger the value of m, the smaller the cluster membership, 𝑤𝑤 (𝑖𝑖,𝑗𝑗), becomes, which leads to fuzzier clusters. The cluster membership probability itself is calculated as follows: 𝑘𝑘

𝑤𝑤 (𝑖𝑖,𝑗𝑗) = [∑ ( 𝑐𝑐=1

‖𝒙𝒙

(𝑖𝑖)

‖𝒙𝒙(𝑖𝑖)

(𝑗𝑗)

− 𝝁𝝁 ‖2

− 𝝁𝝁(𝑐𝑐) ‖2

)

2 −1 𝑚𝑚−1

]

For example, if we chose three cluster centers, as in the previous k-means example, we could calculate the membership of 𝒙𝒙(𝑖𝑖) belonging to the 𝝁𝝁(𝑗𝑗) cluster as follows:

𝑤𝑤 (𝑖𝑖,𝑗𝑗) = [(

‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(𝑗𝑗) ‖2

‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(1) ‖2

)

2 𝑚𝑚−1

+(

‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(𝑗𝑗) ‖2

‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(2) ‖2

2 𝑚𝑚−1

)

+(

‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(𝑗𝑗) ‖2

‖𝒙𝒙(𝑖𝑖) − 𝝁𝝁(3) ‖2

2 −1 𝑚𝑚−1

)

]

The center, 𝝁𝝁(𝑗𝑗), of a cluster itself is calculated as the mean of all examples weighted 𝑚𝑚 by the degree to which each example belongs to that cluster (𝑤𝑤 (𝑖𝑖,𝑗𝑗) ): (𝑗𝑗)

𝝁𝝁

𝑚𝑚

∑𝑛𝑛𝑖𝑖=1 𝑤𝑤 (𝑖𝑖,𝑗𝑗) 𝒙𝒙(𝑖𝑖) = ∑𝑛𝑛𝑖𝑖=1 𝑤𝑤 (𝑖𝑖,𝑗𝑗)𝑚𝑚

Just by looking at the equation to calculate the cluster memberships, we can say that each iteration in FCM is more expensive than an iteration in k-means. On the other hand, FCM typically requires fewer iterations overall to reach convergence. Unfortunately, the FCM algorithm is currently not implemented in scikit-learn. However, it has been found, in practice, that both k-means and FCM produce very similar clustering outputs, as described in a study (Comparative Analysis of k-means and Fuzzy C-Means Algorithms, S. Ghosh, and S. K. Dubey, IJACSA, 4: 35–38, 2013).

Using the elbow method to find the optimal number of clusters

One of the main challenges in unsupervised learning is that we do not know the definitive answer. We don't have the ground truth class labels in our dataset that allow us to apply the techniques that we used in Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning, in order to evaluate the performance of a supervised model. Thus, to quantify the quality of clustering, we need to use intrinsic metrics—such as the within-cluster SSE (distortion)—to compare the performance of different k-means clusterings. [ 361 ]

Working with Unlabeled Data – Clustering Analysis

Conveniently, we don't need to compute the within-cluster SSE explicitly when we are using scikit-learn, as it is already accessible via the inertia_ attribute after fitting a KMeans model: >>> print('Distortion: %.2f' % km.inertia_) Distortion: 72.48

Based on the within-cluster SSE, we can use a graphical tool, the so-called elbow method, to estimate the optimal number of clusters, k, for a given task. We can say that if k increases, the distortion will decrease. This is because the examples will be closer to the centroids they are assigned to. The idea behind the elbow method is to identify the value of k where the distortion begins to increase most rapidly, which will become clearer if we plot the distortion for different values of k: >>> >>> ... ... ... ... ... ... ... >>> >>> >>> >>> >>>

distortions = [] for i in range(1, 11): km = KMeans(n_clusters=i, init='k-means++', n_init=10, max_iter=300, random_state=0) km.fit(X) distortions.append(km.inertia_) plt.plot(range(1,11), distortions, marker='o') plt.xlabel('Number of clusters') plt.ylabel('Distortion') plt.tight_layout() plt.show()

As you can see in the following plot, the elbow is located at k = 3, so this is evidence that k = 3 is indeed a good choice for this dataset:

[ 362 ]

Chapter 11

Quantifying the quality of clustering via silhouette plots

Another intrinsic metric to evaluate the quality of a clustering is silhouette analysis, which can also be applied to clustering algorithms other than k-means, which we will discuss later in this chapter. Silhouette analysis can be used as a graphical tool to plot a measure of how tightly grouped the examples in the clusters are. To calculate the silhouette coefficient of a single example in our dataset, we can apply the following three steps: 1. Calculate the cluster cohesion, 𝑎𝑎(𝑖𝑖), as the average distance between an example, 𝒙𝒙(𝑖𝑖), and all other points in the same cluster.

2. Calculate the cluster separation, 𝑏𝑏(𝑖𝑖), from the next closest cluster as the average distance between the example, 𝒙𝒙(𝑖𝑖), and all examples in the nearest cluster.

3. Calculate the silhouette, 𝑠𝑠 (𝑖𝑖), as the difference between cluster cohesion and separation divided by the greater of the two, as shown here: 𝑠𝑠 (𝑖𝑖) =

𝑏𝑏(𝑖𝑖) − 𝑎𝑎(𝑖𝑖) 𝑚𝑚𝑚𝑚𝑚𝑚{ 𝑏𝑏(𝑖𝑖) , 𝑎𝑎(𝑖𝑖) }

The silhouette coefficient is bounded in the range –1 to 1. Based on the preceding equation, we can see that the silhouette coefficient is 0 if the cluster separation and cohesion are equal (𝑏𝑏(𝑖𝑖) = 𝑎𝑎(𝑖𝑖)). Furthermore, we get close to an ideal silhouette coefficient of 1 if 𝑏𝑏(𝑖𝑖) ≫ 𝑎𝑎(𝑖𝑖), since 𝑏𝑏(𝑖𝑖) quantifies how dissimilar an example is from other clusters, and 𝑎𝑎(𝑖𝑖) tells us how similar it is to the other examples in its own cluster.

The silhouette coefficient is available as silhouette_samples from scikit-learn's metric module, and optionally, the silhouette_scores function can be imported for convenience. The silhouette_scores function calculates the average silhouette coefficient across all examples, which is equivalent to numpy.mean(silhouette_ samples(...)). By executing the following code, we will now create a plot of the silhouette coefficients for a k-means clustering with k = 3: >>> km = KMeans(n_clusters=3, ... init='k-means++', ... n_init=10, ... max_iter=300, ... tol=1e-04, ... random_state=0) >>> y_km = km.fit_predict(X) >>> import numpy as np

[ 363 ]

Working with Unlabeled Data – Clustering Analysis >>> >>> >>> >>> >>> ... ... >>> >>> >>> ... ... ... ... ... ... ... ... ... ... ... >>> >>> ... ... >>> >>> >>> >>> >>>

from matplotlib import cm from sklearn.metrics import silhouette_samples cluster_labels = np.unique(y_km) n_clusters = cluster_labels.shape[0] silhouette_vals = silhouette_samples(X, y_km, metric='euclidean') y_ax_lower, y_ax_upper = 0, 0 yticks = [] for i, c in enumerate(cluster_labels): c_silhouette_vals = silhouette_vals[y_km == c] c_silhouette_vals.sort() y_ax_upper += len(c_silhouette_vals) color = cm.jet(float(i) / n_clusters) plt.barh(range(y_ax_lower, y_ax_upper), c_silhouette_vals, height=1.0, edgecolor='none', color=color) yticks.append((y_ax_lower + y_ax_upper) / 2.) y_ax_lower += len(c_silhouette_vals) silhouette_avg = np.mean(silhouette_vals) plt.axvline(silhouette_avg, color="red", linestyle="--") plt.yticks(yticks, cluster_labels + 1) plt.ylabel('Cluster') plt.xlabel('Silhouette coefficient') plt.tight_layout() plt.show()

Through a visual inspection of the silhouette plot, we can quickly scrutinize the sizes of the different clusters and identify clusters that contain outliers:

[ 364 ]

Chapter 11

However, as you can see in the preceding silhouette plot, the silhouette coefficients are not even close to 0, which is, in this case, an indicator of a good clustering. Furthermore, to summarize the goodness of our clustering, we added the average silhouette coefficient to the plot (dotted line). To see what a silhouette plot looks like for a relatively bad clustering, let's seed the k-means algorithm with only two centroids: >>> km = KMeans(n_clusters=2, ... init='k-means++', ... n_init=10, ... max_iter=300, ... tol=1e-04, ... random_state=0) >>> y_km = km.fit_predict(X) >>> ... ... ... ... ... >>> ... ... ... ... ... ... >>> ... ... ... ... ... >>> >>> >>> >>>

plt.scatter(X[y_km == 0, 0], X[y_km == 0, 1], s=50, c='lightgreen', edgecolor='black', marker='s', label='Cluster 1') plt.scatter(X[y_km == 1, 0], X[y_km == 1, 1], s=50, c='orange', edgecolor='black', marker='o', label='Cluster 2') plt.scatter(km.cluster_centers_[:, 0], km.cluster_centers_[:, 1], s=250, marker='*', c='red', label='Centroids') plt.legend() plt.grid() plt.tight_layout() plt.show()

As you can see in the resulting plot, one of the centroids falls between two of the three spherical groupings of the input data.

[ 365 ]

Working with Unlabeled Data – Clustering Analysis

Although the clustering does not look completely terrible, it is suboptimal:

Please keep in mind that we typically do not have the luxury of visualizing datasets in two-dimensional scatterplots in real-world problems, since we typically work with data in higher dimensions. So, next, we will create the silhouette plot to evaluate the results: >>> >>> >>> ... ... >>> >>> >>> ... ... ... ... ... ... ... ... ... ... ... >>> >>> >>>

cluster_labels = np.unique(y_km) n_clusters = cluster_labels.shape[0] silhouette_vals = silhouette_samples(X, y_km, metric='euclidean') y_ax_lower, y_ax_upper = 0, 0 yticks = [] for i, c in enumerate(cluster_labels): c_silhouette_vals = silhouette_vals[y_km == c] c_silhouette_vals.sort() y_ax_upper += len(c_silhouette_vals) color = cm.jet(float(i) / n_clusters) plt.barh(range(y_ax_lower, y_ax_upper), c_silhouette_vals, height=1.0, edgecolor='none', color=color) yticks.append((y_ax_lower + y_ax_upper) / 2.) y_ax_lower += len(c_silhouette_vals) silhouette_avg = np.mean(silhouette_vals) plt.axvline(silhouette_avg, color="red", linestyle="--") plt.yticks(yticks, cluster_labels + 1) [ 366 ]

Chapter 11 >>> >>> >>> >>>

plt.ylabel('Cluster') plt.xlabel('Silhouette coefficient') plt.tight_layout() plt.show()

As you can see in the resulting plot, the silhouettes now have visibly different lengths and widths, which is evidence of a relatively bad or at least suboptimal clustering:

Organizing clusters as a hierarchical tree In this section, we will take a look at an alternative approach to prototype-based clustering: hierarchical clustering. One advantage of the hierarchical clustering algorithm is that it allows us to plot dendrograms (visualizations of a binary hierarchical clustering), which can help with the interpretation of the results by creating meaningful taxonomies. Another advantage of this hierarchical approach is that we do not need to specify the number of clusters upfront.

The two main approaches to hierarchical clustering are agglomerative and divisive hierarchical clustering. In divisive hierarchical clustering, we start with one cluster that encompasses the complete dataset, and we iteratively split the cluster into smaller clusters until each cluster only contains one example. In this section, we will focus on agglomerative clustering, which takes the opposite approach. We start with each example as an individual cluster and merge the closest pairs of clusters until only one cluster remains.

[ 367 ]

Working with Unlabeled Data – Clustering Analysis

Grouping clusters in bottom-up fashion

The two standard algorithms for agglomerative hierarchical clustering are single linkage and complete linkage. Using single linkage, we compute the distances between the most similar members for each pair of clusters and merge the two clusters for which the distance between the most similar members is the smallest. The complete linkage approach is similar to single linkage but, instead of comparing the most similar members in each pair of clusters, we compare the most dissimilar members to perform the merge. This is shown in the following diagram:

Alternative types of linkages Other commonly used algorithms for agglomerative hierarchical clustering include average linkage and Ward's linkage. In average linkage, we merge the cluster pairs based on the minimum average distances between all group members in the two clusters. In Ward's linkage, the two clusters that lead to the minimum increase of the total within-cluster SSE are merged.

In this section, we will focus on agglomerative clustering using the complete linkage approach. Hierarchical complete linkage clustering is an iterative procedure that can be summarized by the following steps: 1. Compute the distance matrix of all examples. 2. Represent each data point as a singleton cluster. 3. Merge the two closest clusters based on the distance between the most dissimilar (distant) members. 4. Update the similarity matrix. 5. Repeat steps 2-4 until one single cluster remains. [ 368 ]

Chapter 11

Next, we will discuss how to compute the distance matrix (step 1). But first, let's generate a random data sample to work with. The rows represent different observations (IDs 0-4), and the columns are the different features (X, Y, Z) of those examples: >>> >>> >>> >>> >>> >>> >>> >>>

import pandas as pd import numpy as np np.random.seed(123) variables = ['X', 'Y', 'Z'] labels = ['ID_0', 'ID_1', 'ID_2', 'ID_3', 'ID_4'] X = np.random.random_sample([5, 3])*10 df = pd.DataFrame(X, columns=variables, index=labels) df

After executing the preceding code, we should now see the following data frame containing the randomly generated examples:

Performing hierarchical clustering on a distance matrix

To calculate the distance matrix as input for the hierarchical clustering algorithm, we will use the pdist function from SciPy's spatial.distance submodule: >>> from scipy.spatial.distance import pdist, squareform >>> row_dist = pd.DataFrame(squareform( ... pdist(df, metric='euclidean')), ... columns=labels, index=labels) >>> row_dist

Using the preceding code, we calculated the Euclidean distance between each pair of input examples in our dataset based on the features X, Y, and Z. [ 369 ]

Working with Unlabeled Data – Clustering Analysis

We provided the condensed distance matrix—returned by pdist—as input to the squareform function to create a symmetrical matrix of the pair-wise distances, as shown here:

Next, we will apply the complete linkage agglomeration to our clusters using the linkage function from SciPy's cluster.hierarchy submodule, which returns a socalled linkage matrix. However, before we call the linkage function, let's take a careful look at the function documentation: >>> from scipy.cluster.hierarchy import linkage >>> help(linkage) [...] Parameters: y : ndarray A condensed or redundant distance matrix. A condensed distance matrix is a flat array containing the upper triangular of the distance matrix. This is the form that pdist returns. Alternatively, a collection of m observation vectors in n dimensions may be passed as an m by n array. method : str, optional The linkage algorithm to use. See the Linkage Methods section below for full descriptions. metric : str, optional The distance metric to use. See the distance.pdist function for a list of valid distance metrics. Returns: Z : ndarray The hierarchical clustering encoded as a linkage matrix. [...] [ 370 ]

Chapter 11

Based on the function description, we understand that we can use a condensed distance matrix (upper triangular) from the pdist function as an input attribute. Alternatively, we could also provide the initial data array and use the 'euclidean' metric as a function argument in linkage. However, we should not use the squareform distance matrix that we defined earlier, since it would yield different distance values than expected. To sum it up, the three possible scenarios are listed here: • Incorrect approach: Using the squareform distance matrix as shown in the following code snippet leads to incorrect results: >>> row_clusters = linkage(row_dist, ... method='complete', ... metric='euclidean')

• Correct approach: Using the condensed distance matrix as shown in the following code example yields the correct linkage matrix: >>> row_clusters = linkage(pdist(df, metric='euclidean'), ... method='complete')

• Correct approach: Using the complete input example matrix (the so-called design matrix) as shown in the following code snippet also leads to a correct linkage matrix similar to the preceding approach: >>> row_clusters = linkage(df.values, ... method='complete', ... metric='euclidean')

To take a closer look at the clustering results, we can turn those results into a pandas DataFrame (best viewed in a Jupyter Notebook) as follows: >>> pd.DataFrame(row_clusters, ... columns=['row label 1', ... 'row label 2', ... 'distance', ... 'no. of items in clust.'], ... index=['cluster %d' % (i + 1) for i in ... range(row_clusters.shape[0])])

As shown in the following screenshot, the linkage matrix consists of several rows where each row represents one merge. The first and second columns denote the most dissimilar members in each cluster, and the third column reports the distance between those members.

[ 371 ]

Working with Unlabeled Data – Clustering Analysis

The last column returns the count of the members in each cluster:

Now that we have computed the linkage matrix, we can visualize the results in the form of a dendrogram: >>> from scipy.cluster.hierarchy import dendrogram >>> # make dendrogram black (part 1/2) >>> # from scipy.cluster.hierarchy import set_link_color_palette >>> # set_link_color_palette(['black']) >>> row_dendr = dendrogram(row_clusters, ... labels=labels, ... # make dendrogram black (part 2/2) ... # color_threshold=np.inf ... ) >>> plt.tight_layout() >>> plt.ylabel('Euclidean distance') >>> plt.show()

If you are executing the preceding code or reading an e-book version of this book, you will notice that the branches in the resulting dendrogram are shown in different colors. The color scheme is derived from a list of Matplotlib colors that are cycled for the distance thresholds in the dendrogram. For example, to display the dendrograms in black, you can uncomment the respective sections that were inserted in the preceding code:

[ 372 ]

Chapter 11

Such a dendrogram summarizes the different clusters that were formed during the agglomerative hierarchical clustering; for example, you can see that the examples ID_0 and ID_4, followed by ID_1 and ID_2, are the most similar ones based on the Euclidean distance metric.

Attaching dendrograms to a heat map

In practical applications, hierarchical clustering dendrograms are often used in combination with a heat map, which allows us to represent the individual values in the data array or matrix containing our training examples with a color code. In this section, we will discuss how to attach a dendrogram to a heat map plot and order the rows in the heat map correspondingly. However, attaching a dendrogram to a heat map can be a little bit tricky, so let's go through this procedure step by step: 1. We create a new figure object and define the x axis position, y axis position, width, and height of the dendrogram via the add_axes attribute. Furthermore, we rotate the dendrogram 90 degrees counter-clockwise. The code is as follows: >>> >>> >>> ... >>> >>>

fig = plt.figure(figsize=(8, 8), facecolor='white') axd = fig.add_axes([0.09, 0.1, 0.2, 0.6]) row_dendr = dendrogram(row_clusters, orientation='left') # note: for matplotlib < v1.5.1, please use # orientation='right'

[ 373 ]

Working with Unlabeled Data – Clustering Analysis

2. Next, we reorder the data in our initial DataFrame according to the clustering labels that can be accessed from the dendrogram object, which is essentially a Python dictionary, via the leaves key. The code is as follows: >>> df_rowclust = df.iloc[row_dendr['leaves'][::-1]]

3. Now, we construct the heat map from the reordered DataFrame and position it next to the dendrogram: >>> axm = fig.add_axes([0.23, 0.1, 0.6, 0.6]) >>> cax = axm.matshow(df_rowclust, ... interpolation='nearest', ... cmap='hot_r')

4. Finally, we modify the aesthetics of the dendrogram by removing the axis ticks and hiding the axis spines. Also, we add a color bar and assign the feature and data record names to the x and y axis tick labels, respectively: >>> >>> >>> ... >>> >>> >>> >>>

axd.set_xticks([]) axd.set_yticks([]) for i in axd.spines.values(): i.set_visible(False) fig.colorbar(cax) axm.set_xticklabels([''] + list(df_rowclust.columns)) axm.set_yticklabels([''] + list(df_rowclust.index)) plt.show()

After following the previous steps, the heat map should be displayed with the dendrogram attached:

[ 374 ]

Chapter 11

As you can see, the order of rows in the heat map reflects the clustering of the examples in the dendrogram. In addition to a simple dendrogram, the color-coded values of each example and feature in the heat map provide us with a nice summary of the dataset.

Applying agglomerative clustering via scikitlearn In the previous subsection, you saw how to perform agglomerative hierarchical clustering using SciPy. However, there is also an AgglomerativeClustering implementation in scikit-learn, which allows us to choose the number of clusters that we want to return. This is useful if we want to prune the hierarchical cluster tree. By setting the n_cluster parameter to 3, we will now cluster the input examples into three groups using the same complete linkage approach based on the Euclidean distance metric, as before: >>> from sklearn.cluster import AgglomerativeClustering >>> ac = AgglomerativeClustering(n_clusters=3, ... affinity='euclidean', ... linkage='complete') >>> labels = ac.fit_predict(X) >>> print('Cluster labels: %s' % labels) Cluster labels: [1 0 0 2 1]

Looking at the predicted cluster labels, we can see that the first and the fifth examples (ID_0 and ID_4) were assigned to one cluster (label 1), and the examples ID_1 and ID_2 were assigned to a second cluster (label 0). The example ID_3 was put into its own cluster (label 2). Overall, the results are consistent with the results that we observed in the dendrogram. We should note, though, that ID_3 is more similar to ID_4 and ID_0 than to ID_1 and ID_2, as shown in the preceding dendrogram figure; this is not clear from scikit-learn's clustering results. Let's now rerun the AgglomerativeClustering using n_cluster=2 in the following code snippet: >>> ac = AgglomerativeClustering(n_clusters=2, ... affinity='euclidean', ... linkage='complete') >>> labels = ac.fit_predict(X) >>> print('Cluster labels: %s' % labels) Cluster labels: [0 1 1 0 0]

As you can see, in this pruned clustering hierarchy, label ID_3 was assigned to the same cluster as ID_0 and ID_4, as expected.

[ 375 ]

Working with Unlabeled Data – Clustering Analysis

Locating regions of high density via DBSCAN

Although we can't cover the vast amount of different clustering algorithms in this chapter, let's at least include one more approach to clustering: density-based spatial clustering of applications with noise (DBSCAN), which does not make assumptions about spherical clusters like k-means, nor does it partition the dataset into hierarchies that require a manual cut-off point. As its name implies, density-based clustering assigns cluster labels based on dense regions of points. In DBSCAN, the notion of density is defined as the number of points within a specified radius, 𝜀𝜀. According to the DBSCAN algorithm, a special label is assigned to each example (data point) using the following criteria: • A point is considered a core point if at least a specified number (MinPts) of neighboring points fall within the specified radius, 𝜀𝜀.

• A border point is a point that has fewer neighbors than MinPts within ε, but lies within the 𝜀𝜀 radius of a core point.

• All other points that are neither core nor border points are considered noise points.

After labeling the points as core, border, or noise, the DBSCAN algorithm can be summarized in two simple steps: 1. Form a separate cluster for each core point or connected group of core points. (Core points are connected if they are no farther away than 𝜀𝜀.) 2. Assign each border point to the cluster of its corresponding core point.

[ 376 ]

Chapter 11

To get a better understanding of what the result of DBSCAN can look like, before jumping to the implementation, let's summarize what we have just learned about core points, border points, and noise points in the following figure:

One of the main advantages of using DBSCAN is that it does not assume that the clusters have a spherical shape as in k-means. Furthermore, DBSCAN is different from k-means and hierarchical clustering in that it doesn't necessarily assign each point to a cluster but is capable of removing noise points. For a more illustrative example, let's create a new dataset of half-moon-shaped structures to compare k-means clustering, hierarchical clustering, and DBSCAN: >>> >>> ... ... >>> >>> >>>

from sklearn.datasets import make_moons X, y = make_moons(n_samples=200, noise=0.05, random_state=0) plt.scatter(X[:, 0], X[:, 1]) plt.tight_layout() plt.show()

[ 377 ]

Working with Unlabeled Data – Clustering Analysis

As you can see in the resulting plot, there are two visible, half-moon-shaped groups consisting of 100 examples (data points) each:

We will start by using the k-means algorithm and complete linkage clustering to see if one of those previously discussed clustering algorithms can successfully identify the half-moon shapes as separate clusters. The code is as follows: >>> >>> ... >>> >>> ... ... ... ... ... ... >>> ... ... ... ... ... ... >>> >>> ... ... >>>

f, (ax1, ax2) = plt.subplots(1, 2, figsize=(8, 3)) km = KMeans(n_clusters=2, random_state=0) y_km = km.fit_predict(X) ax1.scatter(X[y_km == 0, 0], X[y_km == 0, 1], c='lightblue', edgecolor='black', marker='o', s=40, label='cluster 1') ax1.scatter(X[y_km == 1, 0], X[y_km == 1, 1], c='red', edgecolor='black', marker='s', s=40, label='cluster 2') ax1.set_title('K-means clustering') ac = AgglomerativeClustering(n_clusters=2, affinity='euclidean', linkage='complete') y_ac = ac.fit_predict(X) [ 378 ]

Chapter 11 >>> ... ... ... ... ... ... >>> ... ... ... ... ... ... >>> >>> >>> >>>

ax2.scatter(X[y_ac == 0, 0], X[y_ac == 0, 1], c='lightblue', edgecolor='black', marker='o', s=40, label='Cluster 1') ax2.scatter(X[y_ac == 1, 0], X[y_ac == 1, 1], c='red', edgecolor='black', marker='s', s=40, label='Cluster 2') ax2.set_title('Agglomerative clustering') plt.legend() plt.tight_layout() plt.show()

Based on the visualized clustering results, we can see that the k-means algorithm was unable to separate the two clusters, and also, the hierarchical clustering algorithm was challenged by those complex shapes:

Finally, let's try the DBSCAN algorithm on this dataset to see if it can find the two half-moon-shaped clusters using a density-based approach: >>> >>> ... ... >>> >>> ...

from sklearn.cluster import DBSCAN db = DBSCAN(eps=0.2, min_samples=5, metric='euclidean') y_db = db.fit_predict(X) plt.scatter(X[y_db == 0, 0], X[y_db == 0, 1], [ 379 ]

Working with Unlabeled Data – Clustering Analysis ... ... ... ... ... >>> ... ... ... ... ... ... >>> >>> >>>

c='lightblue', edgecolor='black', marker='o', s=40, label='Cluster 1') plt.scatter(X[y_db == 1, 0], X[y_db == 1, 1], c='red', edgecolor='black', marker='s', s=40, label='Cluster 2') plt.legend() plt.tight_layout() plt.show()

The DBSCAN algorithm can successfully detect the half-moon shapes, which highlights one of the strengths of DBSCAN – clustering data of arbitrary shapes:

However, we should also note some of the disadvantages of DBSCAN. With an increasing number of features in our dataset—assuming a fixed number of training examples—the negative effect of the curse of dimensionality increases. This is especially a problem if we are using the Euclidean distance metric. However, the problem of the curse of dimensionality is not unique to DBSCAN: it also affects other clustering algorithms that use the Euclidean distance metric, for example, k-means and hierarchical clustering algorithms. In addition, we have two hyperparameters in DBSCAN (MinPts and 𝜀𝜀) that need to be optimized to yield good clustering results. Finding a good combination of MinPts and 𝜀𝜀 can be problematic if the density differences in the dataset are relatively large. [ 380 ]

Chapter 11

Graph-based clustering So far, we have seen three of the most fundamental categories of clustering algorithms: prototype-based clustering with k-means, agglomerative hierarchical clustering, and densitybased clustering via DBSCAN. However, there is also a fourth class of more advanced clustering algorithms that we have not covered in this chapter: graph-based clustering. Probably the most prominent members of the graph-based clustering family are the spectral clustering algorithms. Although there are many different implementations of spectral clustering, what they all have in common is that they use the eigenvectors of a similarity or distance matrix to derive the cluster relationships. Since spectral clustering is beyond the scope of this book, you can read the excellent tutorial by Ulrike von Luxburg to learn more about this topic (A tutorial on spectral clustering, U. Von Luxburg, Statistics and Computing, 17(4): 395– 416, 2007). It is freely available from arXiv at http://arxiv. org/pdf/0711.0189v1.pdf.

Note that, in practice, it is not always obvious which clustering algorithm will perform best on a given dataset, especially if the data comes in multiple dimensions that make it hard or impossible to visualize. Furthermore, it is important to emphasize that a successful clustering does not only depend on the algorithm and its hyperparameters; rather, the choice of an appropriate distance metric and the use of domain knowledge that can help to guide the experimental setup can be even more important. In the context of the curse of dimensionality, it is thus common practice to apply dimensionality reduction techniques prior to performing clustering. Such dimensionality reduction techniques for unsupervised datasets include principal component analysis and radial basis function kernel principal component analysis, which we covered in Chapter 5, Compressing Data via Dimensionality Reduction. Also, it is particularly common to compress datasets down to two-dimensional subspaces, which allows us to visualize the clusters and assigned labels using two-dimensional scatterplots, which are particularly helpful for evaluating the results.

[ 381 ]

Working with Unlabeled Data – Clustering Analysis

Summary

In this chapter, you learned about three different clustering algorithms that can help us with the discovery of hidden structures or information in data. We started this chapter with a prototype-based approach, k-means, which clusters examples into spherical shapes based on a specified number of cluster centroids. Since clustering is an unsupervised method, we do not enjoy the luxury of ground truth labels to evaluate the performance of a model. Thus, we used intrinsic performance metrics, such as the elbow method or silhouette analysis, as an attempt to quantify the quality of clustering. We then looked at a different approach to clustering: agglomerative hierarchical clustering. Hierarchical clustering does not require specifying the number of clusters upfront, and the result can be visualized in a dendrogram representation, which can help with the interpretation of the results. The last clustering algorithm that we covered in this chapter was DBSCAN, an algorithm that groups points based on local densities and is capable of handling outliers and identifying non-globular shapes. After this excursion into the field of unsupervised learning, it is now about time to introduce some of the most exciting machine learning algorithms for supervised learning: multilayer artificial neural networks. After their recent resurgence, neural networks are once again the hottest topic in machine learning research. Thanks to recently developed deep learning algorithms, neural networks are considered state-of-the-art for many complex tasks such as image classification and speech recognition. In Chapter 12, Implementing a Multilayer Artificial Neural Network from Scratch, we will construct our own multilayer neural network. In Chapter 13, Parallelizing Neural Network Training with TensorFlow, we will work with the TensorFlow library, which specializes in training neural network models with multiple layers very efficiently by utilizing graphics processing units.

[ 382 ]

Implementing a Multilayer Artificial Neural Network from Scratch As you may know, deep learning is getting a lot of attention from the press and is without any doubt the hottest topic in the machine learning field. Deep learning can be understood as a subfield of machine learning that is concerned with training artificial neural networks (NNs) with many layers efficiently. In this chapter, you will learn the basic concepts of artificial NNs so that you are well equipped for the following chapters, which will introduce advanced Python-based deep learning libraries and deep neural network (DNN) architectures that are particularly well suited for image and text analyses. The topics that we will cover in this chapter are as follows: • Gaining a conceptual understanding of multilayer NNs • Implementing the fundamental backpropagation algorithm for NN training from scratch • Training a basic multilayer NN for image classification

Modeling complex functions with artificial neural networks

At the beginning of this book, we started our journey through machine learning algorithms with artificial neurons in Chapter 2, Training Simple Machine Learning Algorithms for Classification. Artificial neurons represent the building blocks of the multilayer artificial NNs that we will discuss in this chapter.

[ 383 ]

Implementing a Multilayer Artificial Neural Network from Scratch

The basic concept behind artificial NNs was built upon hypotheses and models of how the human brain works to solve complex problem tasks. Although artificial NNs have gained a lot of popularity in recent years, early studies of NNs go back to the 1940s when Warren McCulloch and Walter Pitts first described how neurons could work. (A logical calculus of the ideas immanent in nervous activity, W. S. McCulloch and W. Pitts. The Bulletin of Mathematical Biophysics, 5(4):115–133, 1943.) However, in the decades that followed the first implementation of the McCullochPitts neuron model—Rosenblatt's perceptron in the 1950s — many researchers and machine learning practitioners slowly began to lose interest in NNs since no one had a good solution for training an NN with multiple layers. Eventually, interest in NNs was rekindled in 1986 when D.E. Rumelhart, G.E. Hinton, and R.J. Williams were involved in the (re)discovery and popularization of the backpropagation algorithm to train NNs more efficiently, which we will discuss in more detail later in this chapter (Learning representations by back-propagating errors, D. E. Rumelhart, G. E. Hinton, R. J. Williams, Nature, 323 (6088): 533–536, 1986). Readers who are interested in the history of artificial intelligence (AI), machine learning, and NNs are also encouraged to read the Wikipedia article on the so-called AI winters, which are the periods of time where a large portion of the research community lost interest in the study of NNs (https://en.wikipedia.org/wiki/AI_winter). However, NNs are more popular today than ever thanks to the many major breakthroughs that have been made in the previous decade, which resulted in what we now call deep learning algorithms and architectures—NNs that are composed of many layers. NNs are a hot topic not only in academic research but also in big technology companies, such as Facebook, Microsoft, Amazon, Uber, and Google, that invest heavily in artificial NNs and deep learning research. As of today, complex NNs powered by deep learning algorithms are considered the state-of-the-art solutions for complex problem solving such as image and voice recognition. Popular examples of the products in our everyday life that are powered by deep learning are Google's image search and Google Translate—an application for smartphones that can automatically recognize text in images for real-time translation into more than 20 languages. Many exciting applications of DNNs have been developed at major tech companies and the pharmaceutical industry as listed in the following, non-comprehensive list of examples:

[ 384 ]

Chapter 12

• Facebook's DeepFace for tagging images (DeepFace: Closing the Gap to HumanLevel Performance in Face Verification, Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1701–1708, 2014) • Baidu's DeepSpeech, which is able to handle voice queries in Mandarin (DeepSpeech: Scaling up end-to-end speech recognition, A. Hannun, C. Case, J. Casper, B. Catanzaro, G. Diamos, E. Elsen, R. Prenger, S. Satheesh, S. Sengupta, A. Coates, and Andrew Y. Ng, arXiv preprint arXiv:1412.5567, 2014) • Google's new language translation service (Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation, arXiv preprint arXiv:1412.5567, 2016) • Novel techniques for drug discovery and toxicity prediction (Toxicity prediction using Deep Learning, T. Unterthiner, A. Mayr, G. Klambauer, and S. Hochreiter, arXiv preprint arXiv:1503.01445, 2015) • A mobile application that can detect skin cancer with an accuracy similar to professionally trained dermatologists (Dermatologist-level classification of skin cancer with deep neural networks, A. Esteva, B.Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S.Thrun, in Nature 542, no. 7639, 2017, pages 115-118) • Protein 3D structure prediction from gene sequences (De novo structure prediction with deep-learning based scoring, R. Evans, J. Jumper, J. Kirkpatrick, L. Sifre, T.F.G. Green, C. Qin, A. Zidek, A. Nelson, A. Bridgland, H. Penedones, S. Petersen, K. Simonyan, S. Crossan, D.T. Jones, D. Silver, K. Kavukcuoglu, D. Hassabis, and A.W. Senior, in Thirteenth Critical Assessment of Techniques for Protein Structure Prediction, 1-4 December, 2018) • Learning how to drive in dense traffic from purely observational data such as camera video streams (Model-predictive policy learning with uncertainty regularization for driving in dense traffic, M. Henaff, A. Canziani, Y. LeCun, 2019, in Conference Proceedings of the International Conference on Learning Representations, ICLR, 2019)

Single-layer neural network recap

This chapter is all about multilayer NNs, how they work, and how to train them to solve complex problems. However, before we dig deeper into a particular multilayer NN architecture, let's briefly reiterate some of the concepts of single-layer NNs that we introduced in Chapter 2, Training Simple Machine Learning Algorithms for Classification, namely, the ADAptive LInear NEuron (Adaline) algorithm, which is shown in the following figure:

[ 385 ]

Implementing a Multilayer Artificial Neural Network from Scratch

In Chapter 2, Training Simple Machine Learning Algorithms for Classification, we implemented the Adaline algorithm to perform binary classification, and we used the gradient descent optimization algorithm to learn the weight coefficients of the model. In every epoch (pass over the training dataset), we updated the weight vector w using the following update rule:

𝒘𝒘 ∶= 𝒘𝒘 + ∆𝒘𝒘,

where ∆𝒘𝒘 = −𝜂𝜂𝜂𝜂𝜂𝜂(𝒘𝒘)

In other words, we computed the gradient based on the whole training dataset and updated the weights of the model by taking a step into the opposite direction of the gradient ∇J(𝒘𝒘). In order to find the optimal weights of the model, we optimized an objective function that we defined as the sum of squared errors (SSE) cost function J(𝒘𝒘) . Furthermore, we multiplied the gradient by a factor, the learning rate 𝜂𝜂 , which we had to choose carefully to balance the speed of learning against the risk of overshooting the global minimum of the cost function. In gradient descent optimization, we updated all weights simultaneously after each epoch, and we defined the partial derivative for each weight 𝑤𝑤𝑗𝑗 in the weight vector w as follows:

𝜕𝜕 (𝑖𝑖) 𝐽𝐽(𝒘𝒘) = − ∑(𝑦𝑦 (𝑖𝑖) − 𝑎𝑎(𝑖𝑖) ) 𝑥𝑥𝑗𝑗 𝜕𝜕𝑤𝑤𝑗𝑗 𝑖𝑖

Here, 𝑦𝑦 (𝑖𝑖) is the target class label of a particular sample 𝑥𝑥 (𝑖𝑖) , and 𝑎𝑎(𝑖𝑖) is the activation of the neuron, which is a linear function in the special case of Adaline.

[ 386 ]

Chapter 12

Furthermore, we defined the activation function 𝜙𝜙(∙) as follows:

𝜙𝜙(𝑧𝑧) = 𝑧𝑧 = 𝑎𝑎

Here, the net input, z, is a linear combination of the weights that are connecting the input layer to the output layer:

𝑧𝑧 = ∑ 𝑤𝑤𝑗𝑗 𝑥𝑥𝑗𝑗 = 𝒘𝒘𝑇𝑇 𝒙𝒙 𝑗𝑗

While we used the activation 𝜙𝜙(𝑧𝑧) to compute the gradient update, we implemented a threshold function to squash the continuous valued output into binary class labels for prediction:

1 if 𝑔𝑔(𝑧𝑧) ≥ 0; 𝑦𝑦̂ = { −1 otherwise

Single-layer naming convention

Note that although Adaline consists of two layers, one input layer and one output layer, it is called a single-layer network because of its single link between the input and output layers.

Also, we learned about a certain trick to accelerate the model learning, the so-called stochastic gradient descent (SGD) optimization. SGD approximates the cost from a single training sample (online learning) or a small subset of training examples (mini-batch learning). We will make use of this concept later in this chapter when we implement and train a multilayer perceptron (MLP). Apart from faster learning— due to the more frequent weight updates compared to gradient descent—its noisy nature is also regarded as beneficial when training multilayer NNs with nonlinear activation functions, which do not have a convex cost function. Here, the added noise can help to escape local cost minima, but we will discuss this topic in more detail later in this chapter.

Introducing the multilayer neural network architecture

In this section, you will learn how to connect multiple single neurons to a multilayer feedforward NN; this special type of fully connected network is also called MLP.

[ 387 ]

Implementing a Multilayer Artificial Neural Network from Scratch

The following figure illustrates the concept of an MLP consisting of three layers:

The MLP depicted in the preceding figure has one input layer, one hidden layer, and one output layer. The units in the hidden layer are fully connected to the input layer, and the output layer is fully connected to the hidden layer. If such a network has more than one hidden layer, we also call it a deep artificial NN. Adding additional hidden layers We can add any number of hidden layers to the MLP to create deeper network architectures. Practically, we can think of the number of layers and units in an NN as additional hyperparameters that we want to optimize for a given problem task using cross-validation technique, which we discussed in Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning. However, the error gradients, which we will calculate later via backpropagation, will become increasingly small as more layers are added to a network. This vanishing gradient problem makes the model learning more challenging. Therefore, special algorithms have been developed to help train such DNN structures; this is known as deep learning.

[ 388 ]

Chapter 12

As shown in the preceding figure, we denote the ith activation unit in the lth layer (𝑙𝑙) as 𝑎𝑎𝑖𝑖 . To make the math and code implementations a bit more intuitive, we will not use numerical indices to refer to layers, but we will use the in superscript for the input layer, the h superscript for the hidden layer, and the out superscript for the output layer. For instance, 𝑎𝑎𝑖𝑖(𝑖𝑖𝑖𝑖) refers to the ith value in the input layer, 𝑎𝑎𝑖𝑖(ℎ) refers to the ith unit in the hidden layer, and 𝑎𝑎𝑖𝑖(𝑜𝑜𝑜𝑜𝑜𝑜) refers to the ith unit in the output layer. (𝑖𝑖𝑖𝑖) Here, the activation units 𝑎𝑎0 and 𝑎𝑎0(ℎ) are the bias units, which we set equal to 1. The activation of the units in the input layer is just its input plus the bias unit:

𝒂𝒂(𝑖𝑖𝑖𝑖) =

(𝑖𝑖𝑖𝑖)

𝑎𝑎0

(𝑖𝑖𝑖𝑖) 𝑎𝑎1

⋮

(𝑖𝑖𝑖𝑖) [𝑎𝑎𝑚𝑚 ]

=

1

(𝑖𝑖𝑖𝑖) 𝑥𝑥1

⋮

(𝑖𝑖𝑖𝑖) [𝑥𝑥𝑚𝑚 ]

Notational convention for the bias units Later in this chapter, we will implement an MLP using separate vectors for the bias units, which makes code implementation more efficient and easier to read. This concept is also used by TensorFlow, a deep learning library that we will cover in Chapter 13, Parallelizing Neural Network Training with TensorFlow. However, the mathematical equations that will follow would appear more complex or convoluted if we had to work with additional variables for the bias. Note that the computation via appending 1s to the input vector (as shown previously) and using a weight variable as bias is exactly the same as operating with separate bias vectors; it is merely a different convention.

Each unit in layer l is connected to all units in layer l + 1 via a weight coefficient. For example, the connection between the kth unit in layer l to the jth unit in layer (𝑙𝑙) l + 1 will be written as 𝑤𝑤𝑘𝑘,𝑗𝑗 . Referring back to the previous figure, we denote the weight matrix that connects the input to the hidden layer as 𝑾𝑾(ℎ), and we write the matrix that connects the hidden layer to the output layer as 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) .

While one unit in the output layer would suffice for a binary classification task, we saw a more general form of NN in the preceding figure, which allows us to perform multiclass classification via a generalization of the one-versus-all (OvA) technique. To better understand how this works, remember the one-hot representation of categorical variables that we introduced in Chapter 4, Building Good Training Datasets – Data Preprocessing.

[ 389 ]

Implementing a Multilayer Artificial Neural Network from Scratch

For example, we can encode the three class labels in the familiar Iris dataset (0=Setosa, 1=Versicolor, 2=Virginica) as follows:

1 0 0 0 = [0] , 1 = [1] , 2 = [0] 0 0 1

This one-hot vector representation allows us to tackle classification tasks with an arbitrary number of unique class labels present in the training dataset. If you are new to NN representations, the indexing notation (subscripts and superscripts) may look a little bit confusing at first. What may seem overly complicated at first will make much more sense in later sections when we vectorize the NN representation. As introduced earlier, we summarize the weights that connect the input and hidden layers by a matrix 𝑾𝑾(ℎ) ∈ ℝ𝑚𝑚×𝑑𝑑 , where d is the number of hidden units and m is the number of input units including the bias unit. Since it is important to internalize this notation to follow the concepts later in this chapter, let's summarize what we have just learned in a descriptive illustration of a simplified 3-4-3 MLP:

[ 390 ]

Chapter 12

Activating a neural network via forward propagation

In this section, we will describe the process of forward propagation to calculate the output of an MLP model. To understand how it fits into the context of learning an MLP model, let's summarize the MLP learning procedure in three simple steps: 1. Starting at the input layer, we forward propagate the patterns of the training data through the network to generate an output. 2. Based on the network's output, we calculate the error that we want to minimize using a cost function that we will describe later. 3. We backpropagate the error, find its derivative with respect to each weight in the network, and update the model. Finally, after we repeat these three steps for multiple epochs and learn the weights of the MLP, we use forward propagation to calculate the network output and apply a threshold function to obtain the predicted class labels in the one-hot representation, which we described in the previous section. Now, let's walk through the individual steps of forward propagation to generate an output from the patterns in the training data. Since each unit in the hidden layer is connected to all units in the input layers, we first calculate the activation unit of the (ℎ) hidden layer 𝑎𝑎1 as follows: (ℎ)

𝑧𝑧1

(ℎ)

(ℎ)

𝑎𝑎1

(𝑖𝑖𝑖𝑖)

(ℎ)

(𝑖𝑖𝑖𝑖)

(ℎ)

(𝑖𝑖𝑖𝑖)

(ℎ)

= 𝑎𝑎0 𝑤𝑤0,1 +𝑎𝑎1 𝑤𝑤1,1 + ⋯ +𝑎𝑎𝑚𝑚 𝑤𝑤𝑚𝑚,1 (ℎ)

= 𝜙𝜙 (𝑧𝑧1 )

Here, 𝑧𝑧1 is the net input and 𝜙𝜙(∙) is the activation function, which has to be differentiable to learn the weights that connect the neurons using a gradient-based approach. To be able to solve complex problems such as image classification, we need nonlinear activation functions in our MLP model, for example, the sigmoid (logistic) activation function that we remember from the section about logistic regression in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn:

𝜙𝜙(𝑧𝑧) =

1 1 + 𝑒𝑒 −𝑧𝑧

[ 391 ]

Implementing a Multilayer Artificial Neural Network from Scratch

As you may recall, the sigmoid function is an S-shaped curve that maps the net input z onto a logistic distribution in the range 0 to 1, which cuts the y-axis at z = 0, as shown in the following graph:

MLP is a typical example of a feedforward artificial NN. The term feedforward refers to the fact that each layer serves as the input to the next layer without loops, in contrast to recurrent NNs—an architecture that we will discuss later in this chapter and discuss in more detail in Chapter 16, Modeling Sequential Data Using Recurrent Neural Networks. The term multilayer perceptron may sound a little bit confusing since the artificial neurons in this network architecture are typically sigmoid units, not perceptrons. We can think of the neurons in the MLP as logistic regression units that return values in the continuous range between 0 and 1. For purposes of code efficiency and readability, we will now write the activation in a more compact form using the concepts of basic linear algebra, which will allow us to vectorize our code implementation via NumPy rather than writing multiple nested and computationally expensive Python for loops:

𝒛𝒛(ℎ) = 𝒂𝒂(𝑖𝑖𝑖𝑖) 𝑾𝑾(ℎ) 𝒂𝒂(ℎ) = 𝜙𝜙(𝒛𝒛(ℎ) )

Here, 𝒂𝒂(𝑖𝑖𝑖𝑖) is our 1 × 𝑚𝑚 dimensional feature vector of a sample 𝒙𝒙(𝑖𝑖𝑖𝑖) plus a bias unit.

𝑾𝑾(ℎ) is an 𝑚𝑚 × 𝑑𝑑 dimensional weight matrix where d is the number of units in the hidden layer. After matrix-vector multiplication, we obtain the 1 × 𝑑𝑑 dimensional net input vector 𝒛𝒛(ℎ) to calculate the activation 𝒂𝒂(ℎ) (where 𝒂𝒂(ℎ) 𝜖𝜖ℝ1×𝑑𝑑). [ 392 ]

Chapter 12

Furthermore, we can generalize this computation to all n examples in the training dataset:

𝒁𝒁(ℎ) = 𝑨𝑨(𝑖𝑖𝑖𝑖) 𝑾𝑾(ℎ)

Here, 𝑨𝑨(𝑖𝑖𝑖𝑖) is now an 𝑛𝑛 × 𝑚𝑚 matrix, and the matrix-matrix multiplication will result in an 𝑛𝑛 × 𝑑𝑑 dimensional net input matrix 𝒁𝒁(ℎ). Finally, we apply the activation function 𝜙𝜙(∙) to each value in the net input matrix to get the 𝑛𝑛 × 𝑑𝑑 activation matrix the next layer (here, the output layer):

𝑨𝑨(ℎ) = 𝜙𝜙(𝒁𝒁(ℎ) )

Similarly, we can write the activation of the output layer in vectorized form for multiple examples:

𝒁𝒁(𝑜𝑜𝑜𝑜𝑜𝑜) = 𝑨𝑨(ℎ) 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜)

Here, we multiply the 𝑑𝑑 × 𝑡𝑡 matrix 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) (t is the number of output units) by the 𝑛𝑛 × 𝑑𝑑 dimensional matrix 𝑨𝑨(ℎ) to obtain the 𝑛𝑛 × 𝑡𝑡 dimensional matrix 𝒁𝒁(𝑜𝑜𝑜𝑜𝑜𝑜) (the columns in this matrix represent the outputs for each sample).

Lastly, we apply the sigmoid activation function to obtain the continuous valued output of our network:

𝑨𝑨(𝑜𝑜𝑜𝑜𝑜𝑜) = 𝜙𝜙(𝒁𝒁(𝑜𝑜𝑜𝑜𝑜𝑜) ),

𝑨𝑨(𝑜𝑜𝑜𝑜𝑜𝑜) ∈ ℝ𝑛𝑛×𝑡𝑡

Classifying handwritten digits

In the previous section, we covered a lot of the theory around NNs, which can be a little bit overwhelming if you are new to this topic. Before we continue with the discussion of the algorithm for learning the weights of the MLP model, backpropagation, let's take a short break from the theory and see an NN in action. Additional resources on backpropagation The NN theory can be quite complex; thus, it is recommended that you refer to two additional resources, which cover some of the concepts that we discuss in this chapter in more detail: • Chapter 6, Deep Feedforward Networks, Deep Learning, I. Goodfellow, Y. Bengio, and A. Courville, MIT Press, 2016 (Manuscripts freely accessible at http://www. deeplearningbook.org).

[ 393 ]

Implementing a Multilayer Artificial Neural Network from Scratch

• Pattern Recognition and Machine Learning, C. M. Bishop and others, Volume 1. Springer New York, 2006. • Lecture slides from the deep learning course at the University of Wisconsin–Madison:

1. https://sebastianraschka.com/pdf/ lecture-notes/stat479ss19/L08_ logistic_slides.pdf

2. https://sebastianraschka.com/pdf/

lecture-notes/stat479ss19/L09_mlp_ slides.pdf

In this section, we will implement and train our first multilayer NN to classify handwritten digits from the popular Mixed National Institute of Standards and Technology (MNIST) dataset that has been constructed by Yann LeCun and others and serves as a popular benchmark dataset for machine learning algorithms (Gradient-Based Learning Applied to Document Recognition, Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, Proceedings of the IEEE, 86(11): 2278-2324, November 1998).

Obtaining and preparing the MNIST dataset

The MNIST dataset is publicly available at http://yann.lecun.com/exdb/mnist/ and consists of the following four parts: • Training dataset images: train-images-idx3-ubyte.gz (9.9 MB, 47 MB unzipped, and 60,000 examples) • Training dataset labels: train-labels-idx1-ubyte.gz (29 KB, 60 KB unzipped, and 60,000 labels) • Test dataset images: t10k-images-idx3-ubyte.gz (1.6 MB, 7.8 MB unzipped, and 10,000 examples) • Test dataset labels: t10k-labels-idx1-ubyte.gz (5 KB, 10 KB unzipped, and 10,000 labels) The MNIST dataset was constructed from two datasets of the US National Institute of Standards and Technology (NIST). The training dataset consists of handwritten digits from 250 different people, 50 percent high school students and 50 percent employees from the Census Bureau. Note that the test dataset contains handwritten digits from different people following the same split.

[ 394 ]

Chapter 12

After you download the files, it is recommended that you unzip them using the Unix/Linux gzip tool from the terminal for efficiency, using the following command in your local MNIST download directory: gzip *ubyte.gz -d

Alternatively, you can use your favorite unzipping tool if you are working with a machine running Microsoft Windows. The images are stored in byte format, and we will read them into NumPy arrays that we will use to train and test our MLP implementation. In order to do that, we will define the following helper function: import os import struct import numpy as np def load_mnist(path, kind='train'): """Load MNIST data from 'path'""" labels_path = os.path.join(path, '%s-labels-idx1-ubyte' % kind) images_path = os.path.join(path, '%s-images-idx3-ubyte' % kind) with open(labels_path, 'rb') as lbpath: magic, n = struct.unpack('>II', lbpath.read(8)) labels = np.fromfile(lbpath, dtype=np.uint8) with open(images_path, 'rb') as imgpath: magic, num, rows, cols = struct.unpack(">IIII", imgpath.read(16)) images = np.fromfile(imgpath, dtype=np.uint8).reshape( len(labels), 784) images = ((images / 255.) - .5) * 2 return images, labels

The load_mnist function returns two arrays, the first being an 𝑛𝑛 × 𝑚𝑚 dimensional NumPy array (images), where n is the number of examples and m is the number of features (here, pixels). The training dataset consists of 60,000 training digits and the test dataset contains 10,000 examples, respectively.

[ 395 ]

Implementing a Multilayer Artificial Neural Network from Scratch

The images in the MNIST dataset consist of 28 × 28 pixels, and each pixel is represented by a grayscale intensity value. Here, we unroll the 28 × 28 pixels into one-dimensional row vectors, which represent the rows in our images array (784 per row or image). The second array (labels) returned by the load_mnist function contains the corresponding target variable, the class labels (integers 0-9) of the handwritten digits. The way we read in the image might seem a little bit strange at first: magic, n = struct.unpack('>II', lbpath.read(8)) labels = np.fromfile(lbpath, dtype=np.uint8)

To understand how those two lines of code work, let's take a look at the dataset description from the MNIST website: [offset] [type] 0000 32 bit integer 0004 32 bit integer 0008 unsigned byte 0009 unsigned byte ........ xxxx unsigned byte

[value] 0x00000801(2049) 60000 ?? ??

[description] magic number (MSB first) number of items label label

??

label

Using the two preceding lines of code, we first read in the magic number, which is a description of the file protocol, as well as the number of items (n) from the file buffer, before we load the following bytes into a NumPy array using the fromfile method. The fmt parameter value, '>II', that we passed as an argument to struct. unpack can be composed into the two following parts: • >: This is big-endian—it defines the order in which a sequence of bytes is stored; if you are unfamiliar with the terms big-endian and littleendian, you can find an excellent article about Endianness on Wikipedia: https://en.wikipedia.org/wiki/Endianness

• I: This is an unsigned integer Finally, we also normalized the pixels values in MNIST to the range –1 to 1 (originally 0 to 255) via the following code line: images = ((images / 255.) - .5) * 2

The reason behind this is that gradient-based optimization is much more stable under these conditions as discussed in Chapter 2, Training Simple Machine Learning Algorithms for Classification. Note that we scaled the images on a pixel-by-pixel basis, which is different from the feature scaling approach that we took in previous chapters. [ 396 ]

Chapter 12

Previously, we derived scaling parameters from the training dataset and used these to scale each column in the training dataset and test dataset. However, when working with image pixels, centering them at zero and rescaling them to a [–1, 1] range is also common and usually works well in practice. Batch normalization A commonly used trick for improving convergence in gradient-based optimization through input scaling is batch normalization, which is an advanced topic that we will cover in Chapter 17, Generative Adversarial Networks for Synthesizing New Data. Also, you can read more about batch normalization in the excellent research article Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift by Sergey Ioffe and Christian Szegedy (2015, https://arxiv.org/ abs/1502.03167).

By executing the following code, we will now load the 60,000 training instances as well as the 10,000 test examples from the local directory where we unzipped the MNIST dataset. (In the following code snippet, it is assumed that the downloaded MNIST files were unzipped to the same directory in which this code was executed.) >>> X_train, y_train = load_mnist('', kind='train') >>> print('Rows: %d, columns: %d' ... % (X_train.shape[0], X_train.shape[1])) Rows: 60000, columns: 784 >>> X_test, y_test = load_mnist('', kind='t10k') >>> print('Rows: %d, columns: %d' ... % (X_test.shape[0], X_test.shape[1])) Rows: 10000, columns: 784

To get an idea of how those images in MNIST look, let's visualize examples of the digits 0-9 after reshaping the 784-pixel vectors from our feature matrix into the original 28 × 28 image that we can plot via Matplotlib's imshow function: >>> import matplotlib.pyplot as plt

>>> fig, ax = plt.subplots(nrows=2, ncols=5, ... sharex=True, sharey=True) >>> ax = ax.flatten() >>> for i in range(10): ... img = X_train[y_train == i][0].reshape(28, 28) ... ax[i].imshow(img, cmap='Greys') >>> ax[0].set_xticks([]) [ 397 ]

Implementing a Multilayer Artificial Neural Network from Scratch >>> ax[0].set_yticks([]) >>> plt.tight_layout() >>> plt.show()

We should now see a plot of the 2 × 5 subfigures showing a representative image of each unique digit:

In addition, let's also plot multiple examples of the same digit to see how different the handwriting for each really is: >>> ... ... ... >>> >>> ... ... >>> >>> >>> >>>

fig, ax = plt.subplots(nrows=5, ncols=5, sharex=True, sharey=True) ax = ax.flatten() for i in range(25): img = X_train[y_train == 7][i].reshape(28, 28) ax[i].imshow(img, cmap='Greys') ax[0].set_xticks([]) ax[0].set_yticks([]) plt.tight_layout() plt.show()

After executing the code, we should now see the first 25 variants of the digit 7:

[ 398 ]

Chapter 12

After we've gone through all the previous steps, it is a good idea to save the scaled images in a format that we can load more quickly into a new Python session to avoid the overhead of reading in and processing the data again. When we are working with NumPy arrays, an efficient and convenient method to save multidimensional arrays to disk is NumPy's savez function. (The official documentation can be found here: https://docs.scipy.org/doc/numpy/reference/generated/numpy. savez.html.) In short, the savez function is analogous to Python's pickle module, which we used in Chapter 9, Embedding a Machine Learning Model into a Web Application, but optimized for storing NumPy arrays. The savez function creates zipped archives of our data, producing .npz files that contain files in the .npy format; if you want to learn more about this format, you can find a nice explanation, including a discussion about advantages and disadvantages, in the NumPy documentation: https://docs.scipy. org/doc/numpy/neps/npy-format.html. Furthermore, instead of using savez, we will use savez_compressed, which uses the same syntax as savez, but further compresses the output file down to substantially smaller file sizes (approximately 22 MB versus approximately 400 MB in this case). The following code snippet will save both the training and test datasets to the archive file mnist_scaled.npz: >>> import numpy as np >>> np.savez_compressed('mnist_scaled.npz', ... X_train=X_train, ... y_train=y_train, ... X_test=X_test, ... y_test=y_test)

After we create the .npz files, we can load the preprocessed MNIST image arrays using NumPy's load function as follows: >>> mnist = np.load('mnist_scaled.npz')

The mnist variable now refers to an object that can access the four data arrays that we provided as keyword arguments to the savez_compressed function. These input arrays are now listed under the files attribute list of the mnist object: >>> mnist.files ['X_train', 'y_train', 'X_test', 'y_test']

For instance, to load the training data into our current Python session, we will access the X_train array as follows (similar to a Python dictionary): >>> X_train = mnist['X_train']

[ 399 ]

Implementing a Multilayer Artificial Neural Network from Scratch

Using a list comprehension, we can retrieve all four data arrays as follows: >>> X_train, y_train, X_test, y_test = [mnist[f] for ... f in mnist.files]

Note that while the preceding np.savez_compressed and np.load examples are not essential for executing the code in this chapter, they serve as a demonstration of how to save and load NumPy arrays conveniently and efficiently. Loading MNIST using scikit-learn Using scikit-learn's new fetch_openml function, it is now also possible to load the MNIST dataset more conveniently. For example, you can use the following code to create a 50,000-example training dataset and and a 10,000-example test dataset by fetching the dataset from https://www.openml. org/d/554: >>> from sklearn.datasets import fetch_openml >>> from sklearn.model_selection import train_test_ split >>> X, y = fetch_openml('mnist_784', version=1, ... return_X_y=True) >>> y = y.astype(int) >>> X = ((X / 255.) - .5) * 2 >>> X_train, X_test, y_train, y_test =\ ... train_test_split( ... X, y, test_size=10000, ... random_state=123, stratify=y)

Please note that the distribution of MNIST records into training and test datasets will be different from the manual approach outlined in this section. Thus, you will observe slightly different results in the following sections if you load the dataset using the fetch_openml and train_test_split functions.

Implementing a multilayer perceptron

In this subsection, we will now implement an MLP from scratch to classify the images in the MNIST dataset. To keep things simple, we will implement an MLP with only one hidden layer. Since the approach may seem a little bit complicated at first, you are encouraged to download the sample code for this chapter from the Packt Publishing website or from GitHub (https://github.com/rasbt/ python-machine-learning-book-3rd-edition) so that you can view this MLP implementation annotated with comments and syntax highlighting for better readability.

[ 400 ]

Chapter 12

If you are not running the code from the accompanying Jupyter Notebook file or don't have access to the Internet, copy the NeuralNetMLP code from this chapter into a Python script file in your current working directory (for example, neuralnet.py), which you can then import into your current Python session via the following command: from neuralnet import NeuralNetMLP

The code will contain parts that we have not talked about yet, such as the backpropagation algorithm, but most of the code should look familiar to you based on the Adaline implementation in Chapter 2, Training Simple Machine Learning Algorithms for Classification, and the discussion of forward propagation in earlier sections. Do not worry if not all of the code makes immediate sense to you; we will follow up on certain parts later in this chapter. However, going over the code at this stage can make it easier to follow the theory later. The following is the implementation of an MLP: import numpy as np import sys class NeuralNetMLP(object): """ Feedforward neural network / Multi-layer perceptron classifier.

Parameters -----------n_hidden : int (default: 30) Number of hidden units. l2 : float (default: 0.) Lambda value for L2-regularization. No regularization if l2=0. (default) epochs : int (default: 100) Number of passes over the training set. eta : float (default: 0.001) Learning rate. shuffle : bool (default: True) Shuffles training data every epoch if True to prevent circles. minibatch_size : int (default: 1) Number of training examples per minibatch. seed : int (default: None)

[ 401 ]

Implementing a Multilayer Artificial Neural Network from Scratch Random seed for initializing weights and shuffling. Attributes ----------eval_ : dict Dictionary collecting the cost, training accuracy, and validation accuracy for each epoch during training. """ def __init__(self, n_hidden=30, l2=0., epochs=100, eta=0.001, shuffle=True, minibatch_size=1, seed=None): self.random = np.random.RandomState(seed) self.n_hidden = n_hidden self.l2 = l2 self.epochs = epochs self.eta = eta self.shuffle = shuffle self.minibatch_size = minibatch_size def _onehot(self, y, n_classes): """Encode labels into one-hot representation Parameters -----------y : array, shape = [n_examples] Target values. Returns ----------onehot : array, shape = (n_examples, n_labels) """ onehot = np.zeros((n_classes, y.shape[0])) for idx, val in enumerate(y.astype(int)): onehot[val, idx] = 1. return onehot.T def _sigmoid(self, z): """Compute logistic function (sigmoid)""" return 1. / (1. + np.exp(-np.clip(z, -250, 250))) def _forward(self, X):

[ 402 ]

Chapter 12 """Compute forward propagation step""" # step 1: net input of hidden layer # [n_examples, n_features] dot [n_features, n_hidden] # -> [n_examples, n_hidden] z_h = np.dot(X, self.w_h) + self.b_h # step 2: activation of hidden layer a_h = self._sigmoid(z_h) # step 3: net input of output layer # [n_examples, n_hidden] dot [n_hidden, n_classlabels] # -> [n_examples, n_classlabels] z_out = np.dot(a_h, self.w_out) + self.b_out # step 4: activation output layer a_out = self._sigmoid(z_out) return z_h, a_h, z_out, a_out def _compute_cost(self, y_enc, output): """Compute cost function. Parameters ---------y_enc : array, shape = (n_examples, n_labels) one-hot encoded class labels. output : array, shape = [n_examples, n_output_units] Activation of the output layer (forward propagation) Returns --------cost : float Regularized cost """ L2_term = (self.l2 * (np.sum(self.w_h ** 2.) + np.sum(self.w_out ** 2.))) term1 = -y_enc * (np.log(output)) term2 = (1. - y_enc) * np.log(1. - output) cost = np.sum(term1 - term2) + L2_term return cost

[ 403 ]

Implementing a Multilayer Artificial Neural Network from Scratch

def predict(self, X): """Predict class labels Parameters ----------X : array, shape = [n_examples, n_features] Input layer with original features. Returns: ---------y_pred : array, shape = [n_examples] Predicted class labels. """ z_h, a_h, z_out, a_out = self._forward(X) y_pred = np.argmax(z_out, axis=1) return y_pred def fit(self, X_train, y_train, X_valid, y_valid): """ Learn weights from training data. Parameters ----------X_train : array, shape = [n_examples, n_features] Input layer with original features. y_train : array, shape = [n_examples] Target class labels. X_valid : array, shape = [n_examples, n_features] Sample features for validation during training y_valid : array, shape = [n_examples] Sample labels for validation during training Returns: ---------self """ n_output = np.unique(y_train).shape[0] # no. of class #labels n_features = X_train.shape[1] ######################## # Weight initialization [ 404 ]

Chapter 12 ######################## # weights for input -> hidden self.b_h = np.zeros(self.n_hidden) self.w_h = self.random.normal(loc=0.0, scale=0.1, size=(n_features, self.n_hidden)) # weights for hidden -> output self.b_out = np.zeros(n_output) self.w_out = self.random.normal(loc=0.0, scale=0.1, size=(self.n_hidden, n_output)) epoch_strlen = len(str(self.epochs)) # for progr. format. self.eval_ = {'cost': [], 'train_acc': [], 'valid_acc': \ []} y_train_enc = self._onehot(y_train, n_output) # iterate over training epochs for i in range(self.epochs): # iterate over minibatches indices = np.arange(X_train.shape[0]) if self.shuffle: self.random.shuffle(indices) for start_idx in range(0, indices.shape[0] -\ self.minibatch_size +\ 1, self.minibatch_size): batch_idx = indices[start_idx:start_idx +\ self.minibatch_size] # forward propagation z_h, a_h, z_out, a_out = \ self._forward(X_train[batch_idx]) ################## # Backpropagation ################## # [n_examples, n_classlabels] delta_out = a_out - y_train_enc[batch_idx] [ 405 ]

Implementing a Multilayer Artificial Neural Network from Scratch

# [n_examples, n_hidden] sigmoid_derivative_h = a_h * (1. - a_h) # [n_examples, n_classlabels] dot [n_classlabels, # n_hidden] # -> [n_examples, n_hidden] delta_h = (np.dot(delta_out, self.w_out.T) * sigmoid_derivative_h) # [n_features, n_examples] dot [n_examples, # n_hidden] # -> [n_features, n_hidden] grad_w_h = np.dot(X_train[batch_idx].T, delta_h) grad_b_h = np.sum(delta_h, axis=0) # [n_hidden, n_examples] dot [n_examples, # n_classlabels] # -> [n_hidden, n_classlabels] grad_w_out = np.dot(a_h.T, delta_out) grad_b_out = np.sum(delta_out, axis=0) # Regularization and weight updates delta_w_h = (grad_w_h + self.l2*self.w_h) delta_b_h = grad_b_h # bias is not regularized self.w_h -= self.eta * delta_w_h self.b_h -= self.eta * delta_b_h delta_w_out = delta_b_out = self.w_out -= self.b_out -=

(grad_w_out + self.l2*self.w_out) grad_b_out # bias is not regularized self.eta * delta_w_out self.eta * delta_b_out

############# # Evaluation ############# # Evaluation after each epoch during training z_h, a_h, z_out, a_out = self._forward(X_train) cost = self._compute_cost(y_enc=y_train_enc, output=a_out) y_train_pred = self.predict(X_train) y_valid_pred = self.predict(X_valid) [ 406 ]

Chapter 12

train_acc = ((np.sum(y_train == y_train_pred)).astype(np.float) / X_train.shape[0]) valid_acc = ((np.sum(y_valid == y_valid_pred)).astype(np.float) / X_valid.shape[0]) sys.stderr.write('\r%0*d/%d | Cost: %.2f ' '| Train/Valid Acc.: %.2f%%/%.2f%% ' % (epoch_strlen, i+1, self.epochs, cost, train_acc*100, valid_acc*100)) sys.stderr.flush() self.eval_['cost'].append(cost) self.eval_['train_acc'].append(train_acc) self.eval_['valid_acc'].append(valid_acc) return self

After executing this code, we next initialize a new 784-100-10 MLP—an NN with 784 input units (n_features), 100 hidden units (n_hidden), and 10 output units (n_output): >>> nn = NeuralNetMLP(n_hidden=100, ... l2=0.01, ... epochs=200, ... eta=0.0005, ... minibatch_size=100, ... shuffle=True, ... seed=1)

If you read through the NeuralNetMLP code, you've probably already guessed what these parameters are for. Here, you find a short summary of them: • l2: This is the 𝜆𝜆 parameter for L2 regularization to decrease the degree of overfitting. • epochs: This is the number of passes over the training dataset. • eta: This is the learning rate 𝜂𝜂.

• shuffle: This is for shuffling the training set prior to every epoch to prevent the algorithm getting stuck in circles.

• seed: This is a random seed for shuffling and weight initialization.

[ 407 ]

Implementing a Multilayer Artificial Neural Network from Scratch

• minibatch_size: This is the number of training examples in each minibatch when splitting the training data in each epoch for SGD. The gradient is computed for each mini-batch separately instead of the entire training data for faster learning. Next, we train the MLP using 55,000 examples from the already shuffled MNIST training dataset and use the remaining 5,000 examples for validation during training. Note that training the NN may take up to five minutes on standard desktop computer hardware. As you may have noticed from the preceding code, we implemented the fit method so that it takes four input arguments: training images, training labels, validation images, and validation labels. In NN training, it is really useful to compare training and validation accuracy, which helps us judge whether the network model performs well, given the architecture and hyperparameters. For example, if we observe a low training and validation accuracy, there is likely an issue with the training dataset, or the hyperparameters settings are not ideal. A relatively large gap between the training and the validation accuracy indicated that the model is likely overfitting the training dataset so that we want to reduce the number of parameters in the model or increase the regularization strength. If both the training and validation accuracies are high, the model is likely to generalize well to new data, for example, the test dataset, which we use for the final model evaluation. In general, training (deep) NNs is relatively expensive compared with the other models we've discussed so far. Thus, we want to stop it early in certain circumstances and start over with different hyperparameter settings. On the other hand, if we find that it increasingly tends to overfit the training data (noticeable by an increasing gap between training and validation dataset performance), we may want to stop the training early as well. Now, to start the training, we execute the following code: >>> nn.fit(X_train=X_train[:55000], ... y_train=y_train[:55000], ... X_valid=X_train[55000:], ... y_valid=y_train[55000:]) 200/200 | Cost: 5065.78 | Train/Valid Acc.: 99.28%/97.98%

[ 408 ]

Chapter 12

In our NeuralNetMLP implementation, we also defined an eval_ attribute that collects the cost, training, and validation accuracy for each epoch so that we can visualize the results using Matplotlib: >>> >>> >>> >>> >>>

import matplotlib.pyplot as plt plt.plot(range(nn.epochs), nn.eval_['cost']) plt.ylabel('Cost') plt.xlabel('Epochs') plt.show()

The preceding code plots the cost over the 200 epochs, as shown in the following graph:

As we can see, the cost decreased substantially during the first 100 epochs and seems to slowly converge in the last 100 epochs. However, the small slope between epoch 175 and epoch 200 indicates that the cost would further decrease with a training over additional epochs. Next, let's take a look at the training and validation accuracy: >>> ... >>> ... >>> >>> >>> >>>

plt.plot(range(nn.epochs), nn.eval_['train_acc'], label='training') plt.plot(range(nn.epochs), nn.eval_['valid_acc'], label='validation', linestyle='--') plt.ylabel('Accuracy') plt.xlabel('Epochs') plt.legend(loc='lower right') plt.show()

[ 409 ]

Implementing a Multilayer Artificial Neural Network from Scratch

The preceding code examples plot those accuracy values over the 200 training epochs, as shown in the following figure:

The plot reveals that the gap between training and validation accuracy increases as we train for more epochs. At approximately the 50th epoch, the training and validation accuracy values are equal, and then, the network starts overfitting the training data. Note that this example was chosen deliberately to illustrate the effect of overfitting and demonstrate why it is useful to compare the validation and training accuracy values during training. One way to decrease the effect of overfitting is to increase the regularization strength—for example, by setting l2=0.1. Another useful technique to tackle overfitting in NNs is dropout, which will be covered in Chapter 15, Classifying Images with Deep Convolutional Neural Networks. Finally, let's evaluate the generalization performance of the model by calculating the prediction accuracy on the test dataset: >>> y_test_pred = nn.predict(X_test) >>> acc = (np.sum(y_test == y_test_pred) ... .astype(np.float) / X_test.shape[0]) >>> print('Test accuracy: %.2f%%' % (acc * 100)) Test accuracy: 97.54%

Despite the slight overfitting on the training data, our relatively simple one-hiddenlayer NN achieved a relatively good performance on the test dataset, similar to the validation dataset accuracy (97.98 percent).

[ 410 ]

Chapter 12

To further fine-tune the model, we could change the number of hidden units, values of the regularization parameters, and the learning rate, or use various other tricks that have been developed over the years but are beyond the scope of this book. In Chapter 15, Classifying Images with Deep Convolutional Neural Networks, you will learn about a different NN architecture that is known for its good performance on image datasets. Also, the chapter will introduce additional performance-enhancing tricks such as adaptive learning rates, more sophisticated SGD-based optimization algorithms, batch normalization, and dropout. Other common tricks that are beyond the scope of the following chapters include: • Adding skip-connections, which are the main contribution of residual NNs (Deep residual learning for image recognition. K. He, X. Zhang, S. Ren, J. Sun (2016). In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770-778) • Using learning rate schedulers that change the learning rate during training (Cyclical learning rates for training neural networks. L.N. Smith (2017). In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 464-472) • Attaching loss functions to earlier layers in the networks as it's being done in the popular Inception v3 architecture (Rethinking the Inception architecture for computer vision. C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna (2016). In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818-2826) Lastly, let's take a look at some of the images that our MLP struggles with: >>> miscl_img = X_test[y_test != y_test_pred][:25] >>> correct_lab = y_test[y_test != y_test_pred][:25] >>> miscl_lab = y_test_pred[y_test != y_test_pred][:25] >>> fig, ax = plt.subplots(nrows=5, ... ncols=5, ... sharex=True, ... sharey=True,) >>> ax = ax.flatten() >>> for i in range(25): ... img = miscl_img[i].reshape(28, 28) ... ax[i].imshow(img, ... cmap='Greys', ... interpolation='nearest') ... ax[i].set_title('%d) t: %d p: %d' ... % (i+1, correct_lab[i], miscl_lab[i]))

[ 411 ]

Implementing a Multilayer Artificial Neural Network from Scratch >>> >>> >>> >>>

ax[0].set_xticks([]) ax[0].set_yticks([]) plt.tight_layout() plt.show()

We should now see a 5 × 5 subplot matrix where the first number in the subtitles indicates the plot index, the second number represents the true class label (t), and the third number stands for the predicted class label (p):

As we can see in the preceding figure, some of those images are even challenging for us humans to classify correctly. For example, the 6 in subplot 8 really looks like a carelessly drawn 0, and the 8 in subplot 23 could be a 9 due to the narrow lower part combined with the bold line.

Training an artificial neural network

Now that we have seen an NN in action and have gained a basic understanding of how it works by looking over the code, let's dig a little bit deeper into some of the concepts, such as the logistic cost function and the backpropagation algorithm that we implemented to learn the weights.

Computing the logistic cost function

The logistic cost function that we implemented as the _compute_cost method is actually pretty simple to follow since it is the same cost function that we described in the logistic regression section in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn: [ 412 ]

Chapter 12 𝑛𝑛

𝐽𝐽(𝒘𝒘) = − ∑ 𝑦𝑦 [𝑖𝑖] log(𝑎𝑎[𝑖𝑖] ) + (1 − 𝑦𝑦 [𝑖𝑖] ) log(1 − 𝑎𝑎[𝑖𝑖] ) 𝑖𝑖=1

Here, 𝑎𝑎[𝑖𝑖] is the sigmoid activation of the ith sample in the dataset, which we compute in the forward propagation step:

𝑎𝑎[𝑖𝑖] = 𝜙𝜙(𝑧𝑧 [𝑖𝑖] )

Again, note that in this context, the superscript [i] is an index for training examples, not layers. Now, let's add a regularization term, which allows us to reduce the degree of overfitting. As you recall from earlier chapters, the L2 regularization term is defined as follows (remember that we don't regularize the bias units):

𝐿𝐿2 =

𝜆𝜆‖𝒘𝒘‖22

𝑚𝑚

= 𝜆𝜆 ∑ 𝑤𝑤𝑗𝑗2 𝑗𝑗=1

By adding the L2 regularization term to our logistic cost function, we obtain the following equation: 𝑛𝑛

𝜆𝜆 𝐽𝐽(𝒘𝒘) = − [∑ 𝑦𝑦 [𝑖𝑖] log(𝑎𝑎[𝑖𝑖] ) + (1 − 𝑦𝑦 [𝑖𝑖] ) log(1 − 𝑎𝑎[𝑖𝑖] )] + ‖𝒘𝒘‖22 2 𝑖𝑖=1

Previously, we implemented an MLP for multiclass classification that returns an output vector of t elements that we need to compare to the 𝑡𝑡 × 1 dimensional target vector in the one-hot encoding representation. If we predict the class label of an input image with class label 2, using this MLP, the activation of the third layer and the target may look like this:

𝑎𝑎(𝑜𝑜𝑜𝑜𝑜𝑜)

0.1 0.9 = [ ], ⋮ 0.3

0 1 𝑦𝑦 = [ ] ⋮ 0

Thus, we need to generalize the logistic cost function to all t activation units in our network.

[ 413 ]

Implementing a Multilayer Artificial Neural Network from Scratch

The cost function (without the regularization term) becomes the following: 𝑛𝑛

𝑡𝑡

[𝑖𝑖]

[𝑖𝑖]

[𝑖𝑖]

[𝑖𝑖]

𝐽𝐽(𝑾𝑾) = − ∑ ∑ 𝑦𝑦𝑗𝑗 log(𝑎𝑎𝑗𝑗 ) + (1 − 𝑦𝑦𝑗𝑗 ) log(1 − 𝑎𝑎𝑗𝑗 ) 𝑖𝑖=1 𝑗𝑗=1

Here, again, the superscript [i] is the index of a particular sample in our training dataset. The following generalized regularization term may look a little bit complicated at first, but here we are just calculating the sum of all weights of an l layer (without the bias term) that we added to the first column:

𝐽𝐽(𝑾𝑾) =

𝑛𝑛

𝑡𝑡

[𝑖𝑖] [𝑖𝑖] − [∑ ∑ 𝑦𝑦𝑗𝑗 𝑙𝑙𝑙𝑙𝑙𝑙 (𝑎𝑎𝑗𝑗 ) + 𝑖𝑖=1 𝑗𝑗=1

(1 −

[𝑖𝑖] 𝑦𝑦𝑗𝑗 ) 𝑙𝑙𝑙𝑙𝑙𝑙 (1

−

[𝑖𝑖] 𝑎𝑎𝑗𝑗 )]

𝐿𝐿−1 𝑢𝑢𝑙𝑙 𝑢𝑢𝑙𝑙+1

𝜆𝜆 (𝑙𝑙) 2 + ∑ ∑ ∑ (𝑤𝑤𝑗𝑗,𝑖𝑖 ) 2 𝑙𝑙=1 𝑖𝑖=1 𝑗𝑗=1

Here, 𝑢𝑢𝑙𝑙 refers to the number of units in a given layer l, and the following expression represents the penalty term: 𝐿𝐿−1 𝑢𝑢𝑙𝑙 𝑢𝑢𝑙𝑙+1

𝜆𝜆 (𝑙𝑙) 2 ∑ ∑ ∑ (𝑤𝑤𝑗𝑗,𝑖𝑖 ) 2 𝑙𝑙=1 𝑖𝑖=1 𝑗𝑗=1

Remember that our goal is to minimize the cost function J(W); thus, we need to calculate the partial derivative of the parameters W with respect to each weight for every layer in the network:

𝜕𝜕

(𝑙𝑙)

𝜕𝜕𝑤𝑤𝑗𝑗,𝑖𝑖

𝐽𝐽(𝑾𝑾)

In the next section, we will talk about the backpropagation algorithm, which allows us to calculate those partial derivatives to minimize the cost function. Note that W consists of multiple matrices. In an MLP with one hidden layer, we have the weight matrix, 𝑾𝑾(ℎ), which connects the input to the hidden layer, and 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜), which connects the hidden layer to the output layer. A visualization of the threedimensional tensor W is provided in the following figure:

[ 414 ]

Chapter 12

In this simplified figure, it may seem that both 𝑾𝑾(ℎ) and 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) have the same number of rows and columns, which is typically not the case unless we initialize an MLP with the same number of hidden units, output units, and input features. If this sounds confusing, stay tuned for the next section, where we will discuss the dimensionality of 𝑾𝑾(ℎ) and 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) in more detail in the context of the backpropagation algorithm. Also, you are encouraged to read through the code of the NeuralNetMLP again, which is annotated with helpful comments about the dimensionality of the different matrices and vector transformations. You can obtain the annotated code either from Packt or the book's GitHub repository at https:// github.com/rasbt/python-machine-learning-book-3rd-edition.

Developing your understanding of backpropagation

Although backpropagation was rediscovered and popularized more than 30 years ago (Learning representations by back-propagating errors, D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Nature, 323: 6088, pages 533–536, 1986), it still remains one of the most widely used algorithms to train artificial NNs very efficiently. If you are interested in additional references regarding the history of backpropagation, Juergen Schmidhuber wrote a nice survey article, Who Invented Backpropagation?, which you can find online at http://people.idsia.ch/~juergen/who-inventedbackpropagation.html.

[ 415 ]

Implementing a Multilayer Artificial Neural Network from Scratch

This section will provide both a short, clear summary and the bigger picture of how this fascinating algorithm works before we dive into more mathematical details. In essence, we can think of backpropagation as a very computationally efficient approach to compute the partial derivatives of a complex cost function in multilayer NNs. Here, our goal is to use those derivatives to learn the weight coefficients for parameterizing such a multilayer artificial NN. The challenge in the parameterization of NNs. is that we are typically dealing with a very large number of weight coefficients in a high-dimensional feature space. In contrast to cost functions of single-layer NNs such as Adaline or logistic regression, which we have seen in previous chapters, the error surface of an NN cost function is not convex or smooth with respect to the parameters. There are many bumps in this high-dimensional cost surface (local minima) that we have to overcome in order to find the global minimum of the cost function. You may recall the concept of the chain rule from your introductory calculus classes. The chain rule is an approach to compute the derivative of a complex, nested function, such as f ( g(x)), as follows:

𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑 [𝑓𝑓(𝑔𝑔(𝑥𝑥))] = ∙ 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑

Similarly, we can use the chain rule for an arbitrarily long function composition. For example, let's assume that we have five different functions, f(x), g(x), h(x), u(x), and v(x), and let F be the function composition: F(x) = f(g(h(u(v(x))))). Applying the chain rule, we can compute the derivative of this function as follows:

𝑑𝑑𝑑𝑑 𝑑𝑑 𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑ℎ 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 = 𝐹𝐹(𝑥𝑥) = 𝑓𝑓(𝑔𝑔(ℎ(𝑢𝑢(𝑣𝑣(𝑥𝑥))))) = ∙ ∙ ∙ ∙ 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑ℎ 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑

In the context of computer algebra, a set of techniques has been developed to solve such problems very efficiently, which is also known as automatic differentiation. If you are interested in learning more about automatic differentiation in machine learning applications, read A. G. Baydin and B. A. Pearlmutter's article Automatic Differentiation of Algorithms for Machine Learning, arXiv preprint arXiv:1404.7456, 2014, which is freely available on arXiv at http://arxiv.org/pdf/1404.7456.pdf. Automatic differentiation comes with two modes, the forward and reverse modes; backpropagation is simply a special case of reverse-mode automatic differentiation. The key point is that applying the chain rule in the forward mode could be quite expensive since we would have to multiply large matrices for each layer (Jacobians) that we would eventually multiply by a vector to obtain the output.

[ 416 ]

Chapter 12

The trick of reverse mode is that we start from right to left: we multiply a matrix by a vector, which yields another vector that is multiplied by the next matrix and so on. Matrix-vector multiplication is computationally much cheaper than matrix-matrix multiplication, which is why backpropagation is one of the most popular algorithms used in NN training. A basic calculus refresher To fully understand backpropagation, we need to borrow certain concepts from differential calculus, which is outside the scope of this book. However, you can refer to a review chapter of the most fundamental concepts, which you might find useful in this context. It discusses function derivatives, partial derivatives, gradients, and the Jacobian. This text is freely accessible at https://sebastianraschka.com/pdf/ books/dlb/appendix_d_calculus.pdf. If you are unfamiliar with calculus or need a brief refresher, consider reading this text as an additional supporting resource before reading the next section.

Training neural networks via backpropagation In this section, we will go through the math of backpropagation to understand how you can learn the weights in an NN very efficiently. Depending on how comfortable you are with mathematical representations, the following equations may seem relatively complicated at first.

In a previous section, we saw how to calculate the cost as the difference between the activation of the last layer and the target class label. Now, we will see how the backpropagation algorithm works to update the weights in our MLP model from a mathematical perspective, which we implemented after the # Backpropagation code comment inside the fit method. As we recall from the beginning of this chapter, we first need to apply forward propagation in order to obtain the activation of the output layer, which we formulated as follows:

𝒁𝒁(ℎ)

𝑨𝑨(ℎ)

𝒁𝒁(𝑜𝑜𝑜𝑜𝑜𝑜)

(𝑜𝑜𝑜𝑜𝑜𝑜)

𝑨𝑨

= 𝑨𝑨(𝑖𝑖𝑖𝑖) 𝑾𝑾(ℎ) = 𝜙𝜙(𝒁𝒁(ℎ) )

= 𝑨𝑨(ℎ) 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) = 𝜙𝜙(𝒁𝒁

(𝒐𝒐𝒐𝒐𝒐𝒐)

)

(net input of the hidden layer)

(activation of the hidden layer) (net input of the output layer)

(activation of the output layer)

[ 417 ]

Implementing a Multilayer Artificial Neural Network from Scratch

Concisely, we just forward-propagate the input features through the connection in the network, as shown in the following illustration:

In backpropagation, we propagate the error from right to left. We start by calculating the error vector of the output layer:

𝜹𝜹(𝑜𝑜𝑜𝑜𝑜𝑜) = 𝒂𝒂(𝑜𝑜𝑜𝑜𝑜𝑜) − 𝒚𝒚

Here, y is the vector of the true class labels (the corresponding variable in the NeuralNetMLP code is delta_out). Next, we calculate the error term of the hidden layer: 𝑇𝑇

𝜹𝜹(ℎ) = 𝜹𝜹(𝑜𝑜𝑜𝑜𝑜𝑜) (𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) ) ⊙

𝜕𝜕𝜕𝜕(𝑧𝑧 (ℎ) ) 𝜕𝜕𝑧𝑧 (ℎ)

𝜕𝜕𝜕𝜕(𝑧𝑧 (ℎ) ) is simply the derivative of the sigmoid activation function, which 𝜕𝜕𝑧𝑧 (ℎ) we computed as sigmoid_derivative_h = a_h * (1. - a_h) in the fit method of the NeuralNetMLP:

Here,

𝜕𝜕𝜕𝜕(𝑧𝑧) = (𝑎𝑎(ℎ) ⊙ (1 − 𝑎𝑎(ℎ) )) 𝜕𝜕𝜕𝜕

Note that the ⊙ symbol means element-wise multiplication in this context.

[ 418 ]

Chapter 12

Activation function derivative Although it is not important to follow the next equations, you may be curious how the derivative of the activation function was obtained; it is summarized step by step here:

𝜕𝜕 1 ( ) 𝜕𝜕𝜕𝜕 1 + 𝑒𝑒 −𝑧𝑧 𝑒𝑒 −𝑧𝑧 = (1 + 𝑒𝑒 −𝑧𝑧 )2 2 1 + 𝑒𝑒 −𝑧𝑧 1 = − ( ) (1 + 𝑒𝑒 −𝑧𝑧 )2 1 + 𝑒𝑒 −𝑧𝑧 2 1 1 = − ( ) (1 + 𝑒𝑒 −𝑧𝑧 ) 1 + 𝑒𝑒 −𝑧𝑧

𝜙𝜙 ′ (𝑧𝑧) =

2

= 𝜙𝜙(𝑧𝑧) − (𝜙𝜙(𝑧𝑧))

= 𝜙𝜙(𝑧𝑧)(1 − 𝜙𝜙(𝑧𝑧)) = 𝑎𝑎(1 − a)

Next, we compute the 𝛿𝛿 (ℎ) layer error matrix (delta_h) as follows: 𝑇𝑇

𝛿𝛿 (ℎ) = 𝛿𝛿 (𝑜𝑜𝑜𝑜𝑜𝑜) (𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) ) ⊙ (𝑎𝑎(ℎ) ⊙ (1 − 𝑎𝑎(ℎ) ))

To better understand how we computed this 𝛿𝛿 (ℎ) term, let's walk through it in more 𝑇𝑇 detail. In the preceding equation, we used the transpose (𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) ) of the ℎ × 𝑡𝑡-dimensional matrix 𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜). Here, t is the number of output class labels and h is the number of hidden units. The matrix multiplication between the 𝑇𝑇 𝑛𝑛 × 𝑡𝑡-dimensional 𝜹𝜹(𝑜𝑜𝑜𝑜𝑜𝑜) matrix and the 𝑡𝑡 × ℎ-dimensional matrix (𝑾𝑾(𝑜𝑜𝑜𝑜𝑜𝑜) ) results in an 𝑛𝑛 × ℎ-dimensional matrix that we multiplied element-wise by the sigmoid derivative of the same dimension to obtain the 𝑛𝑛 × ℎ-dimensional matrix 𝛿𝛿 (ℎ).

[ 419 ]

Implementing a Multilayer Artificial Neural Network from Scratch

Eventually, after obtaining the 𝛿𝛿 terms, we can now write the derivation of the cost function as follows:

𝜕𝜕

(𝑜𝑜𝑜𝑜𝑜𝑜) 𝜕𝜕𝑤𝑤𝑖𝑖,𝑗𝑗

𝜕𝜕

(ℎ)

𝜕𝜕𝑤𝑤𝑖𝑖,𝑗𝑗

(ℎ) (𝑜𝑜𝑜𝑜𝑜𝑜)

𝐽𝐽(𝑾𝑾) = 𝑎𝑎𝑗𝑗 𝛿𝛿𝑖𝑖

𝐽𝐽(𝑾𝑾)

(𝑖𝑖𝑖𝑖) (ℎ) 𝛿𝛿𝑖𝑖

= 𝑎𝑎𝑗𝑗

Next, we need to accumulate the partial derivative of every node in each layer and the error of the node in the next layer. However, remember that we need to compute (𝑙𝑙) ∆𝑖𝑖,𝑗𝑗 for every sample in the training dataset. Thus, it is easier to implement it as a vectorized version like in our NeuralNetMLP code implementation:

∆(ℎ)

∆(𝑜𝑜𝑜𝑜𝑜𝑜)

𝑇𝑇

= (𝑨𝑨(𝑖𝑖𝑖𝑖) ) 𝛿𝛿 (ℎ) 𝑇𝑇

= (𝑨𝑨(ℎ) ) 𝛿𝛿 (𝑜𝑜𝑜𝑜𝑜𝑜)

And after we have accumulated the partial derivatives, we can add the following regularization term:

∆(𝑙𝑙) ∶= ∆(𝑙𝑙) + 𝜆𝜆(𝑙𝑙) 𝑾𝑾(𝑙𝑙)

(Please note that the bias units are usually not regularized.) The two previous mathematical equations correspond to the code variables delta_w_h, delta_b_h, delta_w_out, and delta_b_out in NeuralNetMLP. Lastly, after we have computed the gradients, we can update the weights by taking an opposite step toward the gradient for each layer l:

This is implemented as follows: self.w_h -= self.eta * self.b_h -= self.eta * self.w_out -= self.eta self.b_out -= self.eta

𝑾𝑾(𝑙𝑙) ∶= 𝑾𝑾(𝑙𝑙) − 𝜂𝜂∆(𝑙𝑙)

delta_w_h delta_b_h * delta_w_out * delta_b_out

[ 420 ]

Chapter 12

To bring everything together, let's summarize backpropagation in the following figure:

About the convergence in neural networks

You might be wondering why we did not use regular gradient descent but instead used mini-batch learning to train our NN for the handwritten digit classification. You may recall our discussion on SGD that we used to implement online learning. In online learning, we compute the gradient based on a single training example (k = 1) at a time to perform the weight update. Although this is a stochastic approach, it often leads to very accurate solutions with a much faster convergence than regular gradient descent. Mini-batch learning is a special form of SGD where we compute the gradient based on a subset k of the n training examples with 1 < k < n. Minibatch learning has the advantage over online learning that we can make use of our vectorized implementations to improve computational efficiency. However, we can update the weights much faster than in regular gradient descent. Intuitively, you can think of mini-batch learning as predicting the voter turnout of a presidential election from a poll by asking only a representative subset of the population rather than asking the entire population (which would be equal to running the actual election).

[ 421 ]

Implementing a Multilayer Artificial Neural Network from Scratch

Multilayer NNs are much harder to train than simpler algorithms such as Adaline, logistic regression, or support vector machines. In multilayer NNs, we typically have hundreds, thousands, or even billions of weights that we need to optimize. Unfortunately, the output function has a rough surface and the optimization algorithm can easily become trapped in local minima, as shown in the following figure:

Note that this representation is extremely simplified since our NN has many dimensions; it makes it impossible to visualize the actual cost surface for the human eye. Here, we only show the cost surface for a single weight on the x-axis. However, the main message is that we do not want our algorithm to get trapped in local minima. By increasing the learning rate, we can more readily escape such local minima. On the other hand, we also increase the chance of overshooting the global optimum if the learning rate is too large. Since we initialize the weights randomly, we start with a solution to the optimization problem that is typically hopelessly wrong.

A few last words about the neural network implementation

You may be wondering why we went through all of this theory just to implement a simple multilayer artificial network that can classify handwritten digits instead of using an open source Python machine learning library. In fact, we will introduce more complex NN models in the next chapters, which we will train using the open source TensorFlow library (https://www.tensorflow.org).

[ 422 ]

Chapter 12

Although the from-scratch implementation in this chapter seems a bit tedious at first, it was a good exercise for understanding the basics behind backpropagation and NN training, and a basic understanding of algorithms is crucial for applying machine learning techniques appropriately and successfully. Now that you have learned how feedforward NNs work, we are ready to explore more sophisticated DNNs by using TensorFlow, which allows us to construct NNs more efficiently, as we will see in Chapter 13, Parallelizing Neural Network Training with TensorFlow. Over the past two years, since its release in November 2015, TensorFlow has gained a lot of popularity among machine learning researchers, who use it to construct DNNs because of its ability to optimize mathematical expressions for computations on multidimensional arrays utilizing graphics processing units (GPUs). While TensorFlow can be considered a low-level deep learning library, simplifying APIs such as Keras have been developed that make the construction of common deep learning models even more convenient, which we will see in Chapter 13, Parallelizing Neural Network Training with TensorFlow.

Summary

In this chapter, you have learned the basic concepts behind multilayer artificial NNs, which are currently the hottest topic in machine learning research. In Chapter 2, Training Simple Machine Learning Algorithms for Classification, we started our journey with simple single-layer NN structures and now we have connected multiple neurons to a powerful NN architecture to solve complex problems such as handwritten digit recognition. We demystified the popular backpropagation algorithm, which is one of the building blocks of many NN models that are used in deep learning. After learning about the backpropagation algorithm in this chapter, we are well equipped for exploring more complex DNN architectures. In the remaining chapters, we will cover TensorFlow, an open source library geared toward deep learning, which allows us to implement and train multilayer NNs more efficiently.

[ 423 ]

Parallelizing Neural Network Training with TensorFlow In this chapter, we will move on from the mathematical foundations of machine learning and deep learning to focus on TensorFlow. TensorFlow is one of the most popular deep learning libraries currently available, and it lets us implement neural networks (NNs) much more efficiently than any of our previous NumPy implementations. In this chapter, we will start using TensorFlow and see how it brings significant benefits to training performance. This chapter will begin the next stage of our journey into machine learning and deep learning, and we will explore the following topics: • How TensorFlow improves training performance • Working with TensorFlow's Dataset API (tf.data) to build input pipelines and efficient model training • Working with TensorFlow to write optimized machine learning code • Using TensorFlow high-level APIs to build a multilayer NN • Choosing activation functions for artificial NNs • Introducing Keras (tf.keras), a high-level wrapper around TensorFlow that can be used to implement common deep learning architectures conveniently

[ 425 ]

Parallelizing Neural Network Training with TensorFlow

TensorFlow and training performance

TensorFlow can speed up our machine learning tasks significantly. To understand how it can do this, let's begin by discussing some of the performance challenges we typically run into when we run expensive calculations on our hardware. Then, we will take a high-level look at what TensorFlow is and what our learning approach will be in this chapter.

Performance challenges

The performance of computer processors has, of course, been continuously improving in recent years, and that allows us to train more powerful and complex learning systems, which means that we can improve the predictive performance of our machine learning models. Even the cheapest desktop computer hardware that's available right now comes with processing units that have multiple cores. In the previous chapters, we saw that many functions in scikit-learn allow us to spread those computations over multiple processing units. However, by default, Python is limited to execution on one core due to the global interpreter lock (GIL). So, although we, indeed, take advantage of Python's multiprocessing library to distribute our computations over multiple cores, we still have to consider that the most advanced desktop hardware rarely comes with more than eight or 16 such cores. You will recall from Chapter 12, Implementing a Multilayer Artificial Neural Network from Scratch, that we implemented a very simple multilayer perceptron (MLP) with only one hidden layer consisting of 100 units. We had to optimize approximately 80,000 weight parameters ([784*100 + 100] + [100 * 10] + 10 = 79,510) to learn a model for a very simple image classification task. The images in MNIST are rather small (28 × 28), and we can only imagine the explosion in the number of parameters if we wanted to add additional hidden layers or work with images that have higher pixel densities. Such a task would quickly become unfeasible for a single processing unit. The question then becomes, how can we tackle such problems more effectively? The obvious solution to this problem is to use graphics processing units (GPUs), which are real work horses. You can think of a graphics card as a small computer cluster inside your machine. Another advantage is that modern GPUs are relatively cheap compared to the state-of-the-art central processing units (CPUs), as you can see in the following overview:

[ 426 ]

Chapter 13

The sources for the information in the table are the following websites (Date: October 2019): • https://ark.intel.com/content/www/us/en/ark/products/189123/ intel-core-i9-9960x-x-series-processor-22m-cache-up-to-4-50ghz.html

• https://www.nvidia.com/en-us/geforce/graphics-cards/rtx-2080ti/

At 65 percent of the price of a modern CPU, we can get a GPU that has 272 times more cores and is capable of around 10 times more floating-point calculations per second. So, what is holding us back from utilizing GPUs for our machine learning tasks? The challenge is that writing code to target GPUs is not as simple as executing Python code in our interpreter. There are special packages, such as CUDA and OpenCL, that allow us to target the GPU. However, writing code in CUDA or OpenCL is probably not the most convenient environment for implementing and running machine learning algorithms. The good news is that this is what TensorFlow was developed for!

What is TensorFlow?

TensorFlow is a scalable and multiplatform programming interface for implementing and running machine learning algorithms, including convenience wrappers for deep learning. TensorFlow was developed by the researchers and engineers from the Google Brain team. While the main development is led by a team of researchers and software engineers at Google, its development also involves many contributions from the open source community. TensorFlow was initially built for internal use at Google, but it was subsequently released in November 2015 under a permissive open source license. Many machine learning researchers and practitioners from academia and industry have adapted TensorFlow to develop deep learning solutions.

[ 427 ]

Parallelizing Neural Network Training with TensorFlow

To improve the performance of training machine learning models, TensorFlow allows execution on both CPUs and GPUs. However, its greatest performance capabilities can be discovered when using GPUs. TensorFlow supports CUDAenabled GPUs officially. Support for OpenCL-enabled devices is still experimental. However, OpenCL will likely be officially supported in the near future. TensorFlow currently supports frontend interfaces for a number of programming languages. Luckily for us as Python users, TensorFlow's Python API is currently the most complete API, thereby it attracts many machine learning and deep learning practitioners. Furthermore, TensorFlow has an official API in C++. In addition, new tools based on TensorFlow have been released, TensorFlow.js and TensorFlow Lite, that focus on running and deploying machine learning models in a web browser and on mobile and Internet of Things (IoT) devices. The APIs in other languages, such as Java, Haskell, Node.js, and Go, are not stable yet, but the open source community and TensorFlow developers are constantly improving them. TensorFlow is built around a computation graph composed of a set of nodes. Each node represents an operation that may have zero or more input or output. A tensor is created as a symbolic handle to refer to the input and output of these operations. Mathematically, tensors can be understood as a generalization of scalars, vectors, matrices, and so on. More concretely, a scalar can be defined as a rank-0 tensor, a vector can be defined as a rank-1 tensor, a matrix can be defined as a rank-2 tensor, and matrices stacked in a third dimension can be defined as rank-3 tensors. But note that in TensorFlow, the values are stored in NumPy arrays, and the tensors provide references to these arrays. To make the concept of a tensor clearer, consider the following figure, which represents tensors of ranks 0 and 1 in the first row, and tensors of ranks 2 and 3 in the second row:

[ 428 ]

Chapter 13

In the original TensorFlow release, TensorFlow computations relied on constructing a static, directed graph to represent the data flow. As the use of static computation graphs proved to be a major friction point for many users, the TensorFlow library recently received a major overhaul with its 2.0 version, which makes building and training NN models a lot simpler. While TensorFlow 2.0 still supports static computation graphs, it now uses dynamic computation graphs, which allows for more flexibility.

How we will learn TensorFlow

First, we are going to cover TensorFlow's programming model, in particular, creating and manipulating tensors. Then, we will see how to load data and utilize TensorFlow Dataset objects, which will allow us to iterate through a dataset efficiently. In addition, we will discuss the existing, ready-to-use datasets in the tensorflow_datasets submodule and learn how to use them. After learning about these basics, the tf.keras API will be introduced and we will move forward to building machine learning models, learn how to compile and train the models, and learn how to save the trained models on disk for future evaluation.

First steps with TensorFlow

In this section, we will take our first steps in using the low-level TensorFlow API. After installing TensorFlow, we will cover how to create tensors in TensorFlow and different ways of manipulating them, such as changing their shape, data type, and so on.

Installing TensorFlow

Depending on how your system is set up, you can typically just use Python's pip installer and install TensorFlow from PyPI by executing the following from your terminal: pip install tensorflow

This will install the latest stable version, which is 2.0.0 at the time of writing. In order to ensure that the code presented in this chapter can be executed as expected, it is recommended that you use TensorFlow 2.0.0, which can be installed by specifying the version explicitly: pip install tensorflow==[desired-version]

[ 429 ]

Parallelizing Neural Network Training with TensorFlow

In case you want to use GPUs (recommended), you need a compatible NVIDIA graphics card, along with the CUDA Toolkit and the NVIDIA cuDNN library to be installed. If your machine satisfies these requirements, you can install TensorFlow with GPU support, as follows: pip install tensorflow-gpu

For more information about the installation and setup process, please see the official recommendations at https://www.tensorflow.org/install/gpu. Note that TensorFlow is still under active development; therefore, every couple of months, newer versions are released with significant changes. At the time of writing this chapter, the latest TensorFlow version is 2.0. You can verify your TensorFlow version from your terminal, as follows: python -c 'import tensorflow as tf; print(tf.__version__)'

Troubleshooting your installation of TensorFlow If you experience problems with the installation procedure, read more about system- and platform-specific recommendations that are provided at https://www.tensorflow.org/install/. Note that all the code in this chapter can be run on your CPU; using a GPU is entirely optional but recommended if you want to fully enjoy the benefits of TensorFlow. For example, while training some NN models on CPU could take a week, the same models could be trained in just a few hours on a modern GPU. If you have a graphics card, refer to the installation page to set it up appropriately. In addition, you may find this TensorFlowGPU setup guide helpful, which explains how to install the NVIDIA graphics card drivers, CUDA, and cuDNN on Ubuntu (not required but recommended requirements for running TensorFlow on a GPU): https://sebastianraschka.com/ pdf/books/dlb/appendix_h_cloud-computing.pdf. Furthermore, as you will see in Chapter 17, Generative Adversarial Networks for Synthesizing New Data, you can also train your models using a GPU for free via Google Colab.

Creating tensors in TensorFlow

Now, let's consider a few different ways of creating tensors, and then see some of their properties and how to manipulate them. Firstly, we can simply create a tensor from a list or a NumPy array using the tf.convert_to_tensor function as follows: >>> import tensorflow as tf [ 430 ]

Chapter 13 >>> import numpy as np >>> np.set_printoptions(precision=3) >>> a = np.array([1, 2, 3], dtype=np.int32) >>> b = [4, 5, 6] >>> t_a = tf.convert_to_tensor(a) >>> t_b = tf.convert_to_tensor(b) >>> print(t_a) >>> print(t_b) tf.Tensor([1 2 3], shape=(3,), dtype=int32) tf.Tensor([4 5 6], shape=(3,), dtype=int32)

This resulted in tensors t_a and t_b, with their properties, shape=(3,) and dtype=int32, adopted from their source. Similar to NumPy arrays, we can further see these properties: >>> t_ones = tf.ones((2, 3)) >>> t_ones.shape TensorShape([2, 3])

To get access to the values that a tensor refers to, we can simply call the .numpy() method on a tensor: >>> t_ones.numpy() array([[1., 1., 1.], [1., 1., 1.]], dtype=float32)

Finally, creating a tensor of constant values can be done as follows: >>> const_tensor = tf.constant([1.2, 5, np.pi], ... dtype=tf.float32) >>> print(const_tensor) tf.Tensor([1.2 5. 3.142], shape=(3,), dtype=float32)

Manipulating the data type and shape of a tensor

Learning ways to manipulate tensors is necessary to make them compatible for input to a model or an operation. In this section, you will learn how to manipulate tensor data types and shapes via several TensorFlow functions that cast, reshape, transpose, and squeeze.

[ 431 ]

Parallelizing Neural Network Training with TensorFlow

The tf.cast() function can be used to change the data type of a tensor to a desired type: >>> t_a_new = tf.cast(t_a, tf.int64) >>> print(t_a_new.dtype)

As you will see in upcoming chapters, certain operations require that the input tensors have a certain number of dimensions (that is, rank) associated with a certain number of elements (shape). Thus, we might need to change the shape of a tensor, add a new dimension, or squeeze an unnecessary dimension. TensorFlow provides useful functions (or operations) to achieve this, such as tf.transpose(), tf.reshape(), and tf.squeeze(). Let's take a look at some examples: • Transposing a tensor: >>> >>> >>> (3,

t = tf.random.uniform(shape=(3, 5)) t_tr = tf.transpose(t) print(t.shape, ' --> ', t_tr.shape) 5) --> (5, 3)

• Reshaping a tensor (for example, from a 1D vector to a 2D array): >>> >>> >>> (5,

t = tf.zeros((30,)) t_reshape = tf.reshape(t, shape=(5, 6)) print(t_reshape.shape) 6)

• Removing the unnecessary dimensions (dimensions that have size 1, which are not needed): >>> >>> >>> (1,

t = tf.zeros((1, 2, 1, 4, 1)) t_sqz = tf.squeeze(t, axis=(2, 4)) print(t.shape, ' --> ', t_sqz.shape) 2, 1, 4, 1) --> (1, 2, 4)

Applying mathematical operations to tensors Applying mathematical operations, in particular linear algebra operations, is necessary for building most machine learning models. In this subsection, we will cover some widely used linear algebra operations, such as element-wise product, matrix multiplication, and computing the norm of a tensor.

First, let's instantiate two random tensors, one with uniform distribution in the range [–1, 1) and the other with a standard normal distribution: >>> tf.random.set_seed(1) [ 432 ]

Chapter 13 >>> t1 = tf.random.uniform(shape=(5, 2), ... minval=-1.0, maxval=1.0) >>> t2 = tf.random.normal(shape=(5, 2), ... mean=0.0, stddev=1.0)

Notice that t1 and t2 have the same shape. Now, to compute the element-wise product of t1 and t2, we can use the following: >>> t3 = tf.multiply(t1, t2).numpy() >>> print(t3) [[-0.27 -0.874] [-0.017 -0.175] [-0.296 -0.139] [-0.727 0.135] [-0.401 0.004]]

To compute the mean, sum, and standard deviation along a certain axis (or axes), we can use tf.math.reduce_mean(), tf.math.reduce_sum(), and tf.math.reduce_ std(). For example, the mean of each column in t1 can be computed as follows: >>> t4 = tf.math.reduce_mean(t1, axis=0) >>> print(t4) tf.Tensor([0.09 0.207], shape=(2,), dtype=float32)

The matrix-matrix product between t1 and t2 (that is, 𝑡𝑡1 × 𝑡𝑡2⊤, where the superscript T is for transpose) can be computed by using the tf.linalg.matmul() function as follows: >>> t5 = tf.linalg.matmul(t1, >>> print(t5.numpy()) [[-1.144 1.115 -0.87 -0.321 [ 0.248 -0.191 0.25 -0.064 [-0.478 0.407 -0.436 0.022 [ 0.525 -0.234 0.741 -0.593 [-0.099 0.26 0.125 -0.462

t2, transpose_b=True) 0.856] -0.331] 0.527] -1.194] -0.396]]

On the other hand, computing 𝑡𝑡1 × 𝑡𝑡2⊤ is performed by transposing t1, resulting in an array of size 2 × 2: >>> t6 = tf.linalg.matmul(t1, t2, transpose_a=True) >>> print(t6.numpy()) [[-1.711 0.302] [ 0.371 -1.049]]

[ 433 ]

Parallelizing Neural Network Training with TensorFlow

Finally, the tf.norm() function is useful for computing the 𝐿𝐿𝑝𝑝 norm of a tensor. For example, we can calculate the 𝐿𝐿2 norm of t1 as follows: >>> norm_t1 = tf.norm(t1, ord=2, axis=1).numpy() >>> print(norm_t1) [1.046 0.293 0.504 0.96 0.383]

To verify that this code snippet computes the 𝐿𝐿2 norm of t1 correctly, you can compare the results with the following NumPy function: np.sqrt(np.sum(np. square(t1), axis=1)).

Split, stack, and concatenate tensors

In this subsection, we will cover TensorFlow operations for splitting a tensor into multiple tensors, or the reverse: stacking and concatenating multiple tensors into a single one. Assume that we have a single tensor and we want to split it into two or more tensors. For this, TensorFlow provides a convenient tf.split() function, which divides an input tensor into a list of equally-sized tensors. We can determine the desired number of splits as an integer using the argument num_or_size_splits to split a tensor along a desired dimension specified by the axis argument. In this case, the total size of the input tensor along the specified dimension must be divisible by the desired number of splits. Alternatively, we can provide the desired sizes in a list. Let's have a look at an example of both these options: • Providing the number of splits (must be divisible): >>> tf.random.set_seed(1) >>> t = tf.random.uniform((6,)) >>> print(t.numpy()) [0.165 0.901 0.631 0.435 0.292 0.643] >>> t_splits = tf.split(t, num_or_size_splits=3) >>> [item.numpy() for item in t_splits] [array([0.165, 0.901], dtype=float32), array([0.631, 0.435], dtype=float32), array([0.292, 0.643], dtype=float32)]

In this example, a tensor of size 6 was divided into a list of three tensors each with size 2. • Providing the sizes of different splits: Alternatively, instead of defining the number of splits, we can also specify the sizes of the output tensors directly. Here, we are splitting a tensor of size 5 into tensors of sizes 3 and 2: [ 434 ]

Chapter 13 >>> tf.random.set_seed(1) >>> t = tf.random.uniform((5,)) >>> print(t.numpy()) [0.165 0.901 0.631 0.435 0.292] >>> t_splits = tf.split(t, num_or_size_splits=[3, 2]) >>> [item.numpy() for item in t_splits] [array([0.165, 0.901, 0.631], dtype=float32), array([0.435, 0.292], dtype=float32)]

Sometimes, we are working with multiple tensors and need to concatenate or stack them to create a single tensor. In this case, TensorFlow functions such as tf.stack() and tf.concat() come in handy. For example, let's create a 1D tensor, A, containing 1s with size 3 and a 1D tensor, B, containing 0s with size 2 and concatenate them into a 1D tensor, C, of size 5: >>> >>> >>> >>> [1.

A = tf.ones((3,)) B = tf.zeros((2,)) C = tf.concat([A, B], axis=0) print(C.numpy()) 1. 1. 0. 0.]

If we create 1D tensors A and B, both with size 3, then we can stack them together to form a 2D tensor, S: >>> A = tf.ones((3,)) >>> B = tf.zeros((3,)) >>> S = tf.stack([A, B], axis=1) >>> print(S.numpy()) [[1. 0.] [1. 0.] [1. 0.]]

The TensorFlow API has many operations that you can use for building a model, processing your data, and more. However, covering every function is outside the scope of this book, where we will focus on the most essential ones. For the full list of operations and functions, you can refer to the documentation page of TensorFlow at https://www.tensorflow.org/versions/r2.0/api_docs/python/tf.

Building input pipelines using tf.data – the TensorFlow Dataset API

When we are training a deep NN model, we usually train the model incrementally using an iterative optimization algorithm such as stochastic gradient descent, as we have seen in previous chapters. [ 435 ]

Parallelizing Neural Network Training with TensorFlow

As mentioned at the beginning of this chapter, the Keras API is a wrapper around TensorFlow for building NN models. The Keras API provides a method, .fit(), for training the models. In cases where the training dataset is rather small and can be loaded as a tensor into the memory, TensorFlow models (that are built with the Keras API) can directly use this tensor via their .fit() method for training. In typical use cases, however, when the dataset is too large to fit into the computer memory, we will need to load the data from the main storage device (for example, the hard drive or solid-state drive) in chunks, that is, batch by batch (note the use of the term "batch" instead of "mini-batch" in this chapter to stay close to the TensorFlow terminology). In addition, we may need to construct a data-processing pipeline to apply certain transformations and preprocessing steps to our data, such as mean centering, scaling, or adding noise to augment the training procedure and to prevent overfitting. Applying preprocessing functions manually every time can be quite cumbersome. Luckily, TensorFlow provides a special class for constructing efficient and convenient preprocessing pipelines. In this section, we will see an overview of different methods for constructing a TensorFlow Dataset, including dataset transformations and common preprocessing steps.

Creating a TensorFlow Dataset from existing tensors

If the data already exists in the form of a tensor object, a Python list, or a NumPy array, we can easily create a dataset using the tf.data.Dataset.from_tensor_ slices() function. This function returns an object of class Dataset, which we can use to iterate through the individual elements in the input dataset. As a simple example, consider the following code, which creates a dataset from a list of values: >>> a = [1.2, 3.4, 7.5, 4.1, 5.0, 1.0] >>> ds = tf.data.Dataset.from_tensor_slices(a) >>> print(ds)

We can easily iterate through a dataset entry by entry as follows: >>> for item in ds: ... print(item) tf.Tensor(1.2, shape=(), tf.Tensor(3.4, shape=(), tf.Tensor(7.5, shape=(), tf.Tensor(4.1, shape=(), tf.Tensor(5.0, shape=(), tf.Tensor(1.0, shape=(),

dtype=float32) dtype=float32) dtype=float32) dtype=float32) dtype=float32) dtype=float32) [ 436 ]

Chapter 13

If we want to create batches from this dataset, with a desired batch size of 3, we can do this as follows: >>> ds_batch = ds.batch(3) >>> for i, elem in enumerate(ds_batch, 1): ... print('batch {}:'.format(i), elem.numpy()) batch 1: [1.2 3.4 7.5] batch 2: [4.1 5. 1. ]

This will create two batches from this dataset, where the first three elements go into batch #1, and the remaining elements go into batch #2. The .batch() method has an optional argument, drop_remainder, which is useful for cases when the number of elements in the tensor is not divisible by the desired batch size. The default for drop_ remainder is False. We will see more examples illustrating the behavior of this method later in the subsection Shuffle, batch, and repeat.

Combining two tensors into a joint dataset

Often, we may have the data in two (or possibly more) tensors. For example, we could have a tensor for features and a tensor for labels. In such cases, we need to build a dataset that combines these tensors together, which will allow us to retrieve the elements of these tensors in tuples. Assume that we have two tensors, t_x and t_y. Tensor t_x holds our feature values, each of size 3, and t_y stores the class labels. For this example, we first create these two tensors as follows: >>> tf.random.set_seed(1) >>> t_x = tf.random.uniform([4, 3], dtype=tf.float32) >>> t_y = tf.range(4)

Now, we want to create a joint dataset from these two tensors. Note that there is a required one-to-one correspondence between the elements of these two tensors: >>> >>> >>> >>> >>> ... ...

ds_x = tf.data.Dataset.from_tensor_slices(t_x) ds_y = tf.data.Dataset.from_tensor_slices(t_y)

x: x: x: x:

ds_joint = tf.data.Dataset.zip((ds_x, ds_y)) for example in ds_joint: print(' x:', example[0].numpy(), ' y:', example[1].numpy()) [0.165 [0.435 [0.976 [0.605

0.901 0.292 0.435 0.637

0.631] 0.643] 0.66 ] 0.614]

y: y: y: y:

0 1 2 3 [ 437 ]

Parallelizing Neural Network Training with TensorFlow

Here, we first created two separate datasets, namely ds_x and ds_y. We then used the zip function to form a joint dataset. Alternatively, we can create the joint dataset using tf.data.Dataset.from_tensor_slices() as follows: >>> ds_joint = tf.data.Dataset.from_tensor_slices((t_x, t_y)) >>> for example in ds_joint: ... print(' x:', example[0].numpy(), ... ' y:', example[1].numpy()) x: x: x: x:

[0.165 [0.435 [0.976 [0.605

0.901 0.292 0.435 0.637

0.631] 0.643] 0.66 ] 0.614]

y: y: y: y:

0 1 2 3

which results in the same output. Note that a common source of error could be that the element-wise correspondence between the original features (x) and labels (y) might be lost (for example, if the two datasets are shuffled separately). However, once they are merged into one dataset, it is safe to apply these operations. Next, we will see how to apply transformations to each individual element of a dataset. For this, we will use the previous ds_joint dataset and apply featurescaling to scale the values to the range [-1, 1), as currently the values of t_x are in the range [0, 1) based on a random uniform distribution: >>> ds_trans = ds_joint.map(lambda x, y: (x*2-1.0, y)) >>> for example in ds_trans: ... print(' x:', example[0].numpy(), ... ' y:', example[1].numpy()) x: x: x: x:

[-0.67 0.803 [-0.131 -0.416 [ 0.952 -0.13 [ 0.21 0.273

0.262] 0.285] 0.32 ] 0.229]

y: y: y: y:

0 1 2 3

Applying this sort of transformation can be used for a user-defined function. For example, if we have a dataset created from the list of image filenames on disk, we can define a function to load the images from these filenames and apply that function by calling the .map() method. You will see an example of applying multiple transformations to a dataset later in this chapter.

[ 438 ]

Chapter 13

Shuffle, batch, and repeat

As was mentioned in Chapter 2, Training Simple Machine Learning Algorithms for Classification, to train an NN model using stochastic gradient descent optimization, it is important to feed training data as randomly shuffled batches. You have already seen how to create batches by calling the .batch() method of a dataset object. Now, in addition to creating batches, you will see how to shuffle and reiterate over the datasets. We will continue working with the previous ds_joint dataset. First, let's create a shuffled version from the ds_joint dataset: >>> tf.random.set_seed(1) >>> ds = ds_joint.shuffle(buffer_size=len(t_x)) >>> for example in ds: ... print(' x:', example[0].numpy(), ... ' y:', example[1].numpy()) x: x: x: x:

[0.976 [0.435 [0.165 [0.605

0.435 0.292 0.901 0.637

0.66 ] 0.643] 0.631] 0.614]

y: y: y: y:

2 1 0 3

where the rows are shuffled without losing the one-to-one correspondence between the entries in x and y. The .shuffle() method requires an argument called buffer_size, which determines how many elements in the dataset are grouped together before shuffling. The elements in the buffer are randomly retrieved and their place in the buffer is given to the next elements in the original (unshuffled) dataset. Therefore, if we choose a small buffer_size, we may not shuffle the dataset perfectly. If the dataset is small, choosing a relatively small buffer_size may negatively affect the predictive performance of the NN as the dataset may not be completely randomized. In practice, however, it usually does not have a noticeable effect when working with relatively large datasets, which is common in deep learning. Alternatively, to ensure complete randomization during each epoch, we can simply choose a buffer size that is equal to the number of the training examples, as in the preceding code (buffer_size=len(t_x)). You will recall that dividing a dataset into batches for model training is done by calling the .batch() method. Now, let's create such batches from the ds_joint dataset and take a look at what a batch looks like: >>> ds = ds_joint.batch(batch_size=3, ... drop_remainder=False) >>> batch_x, batch_y = next(iter(ds)) >>> print('Batch-x:\n', batch_x.numpy()) [ 439 ]

Parallelizing Neural Network Training with TensorFlow Batch-x: [[0.165 0.901 0.631] [0.435 0.292 0.643] [0.976 0.435 0.66 ]] >>> print('Batch-y: ', batch_y.numpy()) Batch-y: [0 1 2]

In addition, when training a model for multiple epochs, we need to shuffle and iterate over the dataset by the desired number of epochs. So, let's repeat the batched dataset twice: >>> ds = ds_joint.batch(3).repeat(count=2) >>> for i,(batch_x, batch_y) in enumerate(ds): ... print(i, batch_x.shape, batch_y.numpy()) 0 1 2 3

(3, (1, (3, (1,

3) 3) 3) 3)

[0 1 2] [3] [0 1 2] [3]

This results in two copies of each batch. If we change the order of these two operations, that is, first batch and then repeat, the results will be different: >>> ds = ds_joint.repeat(count=2).batch(3) >>> for i,(batch_x, batch_y) in enumerate(ds): ... print(i, batch_x.shape, batch_y.numpy()) 0 (3, 3) [0 1 2] 1 (3, 3) [3 0 1] 2 (2, 3) [2 3]

Notice the difference between the batches. When we first batch and then repeat, we get four batches. On the other hand, when repeat is performed first, three batches are created. Finally, to get a better understanding of how these three operations (batch, shuffle, and repeat) behave, let's experiment with them in different orders. First, we will combine the operations in the following order: (1) shuffle, (2) batch, and (3) repeat: ## Order 1: shuffle -> batch -> repeat >>> tf.random.set_seed(1) >>> ds = ds_joint.shuffle(4).batch(2).repeat(3) >>> for i,(batch_x, batch_y) in enumerate(ds): ... print(i, batch_x.shape, batch_y.numpy()) 0 (2, 3) [2 1] 1 (2, 3) [0 3] [ 440 ]

Chapter 13 2 3 4 5

(2, (2, (2, (2,

3) 3) 3) 3)

[0 [1 [3 [1

3] 2] 0] 2]

Now, let's try a different order: (2) batch, (1) shuffle, and (3) repeat: ## Order 2: batch -> shuffle -> repeat >>> tf.random.set_seed(1) >>> ds = ds_joint.batch(2).shuffle(4).repeat(3) >>> for i,(batch_x, batch_y) in enumerate(ds): ... print(i, batch_x.shape, batch_y.numpy()) 0 (2, 3) [0 1] 1 (2, 3) [2 3] 2 (2, 3) [0 1] 3 (2, 3) [2 3] 4 (2, 3) [2 3] 5 (2, 3) [0 1]

While the first code example (shuffle, batch, repeat) appears to have shuffled the dataset as expected, we can see that in the second case (batch, shuffle, repeat), the elements within a batch were not shuffled at all. We can observe this lack of shuffling by taking a closer look at the tensor containing the target values, y. All batches contain either the pair of values [y=0, y=1] or the remaining pair of values [y=2, y=3]; we do not observe the other possible permutations: [y=2, y=0], [y=1, y=3], and so forth. Note that in order to ensure these results are not coincidental, you may want to repeat this with a higher number than 3. For example, try it with .repeat(20). Now, can you predict what will happen if we use the shuffle operation after repeat, for example, (2) batch, (3) repeat, (1) shuffle? Give it a try. One common source of error is to call .batch() twice in a row on a given dataset. By doing this, retrieving items from the resulting dataset will create a batch of batches of examples. Basically, each time you call .batch() on a dataset, it will increase the rank of the retrieved tensors by one.

Creating a dataset from files on your local storage disk

In this section, we will build a dataset from image files stored on disk. There is an image folder associated with the online content of this chapter. After downloading the folder, you should be able to see six images of cats and dogs in JPEG format. [ 441 ]

Parallelizing Neural Network Training with TensorFlow

This small dataset will show how building a dataset from stored files generally works. To accomplish this, we are going to use two additional modules in TensorFlow: tf.io to read the image file contents, and tf.image to decode the raw contents and image resizing. tf.io and tf.image modules The tf.io and tf.image modules provide a lot of additional and useful functions, which are beyond the scope of the book. You are encouraged to browse through the official documentation to learn more about these functions: https://www.tensorflow.org/versions/r2.0/api_docs/ python/tf/io for tf.io https://www.tensorflow.org/versions/r2.0/api_docs/ python/tf/image for tf.image

Before we start, let's take a look at the content of these files. We will use the pathlib library to generate a list of image files: >>> import pathlib >>> imgdir_path = pathlib.Path('cat_dog_images') >>> file_list = sorted([str(path) for path in ... imgdir_path.glob('*.jpg')]) ['cat_dog_images/dog-03.jpg', 'cat_dog_images/cat-01.jpg', 'cat_dog_ images/cat-02.jpg', 'cat_dog_images/cat-03.jpg', 'cat_dog_images/dog01.jpg', 'cat_dog_images/dog-02.jpg']

Next, we will visualize these image examples using Matplotlib: >>> import matplotlib.pyplot as plt >>> fig = plt.figure(figsize=(10, 5)) >>> for i, file in enumerate(file_list): ... img_raw = tf.io.read_file(file) ... img = tf.image.decode_image(img_raw) ... print('Image shape: ', img.shape) ... ax = fig.add_subplot(2, 3, i+1) ... ax.set_xticks([]); ax.set_yticks([]) ... ax.imshow(img) ... ax.set_title(os.path.basename(file), size=15) >>> plt.tight_layout() >>> plt.show() Image shape: (900, 1200, 3) Image shape: (900, 1200, 3) Image shape: (900, 1200, 3) [ 442 ]

Chapter 13 Image shape: Image shape: Image shape:

(900, 742, 3) (800, 1200, 3) (800, 1200, 3)

The following figure shows the example images:

Just from this visualization and the printed image shapes, we can already see that the images have different aspect ratios. If you print the aspect ratios (or data array shapes) of these images, you will see that some images are 900 pixels high and 1200 pixels wide (900 × 1200), some are 800 × 1200, and one is 900 × 742 . Later, we will preprocess these images to a consistent size. Another point to consider is that the labels for these images are provided within their filenames. So, we extract these labels from the list of filenames, assigning label 1 to dogs and label 0 to cats: >>> labels = [1 if 'dog' in os.path.basename(file) else 0 ... for file in file_list] >>> print(labels) [1, 0, 0, 0, 1, 1]

Now, we have two lists: a list of filenames (or paths of each image) and a list of their labels. In the previous section, you already learned two ways of creating a joint dataset from two tensors. Here, we will use the second approach as follows: >>> ds_files_labels = tf.data.Dataset.from_tensor_slices( ... (file_list, labels)) >>> for item in ds_files_labels: ... print(item[0].numpy(), item[1].numpy()) b'cat_dog_images/dog-03.jpg' 1 b'cat_dog_images/cat-01.jpg' 0 b'cat_dog_images/cat-02.jpg' 0 b'cat_dog_images/cat-03.jpg' 0 b'cat_dog_images/dog-01.jpg' 1 b'cat_dog_images/dog-02.jpg' 1 [ 443 ]

Parallelizing Neural Network Training with TensorFlow

We have called this dataset ds_files_labels, since it has filenames and labels. Next, we need to apply transformations to this dataset: load the image content from its file path, decode the raw content, and resize it to a desired size, for example, 80 × 120. Previously, we saw how to apply a lambda function using the .map() method. However, since we need to apply multiple preprocessing steps this time, we are going to write a helper function instead and use it when calling the .map() method: >>> def load_and_preprocess(path, label): ... image = tf.io.read_file(path) ... image = tf.image.decode_jpeg(image, channels=3) ... image = tf.image.resize(image, [img_height, img_width]) ... image /= 255.0 ... return image, label >>> >>> >>> >>> >>> ... ... ... ... ... >>> >>>

img_width, img_height = 120, 80 ds_images_labels = ds_files_labels.map(load_and_preprocess) fig = plt.figure(figsize=(10, 6)) for i,example in enumerate(ds_images_labels): ax = fig.add_subplot(2, 3, i+1) ax.set_xticks([]); ax.set_yticks([]) ax.imshow(example[0]) ax.set_title('{}'.format(example[1].numpy()), size=15) plt.tight_layout() plt.show()

This results in the following visualization of the retrieved example images, along with their labels:

[ 444 ]

Chapter 13

The load_and_preprocess() function wraps all four steps into a single function, including the loading of the raw content, decoding it, and resizing the images. The function then returns a dataset that we can iterate over and apply other operations that we learned about in the previous sections.

Fetching available datasets from the tensorflow_datasets library

The tensorflow_datasets library provides a nice collection of freely available datasets for training or evaluating deep learning models. The datasets are nicely formatted and come with informative descriptions, including the format of features and labels and their type and dimensionality, as well as the citation of the original paper that introduced the dataset in BibTeX format. Another advantage is that these datasets are all prepared and ready to use as tf.data.Dataset objects, so all the functions we covered in the previous sections can be used directly. So, let's see how to use these datasets in action. First, we need to install the tensorflow_datasets library via pip from the command line: pip install tensorflow-datasets

Now, let's import this module and take a look at the list of available datasets: >>> import tensorflow_datasets as tfds >>> print(len(tfds.list_builders())) 101 >>> print(tfds.list_builders()[:5]) ['abstract_reasoning', 'aflw2k3d', 'amazon_us_reviews', 'bair_robot_ pushing_small', 'bigearthnet']

The preceding code indicates that there are currently 101 datasets available (101 datasets at the time of writing this chapter, but this number will likely increase)— we printed the first five datasets to the command line. There are two ways of fetching a dataset, which we will cover in the following paragraphs by fetching two different datasets: CelebA (celeb_a) and the MNIST digit dataset. The first approach consists of three steps: 1. Calling the dataset builder function 2. Executing the download_and_prepare() method 3. Calling the as_dataset() method

[ 445 ]

Parallelizing Neural Network Training with TensorFlow

Let's work with the first step for the CelebA dataset and print the associated description that is provided within the library: >>> celeba_bldr = tfds.builder('celeb_a') >>> print(celeba_bldr.info.features) FeaturesDict({'image': Image(shape=(218, 178, 3), dtype=tf.uint8), 'landmarks': FeaturesDict({'lefteye_x': Tensor(shape=(), dtype=tf. int64), 'lefteye_y': Tensor(shape=(), dtype=tf.int64), 'righteye_x': Tensor(shape=(), dtype=tf.int64), 'righteye_y': ... >>> print(celeba_bldr.info.features['image']) Image(shape=(218, 178, 3), dtype=tf.uint8) >>> print(celeba_bldr.info.features['attributes'].keys()) dict_keys(['5_o_Clock_Shadow', 'Arched_Eyebrows', ... >>> print(celeba_bldr.info.citation) @inproceedings{conf/iccv/LiuLWT15, added-at = {2018-10-09T00:00:00.000+0200}, author = {Liu, Ziwei and Luo, Ping and Wang, Xiaogang and Tang, Xiaoou}, biburl = {https://www.bibsonomy.org/bibtex/250e4959be61db325d2f02c1d 8cd7bfbb/dblp}, booktitle = {ICCV}, crossref = {conf/iccv/2015}, ee = {http://doi.ieeecomputersociety.org/10.1109/ICCV.2015.425}, interhash = {3f735aaa11957e73914bbe2ca9d5e702}, intrahash = {50e4959be61db325d2f02c1d8cd7bfbb}, isbn = {978-1-4673-8391-2}, keywords = {dblp}, pages = {3730-3738}, publisher = {IEEE Computer Society}, timestamp = {2018-10-11T11:43:28.000+0200}, title = {Deep Learning Face Attributes in the Wild.}, url = {http://dblp.uni-trier.de/db/conf/iccv/iccv2015. html#LiuLWT15}, year = 2015 }

This provides some useful information to understand the structure of this dataset. The features are stored as a dictionary with three keys: 'image', 'landmarks', and 'attributes'.

[ 446 ]

Chapter 13

The 'image' entry refers to the face image of a celebrity; 'landmarks' refers to the dictionary of extracted facial points, such as the position of the eyes, nose, and so on; and 'attributes' is a dictionary of 40 facial attributes for the person in the image, like facial expression, makeup, hair properties, and so on. Next, we will call the download_and_prepare() method. This will download the data and store it on disk in a designated folder for all TensorFlow Datasets. If you have already done this once, it will simply check whether the data is already downloaded so that it does not re-download it if it already exists in the designated location: >>> celeba_bldr.download_and_prepare()

Next, we will instantiate the datasets as follows: >>> datasets = celeba_bldr.as_dataset(shuffle_files=False) >>> datasets.keys() dict_keys(['test', 'train', 'validation'])

This dataset is already split into train, test, and validation datasets. In order to see what the image examples look like, we can execute the following code: >>> ds_train = datasets['train'] >>> assert isinstance(ds_train, tf.data.Dataset) >>> example = next(iter(ds_train)) >>> print(type(example))

>>> print(example.keys()) dict_keys(['image', 'landmarks', 'attributes'])

Note that the elements of this dataset come in a dictionary. If we want to pass this dataset to a supervised deep learning model during training, we have to reformat it as a tuple of (features, label). For the label, we will use the 'Male' category from the attributes. We will do this by applying a transformation via map(): >>> ds_train = ds_train.map(lambda item: ... (item['image'], ... tf.cast(item['attributes']['Male'], tf.int32)))

Finally, let's batch the dataset and take a batch of 18 examples from it to visualize them with their labels: >>> ds_train = ds_train.batch(18) >>> images, labels = next(iter(ds_train)) >>> print(images.shape, labels) (18, 218, 178, 3) tf.Tensor([0 0 0 1 1 1 0 1 1 0 1 1 0 1 0 1 1 1], [ 447 ]

Parallelizing Neural Network Training with TensorFlow shape=(18,), dtype=int32) >>> fig = plt.figure(figsize=(12, 8)) >>> for i,(image,label) in enumerate(zip(images, labels)): ... ax = fig.add_subplot(3, 6, i+1) ... ax.set_xticks([]); ax.set_yticks([]) ... ax.imshow(image) ... ax.set_title('{}'.format(label), size=15) >>> plt.show()

The examples and their labels that are retrieved from ds_train are shown in the following figure:

This was all we needed to do to fetch and use the CelebA image dataset. Next, we will proceed with the second approach for fetching a dataset from tensorflow_datasets. There is a wrapper function called load() that combines the three steps for fetching a dataset in one. Let's see how it can be used to fetch the MNIST digit dataset: >>> mnist, mnist_info = tfds.load('mnist', with_info=True, ... shuffle_files=False) >>> print(mnist_info) tfds.core.DatasetInfo( name='mnist', version=1.0.0, description='The MNIST database of handwritten digits.',

[ 448 ]

Chapter 13 urls=['https://storage.googleapis.com/cvdf-datasets/mnist/'], features=FeaturesDict({ 'image': Image(shape=(28, 28, 1), dtype=tf.uint8), 'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=10) }, total_num_examples=70000, splits={ 'test': , 'train': }, supervised_keys=('image', 'label'), citation=""" @article{lecun2010mnist, title={MNIST handwritten digit database}, author={LeCun, Yann and Cortes, Corinna and Burges, CJ}, journal={ATT Labs [Online]. Availablist}, volume={2}, year={2010} } """, redistribution_info=, ) >>> print(mnist.keys()) dict_keys(['test', 'train'])

As we can see, the MNIST dataset is split into two partitions. Now, we can take the train partition, apply a transformation to convert the elements from a dictionary to a tuple, and visualize 10 examples: >>> ds_train = mnist['train'] >>> ds_train = ds_train.map(lambda item: ... (item['image'], item['label'])) >>> ds_train = ds_train.batch(10) >>> batch = next(iter(ds_train)) >>> print(batch[0].shape, batch[1]) (10, 28, 28, 1) tf.Tensor([8 4 7 7 0 9 0 3 3 3], shape=(10,), dtype=int64) >>> fig = plt.figure(figsize=(15, 6)) >>> for i,(image,label) in enumerate(zip(batch[0], batch[1])): ... ax = fig.add_subplot(2, 5, i+1) ... ax.set_xticks([]); ax.set_yticks([]) ... ax.imshow(image[:, :, 0], cmap='gray_r') ... ax.set_title('{}'.format(label), size=15) >>> plt.show() [ 449 ]

Parallelizing Neural Network Training with TensorFlow

The retrieved example handwritten digits from this dataset are shown as follows:

This concludes our coverage of building and manipulating datasets and fetching datasets from the tensorflow_datasets library. Next, we will see how to build NN models in TensorFlow. TensorFlow style guide Note that the official TensorFlow style guide (https://www. tensorflow.org/community/style_guide) recommends

using two-character spacing for code indents. However, this book uses four characters for indents as it is more consistent with the official Python style guide and also helps with displaying the code syntax highlighting in many text editors correctly, as well as the accompanying Jupyter code notebooks at https://github.com/rasbt/python-machinelearning-book-3rd-edition.

Building an NN model in TensorFlow

So far in this chapter, you have learned about the basic utility components of TensorFlow for manipulating tensors and organizing data into formats that we can iterate over during training. In this section, we will finally implement our first predictive model in TensorFlow. As TensorFlow is a bit more flexible but also more complex than machine learning libraries such as scikit-learn, we will start with a simple linear regression model.

[ 450 ]

Chapter 13

The TensorFlow Keras API (tf.keras)

Keras is a high-level NN API and was originally developed to run on top of other libraries such as TensorFlow and Theano. Keras provides a user-friendly and modular programming interface that allows easy prototyping and the building of complex models in just a few lines of code. Keras can be installed independently from PyPI and then configured to use TensorFlow as its backend engine. Keras is tightly integrated into TensorFlow and its modules are accessible through tf.keras. In TensorFlow 2.0, tf.keras has become the primary and recommended approach for implementing models. This has the advantage that it supports TensorFlowspecific functionalities, such as dataset pipelines using tf.data, which you learned about in the previous section. In this book, we will use the tf.keras module to build NN models. As you will see in the following subsections, the Keras API (tf.keras) makes building an NN model extremely easy. The most commonly used approach for building an NN in TensorFlow is through tf.keras.Sequential(), which allows stacking layers to form a network. A stack of layers can be given in a Python list to a model defined as tf.keras.Sequential(). Alternatively, the layers can be added one by one using the .add() method. Furthermore, tf.keras allows us to define a model by subclassing tf.keras.Model. This gives us more control over the forward pass by defining the call() method for our model class to specify the forward pass explicitly. We will see examples of both of these approaches for building an NN model using the tf.keras API. Finally, as you will see in the following subsections, models built using the tf.keras API can be compiled and trained via the .compile() and .fit() methods.

Building a linear regression model

In this subsection, we will build a simple model to solve a linear regression problem. First, let's create a toy dataset in NumPy and visualize it: >>> X_train = np.arange(10).reshape((10, 1)) >>> y_train = np.array([1.0, 1.3, 3.1, 2.0, 5.0, 6.3, ... 6.6, 7.4, 8.0, 9.0]) >>> >>> >>> >>>

plt.plot(X_train, y_train, 'o', markersize=10) plt.xlabel('x') plt.ylabel('y') plt.show()

[ 451 ]

Parallelizing Neural Network Training with TensorFlow

As a result, the training examples will be shown in a scatterplot as follows:

Next, we will standardize the features (mean centering and dividing by the standard deviation) and create a TensorFlow Dataset: >>> X_train_norm = (X_train - np.mean(X_train))/np.std(X_train) >>> ds_train_orig = tf.data.Dataset.from_tensor_slices( ... (tf.cast(X_train_norm, tf.float32), ... tf.cast(y_train, tf.float32)))

Now, we can define our model for linear regression as 𝑧𝑧 = 𝑤𝑤𝑤𝑤 + 𝑏𝑏. Here, we are going to use the Keras API. tf.keras provides predefined layers for building complex NN models, but to start, you will learn how to define a model from scratch. Later in this chapter, you will see how to use those predefined layers. For this regression problem, we will define a new class derived from the tf.keras. Model class. Subclassing tf.keras.Model allows us to use the Keras tools for exploring a model, training, and evaluation. In the constructor of our class, we will define the parameters of our model, w and b, which correspond to the weight and the bias parameters, respectively. Finally, we will define the call() method to determine how this model uses the input data to generate its output: >>> class MyModel(tf.keras.Model): ... def __init__(self): ... super(MyModel, self).__init__() [ 452 ]

Chapter 13 ... ... ... ... ...

self.w = tf.Variable(0.0, name='weight') self.b = tf.Variable(0.0, name='bias') def call(self, x): return self.w * x + self.b

Next, we will instantiate a new model from the MyModel() class that we can train based on the training data. The TensorFlow Keras API provides a method named .summary() for models that are instantiated from tf.keras.Model, which allows us to get a summary of the model components layer by layer and the number of parameters in each layer. Since we have sub-classed our model from tf.keras. Model, the .summary() method is also available to us. But, in order to be able to call model.summary(), we first need to specify the dimensionality of the input (the number of features) to this model. We can do this by calling model.build() with the expected shape of the input data: >>> model = MyModel() >>> model.build(input_shape=(None, 1)) >>> model.summary() Model: "my_model" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= Total params: 2 Trainable params: 2 Non-trainable params: 0 _________________________________________________________________

Note that we used None as a placeholder for the first dimension of the expected input tensor via model.build(), which allows us to use an arbitrary batch size. However, the number of features is fixed (here 1) as it directly corresponds to the number of weight parameters of the model. Building model layers and parameters after instantiation by calling the .build() method is called late variable creation. For this simple model, we already created the model parameters in the constructor; therefore, specifying the input_shape via build() has no further effect on our parameters, but still it is needed if we want to call model.summary(). After defining the model, we can define the cost function that we want to minimize to find the optimal model weights. Here, we will choose the mean squared error (MSE) as our cost function. Furthermore, to learn the weight parameters of the model, we will use stochastic gradient descent. In this subsection, we will implement this training via the stochastic gradient descent procedure by ourselves, but in the next subsection, we will use the Keras methods compile() and fit() to do the same thing.

[ 453 ]

Parallelizing Neural Network Training with TensorFlow

To implement the stochastic gradient descent algorithm, we need to compute the gradients. Rather than manually computing the gradients, we will use the TensorFlow API tf.GradientTape. We will cover tf.GradientTape and its different behaviors in Chapter 14, Going Deeper – The Mechanics of TensorFlow. The code is as follows: >>> def loss_fn(y_true, y_pred): ... return tf.reduce_mean(tf.square(y_true - y_pred)) >>> def train(model, inputs, outputs, learning_rate): ... with tf.GradientTape() as tape: ... current_loss = loss_fn(model(inputs), outputs) ... dW, db = tape.gradient(current_loss, [model.w, model.b]) ... model.w.assign_sub(learning_rate * dW) ... model.b.assign_sub(learning_rate * db)

Now, we can set the hyperparameters and train the model for 200 epochs. We will create a batched version of the dataset and repeat the dataset with count=None, which will result in an infinitely repeated dataset: >>> >>> >>> >>> >>> >>>

tf.random.set_seed(1) num_epochs = 200 log_steps = 100 learning_rate = 0.001 batch_size = 1 steps_per_epoch = int(np.ceil(len(y_train) / batch_size))

>>> >>> >>> >>>

ds_train ds_train ds_train Ws, bs =

= ds_train_orig.shuffle(buffer_size=len(y_train)) = ds_train.repeat(count=None) = ds_train.batch(1) [], []

>>> for i, batch in enumerate(ds_train): ... if i >= steps_per_epoch * num_epochs: ... # break the infinite loop ... break ... Ws.append(model.w.numpy()) ... bs.append(model.b.numpy()) ... ... bx, by = batch ... loss_val = loss_fn(model(bx), by) ... ... train(model, bx, by, learning_rate=learning_rate) ... if i%log_steps==0: ... print('Epoch {:4d} Step {:2d} Loss {:6.4f}'.format( [ 454 ]

Chapter 13 ...

int(i/steps_per_epoch), i, loss_val))

Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch Epoch

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190

Step Step Step Step Step Step Step Step Step Step Step Step Step Step Step Step Step Step Step Step

0 Loss 43.5600 100 Loss 0.7530 200 Loss 20.1759 300 Loss 23.3976 400 Loss 6.3481 500 Loss 4.6356 600 Loss 0.2411 700 Loss 0.2036 800 Loss 3.8177 900 Loss 0.9416 1000 Loss 0.7035 1100 Loss 0.0348 1200 Loss 0.5404 1300 Loss 0.1170 1400 Loss 0.1195 1500 Loss 0.0944 1600 Loss 0.4670 1700 Loss 2.0695 1800 Loss 0.0020 1900 Loss 0.3612

Let's look at the trained model and plot it. For test data, we will create a NumPy array of values evenly spaced between 0 to 9. Since we trained our model with standardized features, we will also apply the same standardization to the test data: >>> print('Final Parameters: ', model.w.numpy(), model.b.numpy()) Final Parameters: 2.6576622 4.8798566 >>> X_test = np.linspace(0, 9, num=100).reshape(-1, 1) >>> X_test_norm = (X_test - np.mean(X_train)) / np.std(X_train) >>> y_pred = model(tf.cast(X_test_norm, dtype=tf.float32)) >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>>

fig = plt.figure(figsize=(13, 5)) ax = fig.add_subplot(1, 2, 1) plt.plot(X_train_norm, y_train, 'o', markersize=10) plt.plot(X_test_norm, y_pred, '--', lw=3) plt.legend(['Training examples', 'Linear Reg.'], fontsize=15) ax.set_xlabel('x', size=15) ax.set_ylabel('y', size=15) ax.tick_params(axis='both', which='major', labelsize=15) ax = fig.add_subplot(1, 2, 2) plt.plot(Ws, lw=3) plt.plot(bs, lw=3) [ 455 ]

Parallelizing Neural Network Training with TensorFlow >>> >>> >>> >>> >>>

plt.legend(['Weight w', 'Bias unit b'], fontsize=15) ax.set_xlabel('Iteration', size=15) ax.set_ylabel('Value', size=15) ax.tick_params(axis='both', which='major', labelsize=15) plt.show()

The following figure shows the scatterplot of the training examples and the trained linear regression model, as well as the convergence history of the weight, w, and the bias unit, b:

Model training via the .compile() and .fit() methods

In the previous example, we saw how to train a model by writing a custom function, train(), and applied stochastic gradient descent optimization. However, writing the train() function can be a repeatable task across different projects. The TensorFlow Keras API provides a convenient .fit() method that can be called on an instantiated model. To show how this works, let's create a new model and compile it by selecting the optimizer, loss function, and evaluation metrics: >>> tf.random.set_seed(1) >>> model = MyModel() >>> model.compile(optimizer='sgd', ... loss=loss_fn, ... metrics=['mae', 'mse'])

Now, we can simply call the fit() method to train the model. We can pass a batched dataset (like ds_train, which was created in the previous example). However, this time you will see that we can pass the NumPy arrays for x and y directly, without needing to create a dataset: [ 456 ]

Chapter 13 >>> model.fit(X_train_norm, y_train, ... epochs=num_epochs, batch_size=batch_size, ... verbose=1) Train on 10 samples Epoch 1/200 10/10 [==============================] - 0s 4ms/sample - loss: 27.8578 - mae: 4.5810 - mse: 27.8578 Epoch 2/200 10/10 [==============================] - 0s 738us/sample - loss: 18.6640 - mae: 3.7395 - mse: 18.6640 ... Epoch 200/200 10/10 [==============================] - 0s 1ms/sample - loss: 0.4139 - mae: 0.4942 - mse: 0.4139

After the model is trained, visualize the results and make sure that they are similar to the results of the previous method.

Building a multilayer perceptron for classifying flowers in the Iris dataset

In the previous example, you saw how to build a model from scratch. We trained this model using stochastic gradient descent optimization. While we started our journey based on the simplest possible example, you can see that defining the model from scratch, even for such a simple case, is neither appealing nor a good practice. TensorFlow instead provides already defined layers through tf.keras.layers that can be readily used as the building blocks of an NN model. In this section, you will learn how to use these layers to solve a classification task using the Iris flower dataset and build a two-layer perceptron using the Keras API. First, let's get the data from tensorflow_datasets: >>> iris, iris_info = tfds.load('iris', with_info=True) >>> print(iris_info)

This prints some information about this dataset (not printed here to save space). However, you will notice in the shown information that this dataset comes with only one partition, so we have to split the dataset into training and testing partitions (and validation for proper machine learning practice) on our own. Let's assume that we want to use two-thirds of the dataset for training and keep the remaining examples for testing. The tensorflow_datasets library provides a convenient tool that allows us to determine slices and splits via the DatasetBuilder object prior to loading a dataset. You can find out more about splits at https://www.tensorflow.org/ datasets/splits. [ 457 ]

Parallelizing Neural Network Training with TensorFlow

An alternative approach is to load the entire dataset first and then use .take() and .skip() to split the dataset to two partitions. If the dataset is not shuffled at first, we can also shuffle the dataset. However, we need to be very careful with this because it can lead to mixing the train/test examples, which is not acceptable in machine learning. To avoid this, we have to set an argument, reshuffle_each_ iteration=False, in the .shuffle() method. The code for splitting the dataset into train/test is as follows: >>> tf.random.set_seed(1) >>> ds_orig = iris['train'] >>> ds_orig = ds_orig.shuffle(150, reshuffle_each_iteration=False) >>> ds_train_orig = ds_orig.take(100) >>> ds_test = ds_orig.skip(100)

Next, as you have already seen in the previous sections, we need to apply a transformation via the .map() method to convert the dictionary to a tuple: >>> ds_train_orig = ds_train_orig.map( ... lambda x: (x['features'], x['label'])) >>> ds_test = ds_test.map( ... lambda x: (x['features'], x['label']))

Now, we are ready to use the Keras API to build a model efficiently. In particular, using the tf.keras.Sequential class, we can stack a few Keras layers and build an NN. You can see the list of all the Keras layers that are already available at https:// www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/layers. For this problem, we are going to use the Dense layer (tf.keras.layers.Dense), which is also known as a fully connected (FC) layer or linear layer, and can be best represented by 𝑓𝑓(𝑤𝑤 × 𝑥𝑥 + 𝑏𝑏) , where x is the input features, w and b are the weight matrix and the bias vector, and f is the activation function. If you think of the layers in an NN, each layer receives its inputs from the preceding layer; therefore, its dimensionality (rank and shape) is fixed. Typically, we need to concern ourselves with the dimensionality of output only when we design an NN architecture. (Note: the first layer is the exception, but TensorFlow/Keras allows us to decide the input dimensionality of the first layer after defining the model through late variable creation.) Here, we want to define a model with two hidden layers. The first one receives an input of four features and projects them to 16 neurons. The second layer receives the output of the previous layer (which has size 16) and projects them to three output neurons, since we have three class labels. This can be done using the Sequential class and the Dense layer in Keras as follows: >>> iris_model = tf.keras.Sequential([

[ 458 ]

Chapter 13 ... ... ... ...

tf.keras.layers.Dense(16, activation='sigmoid', name='fc1', input_shape=(4,)), tf.keras.layers.Dense(3, name='fc2', activation='softmax')])

>>> iris_model.summary() Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= fc1 (Dense) (None, 16) 80 _________________________________________________________________ fc2 (Dense) (None, 3) 51 ================================================================= Total params: 131 Trainable params: 131 Non-trainable params: 0 _________________________________________________________________

Notice that we determined the input shape for the first layer via input_shape=(4,), and therefore, we did not have to call .build() anymore in order to use iris_ model.summary(). The printed model summary indicates that the first layer (fc1) has 80 parameters, and the second layer has 51 parameters. You can verify that by (𝑛𝑛𝑖𝑖𝑖𝑖 + 1) × 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜 , where 𝑛𝑛𝑖𝑖𝑖𝑖 is the number of input units, and 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜 is the number of output units. Recall that for a fully (densely) connected layer, the learnable parameters are the weight matrix of size 𝑛𝑛𝑖𝑖𝑖𝑖 × 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜 and the bias vector of size 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜. Furthermore, notice that we used the sigmoid activation function for the first layer and softmax activation for the last (output) layer. Softmax activation in the last layer is used to support multi-class classification, since here we have three class labels (which is why we have three neurons at the output layer). We will discuss the different activation functions and their applications later in this chapter. Next, we will compile this model to specify the loss function, the optimizer, and the metrics for evaluation: >>> iris_model.compile(optimizer='adam', ... loss='sparse_categorical_crossentropy', ... metrics=['accuracy'])

Now, we can train the model. We will specify the number of epochs to be 100 and the batch size to be 2. In the following code, we will build an infinitely repeating dataset, which will be passed to the fit() method for training the model. In this case, in order for the fit() method to be able to keep track of the epochs, it needs to know the number of steps for each epoch. [ 459 ]

Parallelizing Neural Network Training with TensorFlow

Given the size of our training data (here, 100) and the batch size (batch_size), we can determine the number of steps in each epoch, steps_per_epoch: >>> >>> >>> >>>

num_epochs = 100 training_size = 100 batch_size = 2 steps_per_epoch = np.ceil(training_size / batch_size)

>>> >>> >>> >>>

ds_train ds_train ds_train ds_train

= = = =

ds_train_orig.shuffle(buffer_size=training_size) ds_train.repeat() ds_train.batch(batch_size=batch_size) ds_train.prefetch(buffer_size=1000)

>>> history = iris_model.fit(ds_train, epochs=num_epochs, ... steps_per_epoch=steps_per_epoch, ... verbose=0)

The returned variable history keeps the training loss and the training accuracy (since they were specified as metrics to iris_model.compile()) after each epoch. We can use this to visualize the learning curves as follows: >>> hist = history.history >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>>

fig = plt.figure(figsize=(12, 5)) ax = fig.add_subplot(1, 2, 1) ax.plot(hist['loss'], lw=3) ax.set_title('Training loss', size=15) ax.set_xlabel('Epoch', size=15) ax.tick_params(axis='both', which='major', labelsize=15) ax = fig.add_subplot(1, 2, 2) ax.plot(hist['accuracy'], lw=3) ax.set_title('Training accuracy', size=15) ax.set_xlabel('Epoch', size=15) ax.tick_params(axis='both', which='major', labelsize=15) plt.show()

The learning curves (training loss and training accuracy) are as follows:

[ 460 ]

Chapter 13

Evaluating the trained model on the test dataset

Since we specified 'accuracy' as our evaluation metric in iris_model.compile(), we can now directly evaluate the model on the test dataset: >>> results = iris_model.evaluate(ds_test.batch(50), verbose=0) >>> print('Test loss: {:.4f} Test Acc.: {:.4f}'.format(*results)) Test loss: 0.0692 Test Acc.: 0.9800

Notice that we have to batch the test dataset as well, to ensure that the input to the model has the correct dimension (rank). As we discussed earlier, calling .batch() will increase the rank of the retrieved tensors by 1. The input data for .evaluate() must have one designated dimension for the batch, although here (for evaluation), the size of the batch does not matter. Therefore, if we pass ds_batch.batch(50) to the .evaluate() method, the entire test dataset will be processed in one batch of size 50, but if we pass ds_batch.batch(1), 50 batches of size 1 will be processed.

Saving and reloading the trained model

Trained models can be saved on disk for future use. This can be done as follows: >>> iris_model.save('iris-classifier.h5', ... overwrite=True, ... include_optimizer=True, ... save_format='h5')

[ 461 ]

Parallelizing Neural Network Training with TensorFlow

The first option is the filename. Calling iris_model.save() will save both the model architecture and all the learned parameters. However, if you want to save only the architecture, you can use the iris_model.to_json() method, which saves the model configuration in JSON format. Or if you want to save only the model weights, you can do that by calling iris_model.save_weights(). The save_format can be specified to be either 'h5' for the HDF5 format or 'tf' for TensorFlow format. Now, let's reload the saved model. Since we have saved both the model architecture and the weights, we can easily rebuild and reload the parameters in just one line: >>> iris_model_new = tf.keras.models.load_model('iris-classifier.h5')

Try to verify the model architecture by calling iris_model_new.summary(). Finally, let's evaluate this new model that is reloaded on the test dataset to verify that the results are the same as before: >>> results = iris_model_new.evaluate(ds_test.batch(33), verbose=0) >>> print('Test loss: {:.4f} Test Acc.: {:.4f}'.format(*results)) Test loss: 0.0692 Test Acc.: 0.9800

Choosing activation functions for multilayer neural networks

For simplicity, we have only discussed the sigmoid activation function in the context of multilayer feedforward NNs so far; we used it in the hidden layer as well as the output layer in the MLP implementation in Chapter 12, Implementing a Multilayer Artificial Neural Network from Scratch. 1

Note that in this book, the sigmoidal logistic function, 𝜎𝜎(𝑧𝑧) = 1 + 𝑒𝑒 −𝑧𝑧 , is referred to as the sigmoid function for brevity, which is common in machine learning literature. In the following subsections, you will learn more about alternative nonlinear functions that are useful for implementing multilayer NNs. Technically, we can use any function as an activation function in multilayer NNs as long as it is differentiable. We can even use linear activation functions, such as in Adaline (Chapter 2, Training Simple Machine Learning Algorithms for Classification). However, in practice, it would not be very useful to use linear activation functions for both hidden and output layers, since we want to introduce nonlinearity in a typical artificial NN to be able to tackle complex problems. The sum of linear functions yields a linear function after all.

[ 462 ]

Chapter 13

The logistic (sigmoid) activation function that we used in Chapter 12 probably mimics the concept of a neuron in a brain most closely—we can think of it as the probability of whether a neuron fires. However, the logistic (sigmoid) activation function can be problematic if we have highly negative input, since the output of the sigmoid function will be close to zero in this case. If the sigmoid function returns output that is close to zero, the NN will learn very slowly, and it will be more likely to get trapped in the local minima during training. This is why people often prefer a hyperbolic tangent as an activation function in hidden layers. Before we discuss what a hyperbolic tangent looks like, let's briefly recapitulate some of the basics of the logistic function and look at a generalization that makes it more useful for multilabel classification problems.

Logistic function recap

As was mentioned in the introduction to this section, the logistic function is, in fact, a special case of a sigmoid function. You will recall from the section on logistic regression in Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, that we can use a logistic function to model the probability that sample x belongs to the positive class (class 1) in a binary classification task. The given net input, z, is shown in the following equation: 𝑚𝑚

𝑧𝑧 = 𝑤𝑤0 𝑥𝑥0 + 𝑤𝑤1 𝑥𝑥1 + ⋯ + 𝑤𝑤𝑚𝑚 𝑥𝑥𝑚𝑚 = ∑ 𝑤𝑤𝑖𝑖 𝑥𝑥𝑖𝑖 = 𝑤𝑤 𝑇𝑇 𝑥𝑥 𝑖𝑖=0

The logistic (sigmoid) function will compute the following:

𝜙𝜙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 (𝑧𝑧) =

1 1 + 𝑒𝑒 −𝑧𝑧

Note that 𝑤𝑤0 is the bias unit (y-axis intercept, which means 𝑥𝑥0 = 1). To provide a more concrete example, let's take a model for a two-dimensional data point, x, and a model with the following weight coefficients assigned to the w vector: >>> import numpy as np >>> X = np.array([1, 1.4, 2.5]) ## first value must be 1 >>> w = np.array([0.4, 0.3, 0.5]) >>> def net_input(X, w): ... return np.dot(X, w)

[ 463 ]

Parallelizing Neural Network Training with TensorFlow >>> def logistic(z): ... return 1.0 / (1.0 + np.exp(-z)) >>> def logistic_activation(X, w): ... z = net_input(X, w) ... return logistic(z) >>> print('P(y=1|x) = %.3f' % logistic_activation(X, w)) P(y=1|x) = 0.888

If we calculate the net input (z) and use it to activate a logistic neuron with those particular feature values and weight coefficients, we get a value of 0.888, which we can interpret as an 88.8 percent probability that this particular sample, x, belongs to the positive class. In Chapter 12, we used the one-hot encoding technique to represent multiclass ground truth labels and designed the output layer consisting of multiple logistic activation units. However, as will be demonstrated by the following code example, an output layer consisting of multiple logistic activation units does not produce meaningful, interpretable probability values: >>> # W : array with shape = (n_output_units, n_hidden_units+1) >>> # note that the first column are the bias units >>> W = np.array([[1.1, 1.2, 0.8, 0.4], ... [0.2, 0.4, 1.0, 0.2], ... [0.6, 1.5, 1.2, 0.7]]) >>> # A : data array with shape = (n_hidden_units + 1, n_samples) >>> # note that the first column of this array must be 1 >>> A = np.array([[1, 0.1, 0.4, 0.6]]) >>> Z = np.dot(W, A[0]) >>> y_probas = logistic(Z) >>> print('Net Input: \n', Z) Net Input: [ 1.78 0.76 1.65] >>> print('Output Units:\n', y_probas) Output Units: [ 0.85569687 0.68135373 0.83889105]

As you can see in the output, the resulting values cannot be interpreted as probabilities for a three-class problem. The reason for this is that they do not sum up to 1. However, this is, in fact, not a big concern if we use our model to predict only the class labels and not the class membership probabilities. One way to predict the class label from the output units obtained earlier is to use the maximum value: [ 464 ]

Chapter 13 >>> y_class = np.argmax(Z, axis=0) >>> print('Predicted class label: %d' % y_class) Predicted class label: 0

In certain contexts, it can be useful to compute meaningful class probabilities for multiclass predictions. In the next section, we will take a look at a generalization of the logistic function, the softmax function, which can help us with this task.

Estimating class probabilities in multiclass classification via the softmax function

In the previous section, you saw how we can obtain a class label using the argmax function. Previously, in the section Building a multilayer perceptron for classifying flowers in the Iris dataset, we determined activation='softmax' in the last layer of the MLP model. The softmax function is a soft form of the argmax function; instead of giving a single class index, it provides the probability of each class. Therefore, it allows us to compute meaningful class probabilities in multiclass settings (multinomial logistic regression). In softmax, the probability of a particular sample with net input z belonging to the ith class can be computed with a normalization term in the denominator, that is, the sum of the exponentially weighted linear functions:

𝑝𝑝(𝑧𝑧) = 𝜙𝜙(𝑧𝑧) =

𝑒𝑒 𝑧𝑧𝑖𝑖 𝑧𝑧𝑗𝑗 ∑𝑀𝑀 𝑗𝑗=1 𝑒𝑒

To see softmax in action, let's code it up in Python:

>>> def softmax(z): ... return np.exp(z) / np.sum(np.exp(z)) >>> y_probas = softmax(Z) >>> print('Probabilities:\n', y_probas) Probabilities: [ 0.44668973 0.16107406 0.39223621] >>> np.sum(y_probas) 1.0

As you can see, the predicted class probabilities now sum up to 1, as we would expect. It is also notable that the predicted class label is the same as when we applied the argmax function to the logistic output.

[ 465 ]

Parallelizing Neural Network Training with TensorFlow

It may help to think of the result of the softmax function as a normalized output that is useful for obtaining meaningful class-membership predictions in multiclass settings. Therefore, when we build a multiclass classification model in TensorFlow, we can use the tf.keras.activations.softmax() function to estimate the probabilities of each class membership for an input batch of examples. To see how we can use the softmax activation function in TensorFlow, in the following code, we will convert Z to a tensor, with an additional dimension reserved for the batch size: >>> import tensorflow as tf >>> Z_tensor = tf.expand_dims(Z, axis=0) >>> tf.keras.activations.softmax(Z_tensor)

Broadening the output spectrum using a hyperbolic tangent

Another sigmoidal function that is often used in the hidden layers of artificial NNs is the hyperbolic tangent (commonly known as tanh), which can be interpreted as a rescaled version of the logistic function:

𝜙𝜙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 (𝑧𝑧) =

1 1 + 𝑒𝑒 −𝑧𝑧

𝜙𝜙𝑡𝑡𝑡𝑡𝑡𝑡ℎ (𝑧𝑧) = 2 × 𝜙𝜙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 (2𝑧𝑧) − 1 =

𝑒𝑒 𝑧𝑧 − 𝑒𝑒 −𝑧𝑧 𝑒𝑒 𝑧𝑧 + 𝑒𝑒 −𝑧𝑧

The advantage of the hyperbolic tangent over the logistic function is that it has a broader output spectrum ranging in the open interval (–1, 1), which can improve the convergence of the back-propagation algorithm (Neural Networks for Pattern Recognition, C. M. Bishop, Oxford University Press, pages: 500-501, 1995). In contrast, the logistic function returns an output signal ranging in the open interval (0, 1). For a simple comparison of the logistic function and the hyperbolic tangent, let's plot the two sigmoidal functions: >>> import matplotlib.pyplot as plt >>> def tanh(z): ... e_p = np.exp(z) ... e_m = np.exp(-z) ... return (e_p - e_m) / (e_p + e_m)

[ 466 ]

Chapter 13 >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> ... ... >>> ... ... >>> >>> >>>

z = np.arange(-5, 5, 0.005) log_act = logistic(z) tanh_act = tanh(z) plt.ylim([-1.5, 1.5]) plt.xlabel('net input $z$') plt.ylabel('activation $\phi(z)$') plt.axhline(1, color='black', linestyle=':') plt.axhline(0.5, color='black', linestyle=':') plt.axhline(0, color='black', linestyle=':') plt.axhline(-0.5, color='black', linestyle=':') plt.axhline(-1, color='black', linestyle=':') plt.plot(z, tanh_act, linewidth=3, linestyle='--', label='tanh') plt.plot(z, log_act, linewidth=3, label='logistic') plt.legend(loc='lower right') plt.tight_layout() plt.show()

As you can see, the shapes of the two sigmoidal curves look very similar; however, the tanh function has double the output space of the logistic function:

Note that we previously implemented the logistic and tanh functions verbosely for the purpose of illustration. In practice, we can use NumPy's tanh function. [ 467 ]

Parallelizing Neural Network Training with TensorFlow

Alternatively, when building an NN model, we can use tf.keras.activations.

tanh() in TensorFlow to achieve the same results:

>>> np.tanh(z) array([-0.9999092 , -0.99990829, -0.99990737, ..., 0.99990737, 0.99990829])

0.99990644,

>>> tf.keras.activations.tanh(z)

In addition, the logistic function is available in SciPy's special module: >>> from scipy.special import expit >>> expit(z) array([0.00669285, 0.00672617, 0.00675966, ..., 0.99320669, 0.99324034, 0.99327383])

Similarly, we can use the tf.keras.activations.sigmoid() function in TensorFlow to do the same computation, as follows: >>> tf.keras.activations.sigmoid(z)

Rectified linear unit activation

Rectified linear unit (ReLU) is another activation function that is often used in deep NNs. Before we delve into ReLU, we should step back and understand the vanishing gradient problem of tanh and logistic activations. To understand this problem, let's assume that we initially have the net input 𝑧𝑧1 = 20 , which changes to 𝑧𝑧2 = 25. Computing the tanh activation, we get 𝜙𝜙(𝑧𝑧1 ) = 1.0 and 𝜙𝜙(𝑧𝑧2 ) = 1.0, which shows no change in the output (due to the asymptotic behavior of the tanh function and numerical errors).

This means that the derivative of activations with respect to the net input diminishes as z becomes large. As a result, learning weights during the training phase becomes very slow because the gradient terms may be very close to zero. ReLU activation addresses this issue. Mathematically, ReLU is defined as follows:

𝜙𝜙(𝑧𝑧) = 𝑚𝑚𝑚𝑚𝑚𝑚(0, 𝑧𝑧) [ 468 ]

Chapter 13

ReLU is still a nonlinear function that is good for learning complex functions with NNs. Besides this, the derivative of ReLU, with respect to its input, is always 1 for positive input values. Therefore, it solves the problem of vanishing gradients, making it suitable for deep NNs. In TensorFlow, we can apply the ReLU activation as follows: >>> tf.keras.activations.tanh(z)

,

We will use the ReLU activation function in the next chapter as an activation function for multilayer convolutional NNs. Now that we know more about the different activation functions that are commonly used in artificial NNs, let's conclude this section with an overview of the different activation functions that we encountered in this book:

[ 469 ]

Parallelizing Neural Network Training with TensorFlow

You can find the list of all activation functions available in the Keras API at https:// www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/activations.

Summary

In this chapter, you learned how to use TensorFlow, an open source library for numerical computations, with a special focus on deep learning. While TensorFlow is more inconvenient to use than NumPy, due to its additional complexity to support GPUs, it allows us to define and train large, multilayer NNs very efficiently. Also, you learned about using the TensorFlow Keras API to build complex machine learning and NN models and run them efficiently. We explored model building in TensorFlow by defining a model from scratch via subclassing the tf.keras.Model class. Implementing models can be tedious when we have to program at the level of matrix-vector multiplications and define every detail of each operation. However, the advantage is that this allows us, as developers, to combine such basic operations and build more complex models. We then explored tf.keras.layers, which makes building NN models a lot easier than implementing them from scratch. Finally, you learned about different activation functions and understood their behaviors and applications. Specifically, in this chapter, we covered tanh, softmax, and ReLU. In the next chapter, we'll continue our journey and dive deeper into TensorFlow, where we'll find ourselves working with TensorFlow function decoration and TensorFlow Estimators. Along the way, you'll learn many new concepts, such as variables and feature columns.

[ 470 ]

Going Deeper – The Mechanics of TensorFlow In Chapter 13, Parallelizing Neural Network Training with TensorFlow, we covered how to define and manipulate tensors and worked with the tf.data API to build input pipelines. We further built and trained a multilayer perceptron to classify the Iris dataset using the TensorFlow Keras API (tf.keras). Now that we have some hands-on experience with TensorFlow neural network (NN) training and machine learning, it's time to take a deeper dive into the TensorFlow library and explore its rich set of features, which will allow us to implement more advanced deep learning models in upcoming chapters. In this chapter, we will use different aspects of TensorFlow's API to implement NNs. In particular, we will again use the Keras API, which provides multiple layers of abstraction to make the implementation of standard architectures very convenient. TensorFlow also allows us to implement custom NN layers, which is very useful in research-oriented projects that require more customization. Later in this chapter, we will implement such a custom layer. To illustrate the different ways of model building using the Keras API, we will also consider the classic exclusive or (XOR) problem. Firstly, we will build multilayer perceptrons using the Sequential class. Then, we will consider other methods, such as subclassing tf.keras.Model for defining custom layers. Finally, we will cover tf.estimator, a high-level TensorFlow API that encapsulates the machine learning steps from raw input to prediction. The topics that we will cover are as follows: • Understanding and working with TensorFlow graphs and migration to TensorFlow v2 • Function decoration for graph compilation [ 471 ]

Going Deeper – The Mechanics of TensorFlow

• Working with TensorFlow variables • Solving the classic XOR problem and understanding model capacity • Building complex NN models using Keras' Model class and the Keras functional API • Computing gradients using automatic differentiation and tf.GradientTape • Working with TensorFlow Estimators

The key features of TensorFlow

TensorFlow provides us with a scalable, multiplatform programming interface for implementing and running machine learning algorithms. The TensorFlow API has been relatively stable and mature since its 1.0 release in 2017, but it just experienced a major redesign with its recent 2.0 release in 2019, which we are using in this book. Since its initial release in 2015, TensorFlow has become the most widely adopted deep learning library. However, one of its main friction points was that it was built around static computation graphs. Static computation graphs have certain advantages, such as better graph optimizations behind the scenes and support for a wider range of hardware devices; however, static computation graphs require separate graph declaration and graph evaluation steps, which make it cumbersome for users to develop and work with NNs interactively. Taking all the user feedback to heart, the TensorFlow team decided to make dynamic computation graphs the default in TensorFlow 2.0, which makes the development and training of NNs much more convenient. In the next section, we will cover some of the important changes from TensorFlow v1.x to v2. Dynamic computation graphs allow for interleaving the graph declaration and graph evaluation steps such that TensorFlow 2.0 feels much more natural for Python and NumPy users compared to previous versions of TensorFlow. However, note that TensorFlow 2.0 still allows users to use the "old" TensorFlow v1.x API via the tf.compat submodule. This helps users to transition their code bases more smoothly to the new TensorFlow v2 API. A key feature of TensorFlow, which was also noted in Chapter 13, Parallelizing Neural Network Training with TensorFlow, is its ability to work with single or multiple graphical processing units (GPUs). This allows users to train deep learning models very efficiently on large datasets and large-scale systems. While TensorFlow is an open source library and can be freely used by everyone, its development is funded and supported by Google. This involves a large team of software engineers who expand and improve the library continuously. Since TensorFlow is an open source library, it also has strong support from other developers outside of Google, who avidly contribute and provide user feedback. [ 472 ]

Chapter 14

This has made the TensorFlow library more useful to both academic researchers and developers. A further consequence of these factors is that TensorFlow has extensive documentation and tutorials to help new users. Last, but not least, TensorFlow supports mobile deployment, which also makes it a very suitable tool for production.

TensorFlow's computation graphs: migrating to TensorFlow v2

TensorFlow performs its computations based on a directed acyclic graph (DAG). In TensorFlow v1.x, such graphs could be explicitly defined in the low-level API, although this was not trivial for large and complex models. In this section, we will see how these graphs can be defined for a simple arithmetic computation. Then, we will see how to migrate a graph to TensorFlow v2, the eager execution and dynamic graph paradigm, as well as the function decoration for faster computations.

Understanding computation graphs

TensorFlow relies on building a computation graph at its core, and it uses this computation graph to derive relationships between tensors from the input all the way to the output. Let's say that we have rank 0 (scalar) tensors a, b, and c and we want to evaluate 𝑧𝑧 = 2 × (𝑎𝑎 − 𝑏𝑏) + 𝑐𝑐. This evaluation can be represented as a computation graph, as shown in the following figure:

[ 473 ]

Going Deeper – The Mechanics of TensorFlow

As you can see, the computation graph is simply a network of nodes. Each node resembles an operation, which applies a function to its input tensor or tensors and returns zero or more tensors as the output. TensorFlow builds this computation graph and uses it to compute the gradients accordingly. In the next subsections, we will see some examples of creating a graph for this computation using TensorFlow v1.x and v2 styles.

Creating a graph in TensorFlow v1.x

In the earlier version of the TensorFlow (v1.x) low-level API, this graph had to be explicitly declared. The individual steps for building, compiling, and evaluating such a computation graph in TensorFlow v1.x are as follows: 1. Instantiate a new, empty computation graph 2. Add nodes (tensors and operations) to the computation graph 3. Evaluate (execute) the graph: a. Start a new session b. Initialize the variables in the graph c. Run the computation graph in this session Before we take a look at the dynamic approach in TensorFlow v2, let's look at a simple example that illustrates how to create a graph in TensorFlow v1.x for evaluating 𝑧𝑧 = 2 × (𝑎𝑎 − 𝑏𝑏) + 𝑐𝑐, as shown in the previous figure. The variables a, b, and c are scalars (single numbers), and we define these as TensorFlow constants. A graph can then be created by calling tf.Graph(). Variables, as well as computations, represent the nodes of the graph, which we will define as follows: ## TF v1.x style >>> g = tf.Graph() >>> with g.as_default(): ... a = tf.constant(1, name='a') ... b = tf.constant(2, name='b') ... c = tf.constant(3, name='c') ... z = 2*(a-b) + c

In this code, we first defined graph g via g=tf.Graph(). Then, we added nodes to the graph, g, using with g.as_default(). However, note that if we do not explicitly create a graph, there is always a default graph to which variables and computations will be added automatically.

[ 474 ]

Chapter 14

In TensorFlow v1.x, a session is an environment in which the operations and tensors of a graph can be executed. The Session class was removed from TensorFlow v2; However, for the time being, it is still available via the tf.compat submodule to allow compatibility with TensorFlow v1.x. A session object can be created by calling tf.compat.v1.Session(), which can receive an existing graph (here, g) as an argument, as in Session(graph=g). After launching a graph in a TensorFlow session, we can execute its nodes, that is, evaluate its tensors or execute its operators. Evaluating each individual tensor involves calling its eval() method inside the current session. When evaluating a specific tensor in the graph, TensorFlow has to execute all the preceding nodes in the graph until it reaches the given node of interest. In case there are one or more placeholder variables, we also need to provide values for those through the session's run method, as we will see later in the chapter. After defining the static graph in the previous code snippet, we can execute the graph in a TensorFlow session and evaluate the tensor, z, as follows: ## TF v1.x style >>> with tf.compat.v1.Session(graph=g) as sess: ... print(Result: z =', sess.run(z)) Result: z = 1

Migrating a graph to TensorFlow v2

Next, let's look at how this code can be migrated to TensorFlow v2. TensorFlow v2 uses dynamic (as opposed to static) graphs by default (this is also called eager execution in TensorFlow), which allows us to evaluate an operation on the fly. Therefore, we do not have to explicitly create a graph and a session, which makes the development workflow much more convenient: ## TF v2 style >>> a = tf.constant(1, name='a') >>> b = tf.constant(2, name='b') >>> c = tf.constant(3, name='c') >>> z = 2*(a - b) + c >>> tf.print('Result: z= ', z) Result: z = 1

[ 475 ]

Going Deeper – The Mechanics of TensorFlow

Loading input data into a model: TensorFlow v1.x style

Another important improvement from TensorFlow v1.x to v2 is regarding how data can be loaded into our models. In TensorFlow v2, we can directly feed data in the form of Python variables or NumPy arrays. However, when using the TensorFlow v1.x low-level API, we had to create placeholder variables for providing input data to a model. For the preceding simple computation graph example, 𝑧𝑧 = 2 × (𝑎𝑎 − 𝑏𝑏) + 𝑐𝑐, let's assume that a, b, and c are the input tensors of rank 0. We can then define three placeholders, which we will then use to "feed" data to the model via a so-called feed_dict dictionary, as follows: ## TF-v1.x style >>> g = tf.Graph() >>> with g.as_default(): ... a = tf.compat.v1.placeholder(shape=None, ... dtype=tf.int32, name='tf_a') ... b = tf.compat.v1.placeholder(shape=None, ... dtype=tf.int32, name='tf_b') ... c = tf.compat.v1.placeholder(shape=None, ... dtype=tf.int32, name='tf_c') ... z = 2*(a-b) + c >>> with tf.compat.v1.Session(graph=g) as sess: ... feed_dict={a:1, b:2, c:3} ... print('Result: z =', sess.run(z, feed_dict=feed_dict)) Result: z = 1

Loading input data into a model: TensorFlow v2 style In TensorFlow v2, all this can simply be done by defining a regular Python function with a, b, and c as its input arguments, for example: ## TF-v2 style >>> def compute_z(a, b, c): ... r1 = tf.subtract(a, b) ... r2 = tf.multiply(2, r1) ... z = tf.add(r2, c) ... return z

[ 476 ]

Chapter 14

Now, to carry out the computation, we can simply call this function with Tensor objects as function arguments. Note that TensorFlow functions such as add, subtract, and multiply also allow us to provide inputs of higher ranks in the form of a TensorFlow Tensor object, a NumPy array, or possibly other Python objects, such as lists and tuples. In the following code example, we provide scalar inputs (rank 0), as well as rank 1 and rank 2 inputs, as lists: >>> tf.print('Scalar Inputs:', compute_z(1, 2, 3)) Scalar Inputs: 1 >>> tf.print('Rank 1 Inputs:', compute_z([1], [2], [3])) Rank 1 Inputs: [1] >>> tf.print('Rank 2 Inputs:', compute_z([[1]], [[2]], [[3]])) Rank 2 Inputs: [[1]]

In this section, you saw how migrating to TensorFlow v2 makes the programming style simple and efficient by avoiding explicit graph and session creation steps. Now that we have seen how TensorFlow v1.x compares to TensorFlow v2, we will focus only on TensorFlow v2 for the remainder of this book. Next, we will take a deeper look into decorating Python functions into a graph that allows for faster computation.

Improving computational performance with function decorators

As you saw in the previous section, we can easily write a normal Python function and utilize TensorFlow operations. However, computations via the eager execution (dynamic graph) mode are not as efficient as the static graph execution in TensorFlow v1.x. Thus, TensorFlow v2 provides a tool called AutoGraph that can automatically transform Python code into TensorFlow's graph code for faster execution. In addition, TensorFlow provides a simple mechanism for compiling a normal Python function to a static TensorFlow graph in order to make the computations more efficient. To see how this works in practice, let's work with our previous compute_z function and annotate it for graph compilation using the @tf.function decorator: >>> @tf.function ... def compute_z(a, b, c): ... r1 = tf.subtract(a, b) ... r2 = tf.multiply(2, r1) ... z = tf.add(r2, c) ... return z [ 477 ]

Going Deeper – The Mechanics of TensorFlow

Note that we can use and call this function the same way as before, but now TensorFlow will construct a static graph based on the input arguments. Python supports dynamic typing and polymorphism, so we can define a function such as def f(a, b): return a+b and then call it using integer, float, list, or string inputs (recall that a+b is a valid operation for lists and strings). While TensorFlow graphs require static types and shapes, tf.function supports such a dynamic typing capability. For example, let's call this function with the following inputs: >>> tf.print('Scalar Inputs:', compute_z(1, 2, 3)) >>> tf.print('Rank 1 Inputs:', compute_z([1], [2], [3])) >>> tf.print('Rank 2 Inputs:', compute_z([[1]], [[2]], [[3]]))

This will produce the same outputs as before. Here, TensorFlow uses a tracing mechanism to construct a graph based on the input arguments. For this tracing mechanism, TensorFlow generates a tuple of keys based on the input signatures given for calling the function. The generated keys are as follows: • For tf.Tensor arguments, the key is based on their shapes and dtypes. • For Python types, such as lists, their id() is used to generate cache keys. • For Python primitive values, the cache keys are based on the input values. Upon calling such a decorated function, TensorFlow will check whether a graph with the corresponding key has already been generated. If such a graph does not exist, TensorFlow will generate a new graph and store the new key. On the other hand, if we want to limit the way a function can be called, we can specify its input signature via a tuple of tf.TensorSpec objects when defining the function. For example, let's redefine the previous function, compute_z, and specify that only rank 1 tensors of type tf.int32 are allowed: >>> @tf.function(input_signature=(tf.TensorSpec(shape=[None], ... dtype=tf.int32), ... tf.TensorSpec(shape=[None], ... dtype=tf.int32), ... tf.TensorSpec(shape=[None], ... dtype=tf.int32),)) ... def compute_z(a, b, c): ... r1 = tf.subtract(a, b) ... r2 = tf.multiply(2, r1) ... z = tf.add(r2, c) ... return z

Now, we can call this function using rank 1 tensors (or lists that can be converted to rank 1 tensors): >>> tf.print('Rank 1 Inputs:', compute_z([1], [2], [3])) >>> tf.print('Rank 1 Inputs:', compute_z([1, 2], [2, 4], [3, 6])) [ 478 ]

Chapter 14

However, calling this function using tensors with ranks other than 1 will result in an error since the rank will not match the specified input signature, as follows: >>> tf.print('Rank 0 Inputs:', compute_z(1, 2, 3) ### will result in error >>> tf.print('Rank 2 Inputs:', compute_z([[1], [2]], ... [[2], [4]], ... [[3], [6]])) ### will result in error

In this section, we learned how to annotate a normal Python function so that TensorFlow will compile it into a graph for faster execution. Next, we will look at TensorFlow variables: how to create them and how to use them.

TensorFlow Variable objects for storing and updating model parameters

We covered Tensor objects in Chapter 13, Parallelizing Neural Network Training with TensorFlow. In the context of TensorFlow, a Variable is a special Tensor object that allows us to store and update the parameters of our models during training. A Variable can be created by just calling the tf.Variable class on user-specified initial values. In the following code, we will generate Variable objects of type float32, int32, bool, and string: >>> a = tf.Variable(initial_value=3.14, name='var_a') >>> print(a)

>>> b = tf.Variable(initial_value=[1, 2, 3], name='var_b') >>> print(b)

>>> c = tf.Variable(initial_value=[True, False], dtype=tf.bool) >>> print(c)

>>> d = tf.Variable(initial_value=['abc'], dtype=tf.string) >>> print(d)

[ 479 ]

Going Deeper – The Mechanics of TensorFlow

Notice that we always have to provide the initial values when creating a Variable. Variables have an attribute called trainable, which, by default, is set to True. Higher-level APIs such as Keras will use this attribute to manage the trainable variables and non-trainable ones. You can define a non-trainable Variable as follows: >>> w = tf.Variable([1, 2, 3], trainable=False) >>> print(w.trainable) False

The values of a Variable can be efficiently modified by running some operations such as .assign(), .assign_add() and related methods. Let's take a look at some examples: >>> print(w.assign([3, 1, 4], read_value=True))

>>> w.assign_add([2, -1, 2], read_value=False) >>> print(w.value()) tf.Tensor([5 0 6], shape=(3,), dtype=int32)

When the read_value argument is set to True (which is also the default), these operations will automatically return the new values after updating the current values of the Variable. Setting the read_value to False will suppress the automatic return of the updated value (but the Variable will still be updated in place). Calling w.value() will return the values in a tensor format. Note that we cannot change the shape or type of the Variable during assignment. You will recall that for NN models, initializing model parameters with random weights is necessary to break the symmetry during backpropagation—otherwise, a multilayer NN would be no more useful than a single-layer NN like logistic regression. When creating a TensorFlow Variable, we can also use a random initialization scheme. TensorFlow can generate random numbers based on a variety of distributions via tf.random (see https://www.tensorflow.org/versions/ r2.0/api_docs/python/tf/random). In the following example, we will take a look at some standard initialization methods that are also available in Keras (see https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/ initializers).

So, let's look at how we can create a Variable with Glorot initialization, which is a classic random initialization scheme that was proposed by Xavier Glorot and Yoshua Bengio. For this, we create an operator called init as an object of class GlorotNormal. Then, we call this operator and provide the desired shape of the output tensor: >>> tf.random.set_seed(1) >>> init = tf.keras.initializers.GlorotNormal() [ 480 ]

Chapter 14 >>> tf.print(init(shape=(3,))) [-0.722795904 1.01456821 0.251808226]

Now, we can use this operator to initialize a Variable of shape 2 × 3: >>> v = tf.Variable(init(shape=(2, 3))) >>> tf.print(v) [[0.28982234 -0.782292783 -0.0453658961] [0.960991383 -0.120003454 0.708528221]]

Xavier (or Glorot) initialization In the early development of deep learning, it was observed that random uniform or random normal weight initialization could often result in a poor performance of the model during training. In 2010, Glorot and Bengio investigated the effect of initialization and proposed a novel, more robust initialization scheme to facilitate the training of deep networks. The general idea behind Xavier initialization is to roughly balance the variance of the gradients across different layers. Otherwise, some layers may get too much attention during training while the other layers lag behind. According to the research paper by Glorot and Bengio, if we want to initialize the weights from uniform distribution, we should choose the interval of this uniform distribution as follows:

𝑊𝑊~𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈𝑈 (−

√6

√6

,

√𝑛𝑛𝑖𝑖𝑖𝑖 + 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜 √𝑛𝑛𝑖𝑖𝑖𝑖 + 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜

)

Here, 𝑛𝑛𝑖𝑖𝑖𝑖 is the number of input neurons that are multiplied by the weights, and 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜 is the number of output neurons that feed into the next layer. For initializing the weights from Gaussian (normal) distribution, it is recommended that you choose the standard deviation of this Gaussian to be 𝜎𝜎 =

√2

√𝑛𝑛𝑖𝑖𝑖𝑖 + 𝑛𝑛𝑜𝑜𝑜𝑜𝑜𝑜

.

TensorFlow supports Xavier initialization in both uniform and normal distributions of weights. For more information about Glorot and Bengio's initialization scheme, including the mathematical derivation and proof, read their original paper (Understanding the difficulty of deep feedforward neural networks, Xavier Glorot and Yoshua Bengio, 2010), which is freely available at http://proceedings.mlr.press/v9/ glorot10a/glorot10a.pdf.

[ 481 ]

Going Deeper – The Mechanics of TensorFlow

Now, to put this into the context of a more practical use case, let's see how we can define a Variable inside the base tf.Module class. We will define two variables: a trainable one and a non-trainable one: >>> ... ... ... ... ... ... >>> >>> All

class MyModule(tf.Module): def __init__(self): init = tf.keras.initializers.GlorotNormal() self.w1 = tf.Variable(init(shape=(2, 3)), trainable=True) self.w2 = tf.Variable(init(shape=(1, 2)), trainable=False) m = MyModule() print('All module variables:', [v.shape for v in m.variables]) module variables: [TensorShape([2, 3]), TensorShape([1, 2])]

>>> print('Trainable variable:', [v.shape for v in ... m.trainable_variables]) Trainable variable: [TensorShape([2, 3])]

As you can see in this code example, subclassing the tf.Module class gives us direct access to all variables defined in a given object (here, an instance of our custom MyModule class) via the .variables attribute. Finally, let's look at using variables inside a function decorated with tf.function. When we define a TensorFlow Variable inside a normal function (not decorated), we might expect that a new Variable will be created and initialized each time the function is called. However, tf.function will try to reuse the Variable based on tracing and graph creation. Therefore, TensorFlow does not allow the creation of a Variable inside a decorated function and, as a result, the following code will raise an error: >>> @tf.function ... def f(x): ... w = tf.Variable([1, 2, 3]) >>> f([1]) ValueError: tf.function-decorated function tried to create variables on non-first call.

One way to avoid this problem is to define the Variable outside of the decorated function and use it inside the function: >>> w = tf.Variable(tf.random.uniform((3, 3))) >>> @tf.function ... def compute_z(x): ... return tf.matmul(w, x) [ 482 ]

Chapter 14

>>> x = tf.constant([[1], [2], [3]], dtype=tf.float32) >>> tf.print(compute_z(x))

Computing gradients via automatic differentiation and GradientTape

As you already know, optimizing NNs requires computing the gradients of the cost with respect to the NN weights. This is required for optimization algorithms such as stochastic gradient descent (SGD). In addition, gradients have other applications, such as diagnosing the network to find out why an NN model is making a particular prediction for a test example. Therefore, in this section, we will cover how to compute gradients of a computation with respect to some variables.

Computing the gradients of the loss with respect to trainable variables

TensorFlow supports automatic differentiation, which can be thought of as an implementation of the chain rule for computing gradients of nested functions. When we define a series of operations that results in some output or even intermediate tensors, TensorFlow provides a context for calculating gradients of these computed tensors with respect to its dependent nodes in the computation graph. In order to compute these gradients, we have to "record" the computations via tf.GradientTape. Let's work with a simple example where we will compute 𝑧𝑧 = 𝑤𝑤𝑤𝑤 + 𝑏𝑏 and define the loss as the squared loss between the target and prediction, 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = (𝑦𝑦 − 𝑧𝑧)2. In the more general case, where we may have multiple predictions and targets, we compute 2 the loss as the sum of the squared error, 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = ∑(𝑦𝑦𝑖𝑖 − 𝑧𝑧𝑖𝑖 ) . In order to implement this 𝑖𝑖

computation in TensorFlow, we will define the model parameters, w and b, as variables, and the input, x and y, as tensors. We will place the computation of z and the loss within the tf.GradientTape context: >>> w = tf.Variable(1.0) >>> b = tf.Variable(0.5) >>> print(w.trainable, b.trainable) True True >>> x = tf.convert_to_tensor([1.4]) >>> y = tf.convert_to_tensor([2.1]) >>> with tf.GradientTape() as tape: ... z = tf.add(tf.multiply(w, x), b) [ 483 ]

Going Deeper – The Mechanics of TensorFlow ...

loss = tf.reduce_sum(tf.square(y – z))

>>> dloss_dw = tape.gradient(loss, w) >>> tf.print('dL/dw:', dloss_dw) dL/dw: -0.559999764

When computing the value z, we could think of the required operations, which we recorded to the "gradient tape," as a forward pass in an NN. We used 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 tape.gradient to compute . Since this is a very simple example, we can 𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 obtain the derivatives, = 2𝑥𝑥(𝑤𝑤𝑤𝑤 + 𝑏𝑏 − 𝑦𝑦), symbolically to verify that the 𝜕𝜕𝜕𝜕 computed gradients match the results we obtained in the previous code example: # verifying the computed gradient >>> tf.print(2*x*(w*x+b-y)) [-0.559999764]

Understanding automatic differentiation Automatic differentiation represents a set of computational techniques for computing derivatives or gradients of arbitrary arithmetic operations. During this process, gradients of a computation (expressed as a series of operations) are obtained by accumulating the gradients through repeated applications of the chain rule. To better understand the concept behind automatic differentiation, let's consider a series of computations, 𝑦𝑦 = 𝑓𝑓 (𝑔𝑔(ℎ(𝑥𝑥))), with input x and output y. This can be broken into a series of steps: • 𝑢𝑢0 = 𝑥𝑥

• • •

𝑢𝑢1 = ℎ(𝑥𝑥)

𝑢𝑢2 = 𝑔𝑔(𝑢𝑢1 )

𝑢𝑢3 = 𝑓𝑓(𝑢𝑢2 ) = 𝑦𝑦

𝑑𝑑𝑑𝑑 can be computed in two different ways: 𝑑𝑑𝑑𝑑 𝑑𝑑𝑢𝑢3 𝑑𝑑𝑢𝑢3 𝑑𝑑𝑢𝑢2 forward accumulation, which starts with , and = 𝑑𝑑𝑑𝑑 𝑑𝑑𝑢𝑢2 𝑑𝑑𝑢𝑢0 𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑 𝑑𝑑𝑢𝑢1 reverse accumulation, which starts with Note = 𝑑𝑑𝑢𝑢0 𝑑𝑑𝑢𝑢1 𝑑𝑑𝑢𝑢0

The derivative

that TensorFlow uses the latter, reverse accumulation. [ 484 ]

Chapter 14

Computing gradients with respect to nontrainable tensors

tf.GradientTape automatically supports the gradients for trainable variables. However, for non-trainable variables and other Tensor objects, we need to add an additional modification to the GradientTape called tape.watch() to monitor

those as well. For example, if we are interested in computing as follows: >>> with tf.GradientTape() as tape: ... tape.watch(x) ... z = tf.add(tf.multiply(w, x), b) ... loss = tf.reduce_sum(tf.square(y - z))

𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 , the code will be 𝜕𝜕𝜕𝜕

>>> dloss_dx = tape.gradient(loss, x) >>> tf.print('dL/dx:', dloss_dx) dL/dx: [-0.399999857]

Adversarial examples Computing gradients of the loss with respect to the input example is used for generating adversarial examples (or adversarial attacks). In computer vision, adversarial examples are examples that are generated by adding some small imperceptible noise (or perturbations) to the input example, which results in a deep NN misclassifying them. Covering adversarial examples is beyond the scope of this book, but if you are interested, you can find the original paper by Christian Szegedy et al., titled Intriguing properties of neural networks, at https://arxiv.org/pdf/1312.6199.pdf.

Keeping resources for multiple gradient computations

When we monitor the computations in the context of tf.GradientTape, by default, the tape will keep the resources only for a single gradient computation. For instance, after calling tape.gradient() once, the resources will be released and the tape will be cleared. Hence, if we want to compute more than one gradient, for example, both 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 and , we need to make the tape persistent: 𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕 >>> with tf.GradientTape(persistent=True) as tape: ... z = tf.add(tf.multiply(w, x), b) [ 485 ]

Going Deeper – The Mechanics of TensorFlow ...

loss = tf.reduce_sum(tf.square(y – z))

>>> dloss_dw = tape.gradient(loss, w) >>> tf.print('dL/dw:', dloss_dw) dL/dw: -0.559999764 >>> dloss_db = tape.gradient(loss, b) >>> tf.print('dL/db:', dloss_db) dL/db: -0.399999857

However, keep in mind that this is only needed when we want to compute more than one gradient, as recording and keeping the gradient tape is less memoryefficient compared to releasing the memory after a single gradient computation. This is also why the default setting is persistent=False. Finally, if we are computing gradients of a loss term with respect to the parameters of a model, we can define an optimizer and apply the gradients to optimize the model parameters using the tf.keras API, as follows: >>> optimizer = tf.keras.optimizers.SGD() >>> optimizer.apply_gradients(zip([dloss_dw, dloss_db], [w, b])) >>> tf.print('Updated w:', w) Updated w: 1.0056 >>> tf.print('Updated bias:', b) Updated bias: 0.504

You will recall that the initial weight and bias unit were w = 1.0 and b = 0.5, and applying the gradients of the loss with respect to the model parameters changed the model parameters to w = 1.0056 and b = 0.504.

Simplifying implementations of common architectures via the Keras API

You have already seen some examples of building a feedforward NN model (for instance, a multilayer perceptron) and defining a sequence of layers using Keras' Sequential class. Before we look at different approaches for configuring those layers, let's briefly recap the basic steps by building a model with two densely (fully) connected layers: >>> model = tf.keras.Sequential() >>> model.add(tf.keras.layers.Dense(units=16, activation='relu')) >>> model.add(tf.keras.layers.Dense(units=32, activation='relu'))

[ 486 ]

Chapter 14 >>> ## late variable creation >>> model.build(input_shape=(None, 4)) >>> model.summary() Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= dense (Dense) multiple 80 _________________________________________________________________ dense_1 (Dense) multiple 544 ================================================================= Total params: 624 Trainable params: 624 Non-trainable params: 0 _________________________________________________________________

We specified the input shape with model.build(), instantiating the variables after defining the model for that particular shape. The number of parameters of each layer is displayed: 16 × 4 + 16 = 80 for the first layer, and 16 × 32 + 32 = 544 for the second layer. Once variables (or model parameters) are created, we can access both trainable and non-trainable variables as follows: >>> ## printing variables of the model >>> for v in model.variables: ... print('{:20s}'.format(v.name), v.trainable, v.shape) dense/kernel:0 True (4, 16) dense/bias:0 True (16,) dense_1/kernel:0 True (16, 32) dense_1/bias:0 True (32,)

In this case, each layer has a weight matrix called kernel as well as a bias vector. Next, let's configure these layers, for example, by applying different activation functions, variable initializers, or regularization methods to the parameters. A comprehensive and complete list of available options for these categories can be found in the official documentation: • Choosing activation functions via tf.keras.activations: https://

www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/ activations

• Initializing the layer parameters via tf.keras.initializers: https:// www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/ initializers

• Applying regularization to the layer parameters (to prevent overfitting) via tf.keras.regularizers: https://www.tensorflow.org/versions/r2.0/ api_docs/python/tf/keras/regularizers [ 487 ]

Going Deeper – The Mechanics of TensorFlow

In the following code example, we will configure the first layer by specifying initializers for the kernel and bias variables. Then, we will configure the second layer by specifying an L1 regularizer for the kernel (weight matrix): >>> model = tf.keras.Sequential() >>> model.add( ... tf.keras.layers.Dense( ... units=16, ... activation=tf.keras.activations.relu, ... kernel_initializer= \ ... tf.keras.initializers.glorot_uniform(), ... bias_initializer=tf.keras.initializers.Constant(2.0) ... )) >>> model.add( ... tf.keras.layers.Dense( ... units=32, ... activation=tf.keras.activations.sigmoid, ... kernel_regularizer=tf.keras.regularizers.l1 ... ))

Furthermore, in addition to configuring the individual layers, we can also configure the model when we compile it. We can specify the type of optimizer and the loss function for training, as well as which metrics to use for reporting the performance on the training, validation, and test datasets. Again, a comprehensive list of all available options can be found in the official documentation: • Optimizers via tf.keras.optimizers: https://www.tensorflow.org/ versions/r2.0/api_docs/python/tf/keras/optimizers

• Loss functions via tf.keras.losses: https://www.tensorflow.org/ versions/r2.0/api_docs/python/tf/keras/losses

• Performance metrics via tf.keras.metrics: https://www.tensorflow. org/versions/r2.0/api_docs/python/tf/keras/metrics Choosing a loss function Regarding the choices for optimization algorithms, SGD and Adam are the most widely used methods. The choice of loss function depends on the task; for example, you might use mean square error loss for a regression problem. The family of cross-entropy loss functions supplies the possible choices for classification tasks, which are extensively discussed in Chapter 15, Classifying Images with Deep Convolutional Neural Networks. [ 488 ]

Chapter 14

Furthermore, you can use the techniques you have learned from previous chapters (for example, techniques for model evaluation from Chapter 6, Learning Best Practices for Model Evaluation and Hyperparameter Tuning) combined with the appropriate metrics for the problem. For example, precision and recall, accuracy, area under the curve (AUC), and false negative and false positive scores are appropriate metrics for evaluating classification models.

In this example, we will compile the model using the SGD optimizer, cross-entropy loss for binary classification, and a specific list of metrics, including accuracy, precision, and recall: >>> model.compile( ... optimizer=tf.keras.optimizers.SGD(learning_rate=0.001), ... loss=tf.keras.losses.BinaryCrossentropy(), ... metrics=[tf.keras.metrics.Accuracy(), ... tf.keras.metrics.Precision(), ... tf.keras.metrics.Recall(),])

When we train this model by calling model.fit(...), the history of the loss and the specified metrics for evaluating training and validation performance (if a validation dataset is used) will be returned, which can be used to diagnose the learning behavior. Next, we will look at a more practical example: solving the classic XOR classification problem using the Keras API. First, we will use the tf.keras.Sequential() class to build the model. Along the way, you will also learn about the capacity of a model for handling nonlinear decision boundaries. Then, we will cover other ways of building a model that will give us more flexibility and control over the layers of the network.

Solving an XOR classification problem

The XOR classification problem is a classic problem for analyzing the capacity of a model with regard to capturing the nonlinear decision boundary between two classes. We generate a toy dataset of 200 training examples with two features (𝑥𝑥0 , 𝑥𝑥1 ) drawn from a uniform distribution between [−1, 1). Then, we assign the ground truth label for training example i according to the following rule: (𝑖𝑖)

(𝑖𝑖)

𝑦𝑦 (𝑖𝑖) = { 0 if 𝑥𝑥0 × 𝑥𝑥1 < 0 1 otherwise

[ 489 ]

Going Deeper – The Mechanics of TensorFlow

We will use half of the data (100 training examples) for training and the remaining half for validation. The code for generating the data and splitting it into the training and validation datasets is as follows: >>> import tensorflow as tf >>> import numpy as np >>> import matplotlib.pyplot as plt >>> tf.random.set_seed(1) >>> np.random.seed(1) >>> x = np.random.uniform(low=-1, high=1, size=(200, 2)) >>> y = np.ones(len(x)) >>> y[x[:, 0] * x[:, 1]>> >>> >>> >>>

x_train y_train x_valid y_valid

= = = =

x[:100, :] y[:100] x[100:, :] y[100:]

>>> >>> ... >>> ... >>> >>> >>>

fig = plt.figure(figsize=(6, 6)) plt.plot(x[y==0, 0], x[y==0, 1], 'o', alpha=0.75, markersize=10) plt.plot(x[y==1, 0], x[y==1, 1], '>> tf.random.set_seed(1) >>> ## input layer: >>> inputs = tf.keras.Input(shape=(2,)) >>> >>> >>> >>>

## h1 h2 h3

hidden layers = tf.keras.layers.Dense(units=4, activation='relu')(inputs) = tf.keras.layers.Dense(units=4, activation='relu')(h1) = tf.keras.layers.Dense(units=4, activation='relu')(h2)

>>> ## output: >>> outputs = tf.keras.layers.Dense(units=1, activation='sigmoid')(h3) >>> ## construct a model: >>> model = tf.keras.Model(inputs=inputs, outputs=outputs) >>> model.summary()

Compiling and training this model is similar to what we did previously: >>> ## compile: >>> model.compile( ... optimizer=tf.keras.optimizers.SGD(), ... loss=tf.keras.losses.BinaryCrossentropy(), ... metrics=[tf.keras.metrics.BinaryAccuracy()]) >>> ## train: >>> hist = model.fit( ... x_train, y_train, ... validation_data=(x_valid, y_valid), ... epochs=200, batch_size=2, verbose=0)

[ 495 ]

Going Deeper – The Mechanics of TensorFlow

Implementing models based on Keras' Model class

An alternative way to build complex models is by subclassing tf.keras.Model. In this approach, we create a new class derived from tf.keras.Model and define the function, __init__(), as a constructor. The call() method is used to specify the forward pass. In the constructor function, __init__(), we define the layers as attributes of the class so that they can be accessed via the self reference attribute. Then, in the call() method, we specify how these layers are to be used in the forward pass of the NN. The code for defining a new class that implements the previous model is as follows: >>> class MyModel(tf.keras.Model): ... def __init__(self): ... super(MyModel, self).__init__() ... self.hidden_1 = tf.keras.layers.Dense( ... units=4, activation='relu') ... self.hidden_2 = tf.keras.layers.Dense( ... units=4, activation='relu') ... self.hidden_3 = tf.keras.layers.Dense( ... units=4, activation='relu') ... self.output_layer = tf.keras.layers.Dense( ... units=1, activation='sigmoid') ... ... def call(self, inputs): ... h = self.hidden_1(inputs) ... h = self.hidden_2(h) ... h = self.hidden_3(h) ... return self.output_layer(h)

Notice that we used the same output name, h, for all hidden layers. This makes the code more readable and easier to follow. A model class derived from tf.keras.Model through subclassing inherits general model attributes, such as build(), compile(), and fit(). Therefore, once we define an instance of this new class, we can compile and train it like any other model built by Keras: >>> tf.random.set_seed(1) >>> model = MyModel() >>> model.build(input_shape=(None, 2)) >>> model.summary()

[ 496 ]

Chapter 14

>>> ## compile: >>> model.compile(optimizer=tf.keras.optimizers.SGD(), ... loss=tf.keras.losses.BinaryCrossentropy(), ... metrics=[tf.keras.metrics.BinaryAccuracy()]) >>> ## train: >>> hist = model.fit(x_train, y_train, ... validation_data=(x_valid, y_valid), ... epochs=200, batch_size=2, verbose=0)

Writing custom Keras layers

In cases where we want to define a new layer that is not already supported by Keras, we can define a new class derived from the tf.keras.layers.Layer class. This is especially useful when designing a new layer or customizing an existing layer. To illustrate the concept of implementing custom layers, let's consider a simple example. Imagine we want to define a new linear layer that computes 𝑤𝑤(𝑥𝑥 + 𝜖𝜖) + 𝑏𝑏, where 𝜖𝜖 refers to a random variable as a noise variable. To implement this computation, we define a new class as a subclass of tf.keras.layers.Layer. For this new class, we have to define both the constructor __init__() method and the call() method. In the constructor, we define the variables and other required tensors for our customized layer. We have the option to create variables and initialize them in the constructor if the input_shape is given to the constructor. Alternatively, we can delay the variable initialization (for instance, if we do not know the exact input shape upfront) and delegate it to the build() method for late variable creation. In addition, we can define get_config() for serialization, which means that a model using our custom layer can be efficiently saved using TensorFlow's model saving and loading capabilities. To look at a concrete example, we are going to define a new layer called NoisyLinear, which implements the computation 𝑤𝑤(𝑥𝑥 + 𝜖𝜖) + 𝑏𝑏 , which was mentioned in the preceding paragraph:

>>> class NoisyLinear(tf.keras.layers.Layer): ... def __init__(self, output_dim, noise_stddev=0.1, **kwargs): ... self.output_dim = output_dim ... self.noise_stddev = noise_stddev ... super(NoisyLinear, self).__init__(**kwargs) ... ... def build(self, input_shape): ... self.w = self.add_weight(name='weights', ... shape=(input_shape[1], [ 497 ]

Going Deeper – The Mechanics of TensorFlow ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...

self.output_dim), initializer='random_normal', trainable=True) self.b = self.add_weight(shape=(self.output_dim,), initializer='zeros', trainable=True) def call(self, inputs, training=False): if training: batch = tf.shape(inputs)[0] dim = tf.shape(inputs)[1] noise = tf.random.normal(shape=(batch, dim), mean=0.0, stddev=self.noise_stddev) noisy_inputs = tf.add(inputs, noise) else: noisy_inputs = inputs z = tf.matmul(noisy_inputs, self.w) + self.b return tf.keras.activations.relu(z) def get_config(self): config = super(NoisyLinear, self).get_config() config.update({'output_dim': self.output_dim, 'noise_stddev': self.noise_stddev}) return config

In the constructor, we have added an argument, noise_stddev, to specify the standard deviation for the distribution of 𝜖𝜖 , which is sampled from a Gaussian distribution. Furthermore, notice that in the call() method, we have used an additional argument, training=False. In the context of Keras, the training argument is a special Boolean argument that distinguishes whether a model or layer is used during training (for example, via fit()) or only for prediction (for example, via predict(); this is sometimes also called "inference" or evaluation). One of the main differences between training and prediction is that during prediction, we do not require gradients. Also, there are certain methods that behave differently in training and prediction modes. You will encounter an example of such a method, Dropout, in the upcoming chapters. In the previous code snippet, we also specified that the random vector, 𝜖𝜖, was to be generated and added to the input during training only and not used for inference or evaluation. Before we go a step further and use our custom NoisyLinear layer in a model, let's test it in the context of a simple example. [ 498 ]

Chapter 14

In the following code, we will define a new instance of this layer, initialize it by calling .build(), and execute it on an input tensor. Then, we will serialize it via .get_config() and restore the serialized object via .from_config(): >>> tf.random.set_seed(1) >>> noisy_layer = NoisyLinear(4) >>> noisy_layer.build(input_shape=(None, 4)) >>> x = tf.zeros(shape=(1, 4)) >>> tf.print(noisy_layer(x, training=True)) [[0 0.00821428 0 0]] >>> >>> >>> >>> [[0

## re-building from config: config = noisy_layer.get_config() new_layer = NoisyLinear.from_config(config) tf.print(new_layer(x, training=True)) 0.0108502861 0 0]]

In the previous code snippet, we called the layer two times on the same input tensor. However, note that the outputs differ because the NoisyLinear layer added random noise to the input tensor. Now, let's create a new model similar to the previous one for solving the XOR classification task. As before, we will use Keras' Sequential class, but this time, we will use our NoisyLinear layer as the first hidden layer of the multilayer perceptron. The code is as follows: >>> tf.random.set_seed(1) >>> model = tf.keras.Sequential([ ... NoisyLinear(4, noise_stddev=0.1), ... tf.keras.layers.Dense(units=4, activation='relu'), ... tf.keras.layers.Dense(units=4, activation='relu'), ... tf.keras.layers.Dense(units=1, activation='sigmoid')]) >>> model.build(input_shape=(None, 2)) >>> model.summary() >>> ## compile: >>> model.compile(optimizer=tf.keras.optimizers.SGD(), ... loss=tf.keras.losses.BinaryCrossentropy(), ... metrics=[tf.keras.metrics.BinaryAccuracy()]) >>> ## train: >>> hist = model.fit(x_train, y_train, [ 499 ]

Going Deeper – The Mechanics of TensorFlow ... ... ...

validation_data=(x_valid, y_valid), epochs=200, batch_size=2, verbose=0)

>>> ## Plotting >>> history = hist.history >>> >>> >>> >>> >>> >>>

fig = plt.figure(figsize=(16, 4)) ax = fig.add_subplot(1, 3, 1) plt.plot(history['loss'], lw=4) plt.plot(history['val_loss'], lw=4) plt.legend(['Train loss', 'Validation loss'], fontsize=15) ax.set_xlabel('Epochs', size=15)

>>> >>> >>> >>> >>>

ax = fig.add_subplot(1, 3, 2) plt.plot(history['binary_accuracy'], lw=4) plt.plot(history['val_binary_accuracy'], lw=4) plt.legend(['Train Acc.', 'Validation Acc.'], fontsize=15) ax.set_xlabel('Epochs', size=15)

>>> >>> ... >>> >>> >>> >>> >>>

ax = fig.add_subplot(1, 3, 3) plot_decision_regions(X=x_valid, y=y_valid.astype(np.integer), clf=model) ax.set_xlabel(r'$x_1$', size=15) ax.xaxis.set_label_coords(1, -0.025) ax.set_ylabel(r'$x_2$', size=15) ax.yaxis.set_label_coords(-0.025, 1) plt.show()

The resulting figure will be as follows:

[ 500 ]

Chapter 14

Here, our goal was to learn how to define a new custom layer subclassed from tf.keras.layers.Layer and to use it as we would use any other standard Keras layer. Although, with this particular example, NoisyLinear did not help to improve the performance, please keep in mind that our objective was to mainly learn how to write a customized layer from scratch. In general, writing a new customized layer can be useful in other applications, for example, if you develop a new algorithm that depends on a new layer beyond the existing ones.

TensorFlow Estimators

So far, in this chapter, we have mostly focused on the low-level TensorFlow API. We used decorators to modify functions to compile the computational graphs explicitly for computational efficiency. Then, we worked with the Keras API and implemented feedforward NNs, to which we added customized layers. In this section, we will switch gears and work with TensorFlow Estimators. The tf.estimator API encapsulates the underlying steps in machine learning tasks, such as training, prediction (inference), and evaluation. Estimators are more encapsulated but also more scalable when compared to the previous approaches that we have covered in this chapter. Also, the tf.estimator API adds support for running models on multiple platforms without requiring major code changes, which makes them more suitable for the so-called "production phase" in industry applications. In addition, TensorFlow comes with a selection of off-the-shelf estimators for common machine learning and deep learning architectures that are useful for comparison studies, for example, to quickly assess whether a certain approach is applicable to a particular dataset or problem. In the remaining sections of this chapter, you will learn how to use such pre-made Estimators and how to create an Estimator from an existing Keras model. One of the essential elements of Estimators is defining the feature columns as a mechanism for importing data into an Estimator-based model, which we will cover in the next section.

Working with feature columns

In machine learning and deep learning applications, we can encounter various different types of features: continuous, unordered categorical (nominal), and ordered categorical (ordinal). You will recall that in Chapter 4, Building Good Training Datasets – Data Preprocessing, we covered different types of features and learned how to handle each type. Note that while numeric data can be either continuous or discrete, in the context of the TensorFlow API, "numeric" data specifically refers to continuous data of the floating point type.

[ 501 ]

Going Deeper – The Mechanics of TensorFlow

Sometimes, feature sets are comprised of a mixture of different feature types. While TensorFlow Estimators were designed to handle all these different types of features, we must specify how each feature should be interpreted by the Estimator. For example, consider a scenario with a set of seven different features, as shown in the following figure:

The features shown in the figure (model year, cylinders, displacement, horsepower, weight, acceleration, and origin) were obtained from the Auto MPG dataset, which is a common machine learning benchmark dataset for predicting the fuel efficiency of a car in miles per gallon (MPG). The full dataset and its description are available from UCI's machine learning repository at https://archive.ics.uci.edu/ml/ datasets/auto+mpg. We are going to treat five features from the Auto MPG dataset (number of cylinders, displacement, horsepower, weight, and acceleration) as "numeric" (here, continuous) features. The model year can be regarded as an ordered categorical (ordinal) feature. Lastly, the manufacturing origin can be regarded as an unordered categorical (nominal) feature with three possible discrete values, 1, 2, and 3, which correspond to the US, Europe, and Japan, respectively.

[ 502 ]

Chapter 14

Let's first load the data and apply the necessary preprocessing steps, such as partitioning the dataset into training and test datasets, as well as standardizing the continuous features: >>> import pandas as pd >>> dataset_path = tf.keras.utils.get_file( ... "auto-mpg.data", ... ("http://archive.ics.uci.edu/ml/machine-learning" ... "-databases/auto-mpg/auto-mpg.data")) >>> column_names = [ ... 'MPG', 'Cylinders', 'Displacement', ... 'Horsepower', 'Weight', 'Acceleration', ... 'ModelYear', 'Origin'] >>> ... ... >>> >>> >>>

df = pd.read_csv(dataset_path, names=column_names, na_values = '?', comment='\t', sep=' ', skipinitialspace=True) ## drop the NA rows df = df.dropna() df = df.reset_index(drop=True)

>>> ## train/test splits: >>> import sklearn >>> import sklearn.model_selection >>> df_train, df_test = sklearn.model_selection.train_test_split( ... df, train_size=0.8) >>> train_stats = df_train.describe().transpose() >>> numeric_column_names = [ ... 'Cylinders', 'Displacement', ... 'Horsepower', 'Weight', ... 'Acceleration'] >>> df_train_norm, df_test_norm = df_train.copy(), df_test.copy() >>> for col_name in numeric_column_names: ... mean = train_stats.loc[col_name, 'mean'] ... std = train_stats.loc[col_name, 'std'] ... df_train_norm.loc[:, col_name] = ( ... df_train_norm.loc[:, col_name] - mean)/std ... df_test_norm.loc[:, col_name] = (

[ 503 ]

Going Deeper – The Mechanics of TensorFlow ...

df_test_norm.loc[:, col_name] - mean)/std

>>> df_train_norm.tail()

This results in the following:

The pandas DataFrame that we created via the previous code snippet contains five columns with values of the type float. These columns will constitute the continuous features. In the following code, we will use TensorFlow's feature_column function to transform these continuous features into the feature column data structure that TensorFlow Estimators can work with: >>> numeric_features = [] >>> for col_name in numeric_column_names: ... numeric_features.append( ... tf.feature_column.numeric_column(key=col_name))

Next, let's group the rather fine-grained model year information into buckets to simplify the learning task for the model that we are going to train later. Concretely, we are going to assign each car into one of four "year" buckets, as follows:

0 1 bucket = { 2 3

if year < 73 if 73 ≤ year ≤ 76 if 76 ≤ year ≤ 79 if year ≥ 79

Note that the chosen intervals were selected arbitrarily to illustrate the concepts of "bucketing." In order to group the cars into these buckets, we will first define a numeric feature based on each original model year. Then, these numeric features will be passed to the bucketized_column function for which we will specify three interval cut-off values: [73, 76, 79]. The specified values include the right cut-off value. These cut-off values are used to specify half-closed intervals, for instance, (−∞, 73), [73, 76), [76, 79), and [79, ∞). The code is as follows: [ 504 ]

Chapter 14 >>> feature_year = tf.feature_column.numeric_column(key='ModelYear') >>> bucketized_features = [] >>> bucketized_features.append( ... tf.feature_column.bucketized_column( ... source_column=feature_year, ... boundaries=[73, 76, 79]))

For consistency, we added this bucketized feature column to a Python list, even though the list consists of only one entry. In the following steps, we will merge this list with the lists made from other features, which will then be provided as input to the TensorFlow Estimator-based model. Next, we will proceed with defining a list for the unordered categorical feature, Origin. In TensorFlow, there are different ways of creating a categorical feature column. If the data contains the category names (for example, in string format like "US," "Europe," and "Japan"), then we can use tf.feature_column.categorical_ column_with_vocabulary_list and provide a list of unique, possible category names as input. If the list of possible categories is too large, for example, in a typical text analysis context, then we can use tf.feature_column.categorical_column_ with_vocabulary_file instead. When using this function, we simply provide a file that contains all the categories/words so that we do not have to store a list of all possible words in memory. Moreover, if the features are already associated with an index of categories in the range [0, num_categories), then we can use the tf.feature_column.categorical_column_with_identity function. However, in this case, the feature Origin is given as integer values 1, 2, 3 (as opposed to 0, 1, 2), which does not match the requirement for categorical indexing, as it expects the indices to start from 0. In the following code example, we will proceed with the vocabulary list: >>> feature_origin = tf.feature_column.categorical_column_with_ vocabulary_list( ... key='Origin', ... vocabulary_list=[1, 2, 3])

Certain Estimators, such as DNNClassifier and DNNRegressor, only accept so-called "dense columns." Therefore, the next step is to convert the existing categorical feature column to such a dense column. There are two ways to do this: using an embedding column via embedding_column or an indicator column via indicator_column. An indicator column converts the categorical indices to one-hot encoded vectors, for example, index 0 will be encoded as [1, 0, 0], index 1 will be encoded as [0, 1, 0], and so on. On the other hand, the embedding column maps each index to a vector of random number of the type float, which can be trained.

[ 505 ]

Going Deeper – The Mechanics of TensorFlow

When the number of categories is large, using the embedding column with fewer dimensions than the number of categories can improve the performance. In the following code snippet, we will use the indicator column approach on the categorical feature in order to convert it into the dense format: >>> categorical_indicator_features = [] >>> categorical_indicator_features.append( ... tf.feature_column.indicator_column(feature_origin))

In this section, we have covered the most common approaches for creating feature columns that can be used with TensorFlow Estimators. However, there are several additional feature columns that we haven't discussed, including hashed columns and crossed columns. More information about these other feature columns can be found in the official TensorFlow documentation at https://www.tensorflow.org/ versions/r2.0/api_docs/python/tf/feature_column.

Machine learning with pre-made Estimators Now, after constructing the mandatory feature columns, we can finally utilize TensorFlow's Estimators. Using pre-made Estimators can be summarized in four steps: 1. Define an input function for data loading 2. Convert the dataset into feature columns 3. Instantiate an Estimator (use a pre-made Estimator or create a new one, for example, by converting a Keras model into an Estimator) 4. Use the Estimator methods train(), evaluate(), and predict() Continuing with the Auto MPG example from the previous section, we will apply these four steps to illustrate how we can use Estimators in practice. For the first step, we need to define a function that processes the data and returns a TensorFlow dataset consisting of a tuple that contains the input features and the labels (ground truth MPG values). Note that the features must be in a dictionary format, and the keys of the dictionary must match the feature columns' names. Starting with the first step, we will define the input function for the training data as follows: >>> def train_input_fn(df_train, batch_size=8): ... df = df_train.copy() ... train_x, train_y = df, df.pop('MPG') ... dataset = tf.data.Dataset.from_tensor_slices( ... (dict(train_x), train_y)) ... [ 506 ]

Chapter 14 ... ...

# shuffle, repeat, and batch the examples. return dataset.shuffle(1000).repeat().batch(batch_size)

Notice that we used dict(train_x) in this function to convert the pandas DataFrame object into a Python dictionary. Let's load a batch from this dataset to see how it looks: >>> ds = train_input_fn(df_train_norm) >>> batch = next(iter(ds)) >>> print('Keys:', batch[0].keys()) Keys: dict_keys(['Cylinders', 'Displacement', 'Horsepower', 'Weight', 'Acceleration', 'ModelYear', 'Origin']) >>> print('Batch Model Years:', batch[0]['ModelYear']) Batch Model Years: tf.Tensor([74 71 81 72 82 81 70 74], shape=(8,), dtype=int32)

We also need to define an input function for the test dataset that will be used for evaluation after model training: >>> def eval_input_fn(df_test, batch_size=8): ... df = df_test.copy() ... test_x, test_y = df, df.pop('MPG') ... dataset = tf.data.Dataset.from_tensor_slices( ... (dict(test_x), test_y)) ... return dataset.batch(batch_size)

Now, moving on to step 2, we need to define the feature columns. We have already defined a list containing the continuous features, a list for the bucketized feature column, and a list for the categorical feature column. We can now concatenate these individual lists to a single list containing all feature columns: >>> all_feature_columns = ( ... numeric_features + ... bucketized_features + ... categorical_indicator_features)

For step 3, we need to instantiate a new Estimator. Since predicting MPG values is a typical regression problem, we will use tf.estimator.DNNRegressor. When instantiating the regression Estimator, we will provide the list of feature columns and specify the number of hidden units that we want to have in each hidden layer using the argument hidden_units. Here, we will use two hidden layers, where the first hidden layer has 32 units and the second hidden layer has 10 units: >>> regressor = tf.estimator.DNNRegressor( ... feature_columns=all_feature_columns, ... hidden_units=[32, 10], ... model_dir='models/autompg-dnnregressor/') [ 507 ]

Going Deeper – The Mechanics of TensorFlow

The other argument, model_dir, that we have provided specifies the directory for saving model parameters. One of the advantages of Estimators is that they automatically checkpoint the model during training, so that in case the training of the model crashes for an unexpected reason (like power failure), we can easily load the last saved checkpoint and continue training from there. The checkpoints will also be saved in the directory specified by model_dir. If we do not specify the model_dir argument, the Estimator will create a random temporary folder (for example, in the Linux operating system, a random folder in the /tmp/ directory will be created), which will be used for this purpose. After these three basic setup steps, we can finally use the Estimator for training, evaluation, and, eventually, prediction. The regressor can be trained by calling the train() method, for which we require the previously defined input function: >>> EPOCHS = 1000 >>> BATCH_SIZE = 8 >>> total_steps = EPOCHS * int(np.ceil(len(df_train) / BATCH_SIZE)) >>> print('Training Steps:', total_steps) Training Steps: 40000 >>> regressor.train( ... input_fn=lambda:train_input_fn( ... df_train_norm, batch_size=BATCH_SIZE), ... steps=total_steps)

Calling .train() will automatically save the checkpoints during the training of the model. We can then reload the last checkpoint: >>> reloaded_regressor = tf.estimator.DNNRegressor( ... feature_columns=all_feature_columns, ... hidden_units=[32, 10], ... warm_start_from='models/autompg-dnnregressor/', ... model_dir='models/autompg-dnnregressor/')

Then, in order to evaluate the predictive performance of the trained model, we can use the evaluate() method, as follows: >>> eval_results = reloaded_regressor.evaluate( ... input_fn=lambda:eval_input_fn(df_test_norm, batch_size=8)) >>> print('Average-Loss {:.4f}'.format( ... eval_results['average_loss'])) Average-Loss 15.1866

Finally, to predict the target values on new data points, we can use the predict() method. For the purposes of this example, suppose that the test dataset represents a dataset of new, unlabeled data points in a real-world application. [ 508 ]

Chapter 14

Note that in a real-world prediction task, the input function will only need to return a dataset consisting of features, assuming that the labels are not available. Here, we will simply use the same input function that we used for evaluation to get the predictions for each example: >>> pred_res = regressor.predict( ... input_fn=lambda: eval_input_fn( ... df_test_norm, batch_size=8)) >>> print(next(iter(pred_res))) {'predictions': array([23.747658], dtype=float32)}

While the preceding code snippets conclude the illustration of the four steps that are required for using pre-made Estimators, for practice, let's take a look at another premade Estimator: the boosted tree regressor, tf.estimator.BoostedTreeRegressor. Since, the input functions and the feature columns are already built, we just need to repeat steps 3 and 4. For step 3, we will create an instance of BoostedTreeRegressor and configure it to have 200 trees. Decision tree boosting We already covered the ensemble algorithms, including boosting, in Chapter 7, Combining Different Models for Ensemble Learning. The boosted tree algorithm is a special family of boosting algorithms that is based on the optimization of an arbitrary loss function. Feel free to visit https://medium.com/mlreview/gradientboosting-from-scratch-1e317ae4587d to learn more. >>> boosted_tree = tf.estimator.BoostedTreesRegressor( ... feature_columns=all_feature_columns, ... n_batches_per_layer=20, ... n_trees=200) >>> boosted_tree.train( ... input_fn=lambda:train_input_fn( ... df_train_norm, batch_size=BATCH_SIZE)) >>> eval_results = boosted_tree.evaluate( ... input_fn=lambda:eval_input_fn( ... df_test_norm, batch_size=8)) >>> print('Average-Loss {:.4f}'.format( ... eval_results['average_loss'])) Average-Loss 11.2609

As you can see, the boosted tree regressor achieves lower average loss than the DNNRegressor. For a small dataset like this, this is expected. [ 509 ]

Going Deeper – The Mechanics of TensorFlow

In this section, we covered the essential steps for using TensorFlow's Estimators for regression. In the next subsection, we will take a look at a typical classification example using Estimators.

Using Estimators for MNIST handwritten digit classification

For this classification problem, we are going to use the DNNClassifier Estimator provided by TensorFlow, which lets us implement a multilayer perceptron very conveniently. In the previous section, we covered the four essential steps for using the pre-made Estimators in detail, which we will need to repeat in this section. First, we are going to import the tensorflow_datasets (tfds) submodule, which we can use to load the MNIST dataset and specify the hyperparameters of the model. Estimator API and graph issues Since parts of TensorFlow 2.0 are still a bit rough around the edges, you may encounter the following issue when executing the next code block: RuntimeError: Graph is finalized and cannot be modified. Currently, there is no good solution for this issue, and a suggested workaround is to restart your Python, IPython, or Jupyter Notebook session before executing the next code block.

The setup step includes loading the dataset and specifying hyperparameters (BUFFER_SIZE for shuffling the dataset, BATCH_SIZE for the size of mini-batches, and the number of training epochs): >>> import tensorflow_datasets as tfds >>> import tensorflow as tf >>> import numpy as np >>> >>> >>> >>>

BUFFER_SIZE = 10000 BATCH_SIZE = 64 NUM_EPOCHS = 20 steps_per_epoch = np.ceil(60000 / BATCH_SIZE)

Note that steps_per_epoch determines the number of iterations in each epoch, which is needed for infinitely repeated datasets (as discussed in Chapter 13, Parallelizing Neural Network Training with TensorFlow. Next, we will define a helper function that will preprocess the input image and its label.

[ 510 ]

Chapter 14

Since the input image is originally of the type 'uint8' (in the range [0, 255]), we will use tf.image.convert_image_dtype() to convert its type to tf.float32 (and thereby, within the range [0, 1]): >>> def preprocess(item): ... image = item['image'] ... label = item['label'] ... image = tf.image.convert_image_dtype( ... image, tf.float32) ... image = tf.reshape(image, (-1,)) ... ... return {'image-pixels':image}, label[..., tf.newaxis]

Step 1: Define two input functions (one for training and one for evaluation): >>> ## Step 1: Define the input functions >>> def train_input_fn(): ... datasets = tfds.load(name='mnist') ... mnist_train = datasets['train'] ... ... dataset = mnist_train.map(preprocess) ... dataset = dataset.shuffle(BUFFER_SIZE) ... dataset = dataset.batch(BATCH_SIZE) ... return dataset.repeat() >>> def eval_input_fn(): ... datasets = tfds.load(name='mnist') ... mnist_test = datasets['test'] ... dataset = mnist_test.map(preprocess).batch(BATCH_SIZE) ... return dataset

Notice that the dictionary of features has only one key, 'image-pixels'. We will use this key in the next step. Step 2: Define the feature columns: >>> ## Step 2: feature columns >>> image_feature_column = tf.feature_column.numeric_column( ... key='image-pixels', shape=(28*28))

Note that here, we defined the feature columns of size 784 (that is, 28 × 28 ), which is the size of the input MNIST images after they are flattened. Step 3: Create a new Estimator. Here, we specify two hidden layers: 32 units in the first hidden layer and 16 units in the second.

[ 511 ]

Going Deeper – The Mechanics of TensorFlow

We also specify the number of classes (remember that MNIST consists of 10 different digits, 0-9) using the argument n_classes: >>> ## Step 3: instantiate the estimator >>> dnn_classifier = tf.estimator.DNNClassifier( ... feature_columns=[image_feature_column], ... hidden_units=[32, 16], ... n_classes=10, ... model_dir='models/mnist-dnn/')

Step 4: Use the Estimator for training, evaluation, and prediction: >>> ## Step 4: train and evaluate >>> dnn_classifier.train( ... input_fn=train_input_fn, ... steps=NUM_EPOCHS * steps_per_epoch) >>> eval_result = dnn_classifier.evaluate( ... input_fn=eval_input_fn) >>> print(eval_result) {'accuracy': 0.8957, 'average_loss': 0.3876346, 'loss': 0.38815108, 'global_step': 18760}

So far, you have learned how to use pre-made Estimators and apply them for preliminary assessment to see, for example, whether an existing model is suitable for a particular problem. Besides using pre-made Estimators, we can also create an Estimator by converting a Keras model to an Estimator, which we will do in the next subsection.

Creating a custom Estimator from an existing Keras model Converting a Keras model to an Estimator is useful in academia as well as industry for cases where you have developed a model and want to publish it or share the model with other members in your organization. Such a conversion allows us to access the strengths of Estimators, such as distributed training and automatic checkpointing. In addition, it will make it easy for others to use this model, and particularly to avoid confusions in interpreting the input features by specifying the feature columns and the input function. To learn how we can create our own Estimator from a Keras model, we will work with the previous XOR problem. First, we will regenerate the data and split it into training and validation datasets: [ 512 ]

Chapter 14 >>> tf.random.set_seed(1) >>> np.random.seed(1) >>> >>> >>> >>>

## Create the data x = np.random.uniform(low=-1, high=1, size=(200, 2)) y = np.ones(len(x)) y[x[:, 0] * x[:, 1]>> >>> >>> >>>

x_train y_train x_valid y_valid

= = = =

x[:100, :] y[:100] x[100:, :] y[100:]

Let's also build a Keras model that we want to convert to an Estimator later. We will define the model using the Sequential class as before. This time, we will also add an input layer defined as tf.keras.layers.Input to give a name to the input to this model: >>> model = tf.keras.Sequential([ ... tf.keras.layers.Input(shape=(2,), name='input-features'), ... tf.keras.layers.Dense(units=4, activation='relu'), ... tf.keras.layers.Dense(units=4, activation='relu'), ... tf.keras.layers.Dense(units=4, activation='relu'), ... tf.keras.layers.Dense(1, activation='sigmoid') ... ])

Next, we will go through the four steps that we described in the previous subsection. Steps 1, 2, and 4 will be the same as the ones we used with the pre-made estimators. Note that the key name for the input features that we use in steps 1 and 2 must match with what we defined in the input layer of our model. The code is as follows: >>> ## Step 1: Define the input functions >>> def train_input_fn(x_train, y_train, batch_size=8): ... dataset = tf.data.Dataset.from_tensor_slices( ... ({'input-features':x_train}, y_train.reshape(-1, 1))) ... ... # shuffle, repeat, and batch the examples. ... return dataset.shuffle(100).repeat().batch(batch_size) >>> def eval_input_fn(x_test, y_test=None, batch_size=8): ... if y_test is None: ... dataset = tf.data.Dataset.from_tensor_slices( ... {'input-features':x_test}) ... else: ... dataset = tf.data.Dataset.from_tensor_slices(

[ 513 ]

Going Deeper – The Mechanics of TensorFlow ... ... ... ... ...

({'input-features':x_test}, y_test.reshape(-1, 1)))

# shuffle, repeat, and batch the examples. return dataset.batch(batch_size)

>>> ## Step 2: Define the feature columns >>> features = [ ... tf.feature_column.numeric_column( ... key='input-features:', shape=(2,)) ... ]

For step 3, we will convert the model to an Estimator using tf.keras.estimator.model_to_estimator instead of instantiating one of the pre-made Estimators. Before converting the model, we first need to compile it: >>> model.compile(optimizer=tf.keras.optimizers.SGD(), ... loss=tf.keras.losses.BinaryCrossentropy(), ... metrics=[tf.keras.metrics.BinaryAccuracy()]) >>> my_estimator = tf.keras.estimator.model_to_estimator( ... keras_model=model, ... model_dir='models/estimator-for-XOR/')

Finally, in step 4, we can train our model using the Estimator and evaluate it on the validation dataset: >>> >>> >>> >>>

## Step 4: Use the estimator num_epochs = 200 batch_size = 2 steps_per_epoch = np.ceil(len(x_train) / batch_size)

>>> my_estimator.train( ... input_fn=lambda: train_input_fn(x_train, y_train, batch_size), ... steps=num_epochs * steps_per_epoch) >>> my_estimator.evaluate( ... input_fn=lambda: eval_input_fn(x_valid, y_valid, batch_size)) {'binary_accuracy': 0.96, 'loss': 0.081909806, 'global_step': 10000}

As you can see, converting a Keras model to an Estimator is very straightforward. Doing this allows us to easily benefit from the various Estimator strengths, such as distributed training and automatically saving the checkpoints during training.

[ 514 ]

Chapter 14

Summary

In this chapter, we covered TensorFlow's most essential and useful features. We started by discussing the migration from TensorFlow v1.x to v2. In particular, we used TensorFlow's dynamic computation graph approach, the so-called eager execution mode, which makes implementing computations more convenient compared to using static graphs. We also covered the semantics of defining TensorFlow Variable objects as model parameters, annotating Python functions using the tf.function decorator to improve computational efficiency via graph compilation. After we considered the concept of computing partial derivatives and gradients of arbitrary functions, we covered the Keras API in more detail. It provides us with a user-friendly interface for building more complex deep NN models. Finally, we utilized TensorFlow's tf.estimator API to provide a consistent interface that is typically preferred in production environments. We concluded this chapter by converting a Keras model into a custom Estimator. Now that we have covered the core mechanics of TensorFlow, the next chapter will introduce the concept behind convolutional neural network (CNN) architectures for deep learning. CNNs are powerful models and have shown great performance in the field of computer vision.

[ 515 ]

Classifying Images with Deep Convolutional Neural Networks In the previous chapter, we looked in depth at different aspects of the TensorFlow API, you became familiar with tensors and decorating functions, and you learned how to work with TensorFlow Estimators. In this chapter, you will now learn about convolutional neural networks (CNNs) for image classification. We will start by discussing the basic building blocks of CNNs, using a bottom-up approach. Then, we will take a deeper dive into the CNN architecture and explore how to implement CNNs in TensorFlow. In this chapter, we will cover the following topics: • Convolution operations in one and two dimensions • The building blocks of CNN architectures • Implementing deep CNNs in TensorFlow • Data augmentation techniques for improving the generalization performance • Implementing a face image-based CNN classifier for predicting the gender of a person

[ 517 ]

Classifying Images with Deep Convolutional Neural Networks

The building blocks of CNNs

CNNs are a family of models that were originally inspired by how the visual cortex of the human brain works when recognizing objects. The development of CNNs goes back to the 1990s, when Yann LeCun and his colleagues proposed a novel NN architecture for classifying handwritten digits from images (Handwritten Digit Recognition with a Back-Propagation Network, Y. LeCun, and others, 1989, published at the Neural Information Processing Systems (NeurIPS) conference). The human visual cortex The original discovery of how the visual cortex of our brain functions was made by David H. Hubel and Torsten Wiesel in 1959, when they inserted a microelectrode into the primary visual cortex of an anesthetized cat. Then, they observed that brain neurons respond differently after projecting different patterns of light in front of the cat. This eventually led to the discovery of the different layers of the visual cortex. While the primary layer mainly detects edges and straight lines, higher-order layers focus more on extracting complex shapes and patterns.

Due to the outstanding performance of CNNs for image classification tasks, this particular type of feedforward NN gained a lot of attention and led to tremendous improvements in machine learning for computer vision. Several years later, in 2019, Yann LeCun received the Turing award (the most prestigious award in computer science) for his contributions to the field of artificial intelligence (AI), along with two other researchers, Yoshua Bengio and Geoffrey Hinton, whose names you have already encountered in previous chapters. In the following sections, we will discuss the broader concept of CNNs and why convolutional architectures are often described as "feature extraction layers." Then, we will delve into the theoretical definition of the type of convolution operation that is commonly used in CNNs and walk through examples for computing convolutions in one and two dimensions.

Understanding CNNs and feature hierarchies

Successfully extracting salient (relevant) features is key to the performance of any machine learning algorithm and traditional machine learning models rely on input features that may come from a domain expert or are based on computational feature extraction techniques.

[ 518 ]

Chapter 15

Certain types of NNs, such as CNNs, are able to automatically learn the features from raw data that are most useful for a particular task. For this reason, it's common to consider CNN layers as feature extractors: the early layers (those right after the input layer) extract low-level features from raw data, and the later layers (often fully connected layers like in a multilayer perceptron (MLP)) use these features to predict a continuous target value or class label. Certain types of multilayer NNs, and in particular, deep convolutional NNs, construct a so-called feature hierarchy by combining the low-level features in a layer-wise fashion to form high-level features. For example, if we're dealing with images, then low-level features, such as edges and blobs, are extracted from the earlier layers, which are combined together to form high-level features. These highlevel features can form more complex shapes, such as the general contours of objects like buildings, cats, or dogs. As you can see in the following image, a CNN computes feature maps from an input image, where each element comes from a local patch of pixels in the input image:

(Photo by Alexander Dummer on Unsplash)

This local patch of pixels is referred to as the local receptive field. CNNs will usually perform very well on image-related tasks, and that's largely due to two important ideas: • Sparse connectivity: A single element in the feature map is connected to only a small patch of pixels. (This is very different from connecting to the whole input image as in the case of perceptrons. You may find it useful to look back and compare how we implemented a fully connected network that connected to the whole image in Chapter 12, Implementing a Multilayer Artificial Neural Network from Scratch.) • Parameter-sharing: The same weights are used for different patches of the input image. [ 519 ]

Classifying Images with Deep Convolutional Neural Networks

As a direct consequence of these two ideas, replacing a conventional, fully connected MLP with a convolution layer substantially decreases the number of weights (parameters) in the network and we will see an improvement in the ability to capture salient features. In the context of image data, it makes sense to assume that nearby pixels are typically more relevant to each other than pixels that are far away from each other. Typically, CNNs are composed of several convolutional and subsampling layers that are followed by one or more fully connected layers at the end. The fully connected layers are essentially an MLP, where every input unit, i, is connected to every output unit, j, with weight 𝑤𝑤𝑖𝑖𝑖𝑖 (which we covered in more detail in Chapter 12, Implementing a Multilayer Artificial Neural Network from Scratch). Please note that subsampling layers, commonly known as pooling layers, do not have any learnable parameters; for instance, there are no weights or bias units in pooling layers. However, both the convolutional and fully connected layers have weights and biases that are optimized during training.

In the following sections, we will study convolutional and pooling layers in more detail and see how they work. To understand how convolution operations work, let's start with a convolution in one dimension, which is sometimes used for working with certain types of sequence data, such as text. After discussing one-dimensional convolutions, we will work through the typical two-dimensional ones that are commonly applied to two-dimensional images.

Performing discrete convolutions

A discrete convolution (or simply convolution) is a fundamental operation in a CNN. Therefore, it's important to understand how this operation works. In this section, we will cover the mathematical definition and discuss some of the naive algorithms to compute convolutions of one-dimensional tensors (vectors) and two-dimensional tensors (matrices). Please note that the formulas and descriptions in this section are solely for understanding how convolution operations in CNNs work. Indeed, much more efficient implementations of convolutional operations already exist in packages such as TensorFlow, as you will see later in this chapter. Mathematical notation In this chapter, we will use subscripts to denote the size of a multidimensional array (tensor); for example, 𝐴𝐴𝑛𝑛1 ×𝑛𝑛2 is a twodimensional array of size 𝑛𝑛1 × 𝑛𝑛2. We use brackets, [ ], to denote the indexing of a multidimensional array. [ 520 ]

Chapter 15

For example, A[i, j] refers to the element at index i, j of matrix A. Furthermore, note that we use a special symbol, ∗ , to denote the convolution operation between two vectors or matrices, which is not to be confused with the multiplication operator, *, in Python.

Discrete convolutions in one dimension

Let's start with some basic definitions and notations that we are going to use. A discrete convolution for two vectors, x and w, is denoted by 𝒚𝒚 = 𝒙𝒙 ∗ 𝒘𝒘, in which vector x is our input (sometimes called signal) and w is called the filter or kernel. A discrete convolution is mathematically defined as follows: +∞

𝒚𝒚 = 𝒙𝒙 ∗ 𝒘𝒘 → 𝑦𝑦[𝑖𝑖] = ∑ 𝑥𝑥[𝑖𝑖 − 𝑘𝑘 ] 𝑤𝑤[𝑘𝑘] 𝑘𝑘=−∞

As mentioned earlier, the brackets, [ ], are used to denote the indexing for vector elements. The index, i, runs through each element of the output vector, y. There are two odd things in the preceding formula that we need to clarify: −∞ to +∞ indices and negative indexing for x.

The fact that the sum runs through indices from −∞ to +∞ seems odd, mainly because in machine learning applications, we always deal with finite feature vectors. For example, if x has 10 features with indices 0, 1, 2,…, 8, 9, then indices −∞ ∶ −1 and 10 ∶ +∞ are out of bounds for x. Therefore, to correctly compute the summation shown in the preceding formula, it is assumed that x and w are filled with zeros. This will result in an output vector, y, that also has infinite size, with lots of zeros as well. Since this is not useful in practical situations, x is padded only with a finite number of zeros. This process is called zero-padding or simply padding. Here, the number of zeros padded on each side is denoted by p. An example padding of a one-dimensional vector, x, is shown in the following figure:

[ 521 ]

Classifying Images with Deep Convolutional Neural Networks

Let's assume that the original input, x, and filter, w, have n and m elements, respectively, where 𝑚𝑚 ≤ 𝑛𝑛 . Therefore, the padded vector, 𝒙𝒙𝑝𝑝, has size n + 2p. The practical formula for computing a discrete convolution will change to the following: 𝑘𝑘=𝑚𝑚−1

𝒚𝒚 = 𝒙𝒙 ∗ 𝒘𝒘 → 𝑦𝑦[𝑖𝑖] = ∑ 𝒙𝒙𝑝𝑝 [𝑖𝑖 + 𝑚𝑚 − 𝑘𝑘]𝑤𝑤[𝑘𝑘] 𝑘𝑘=0

Now that we have solved the infinite index issue, the second issue is indexing x with i + m – k. The important point to notice here is that x and w are indexed in different directions in this summation. Computing the sum with one index going in the reverse direction is equivalent to computing the sum with both indices in the forward direction after flipping one of those vectors, x or w, after they are padded. Then, we can simply compute their dot product. Let's assume we flip (rotate) the filter, w, to get the rotated filter, 𝒘𝒘𝑟𝑟 . Then, the dot product, 𝒙𝒙[𝑖𝑖: 𝑖𝑖 + 𝑚𝑚] . 𝒘𝒘𝑟𝑟 , is computed to get one element, y[i], where x[i: i + m] is a patch of x with size m. This operation is repeated like in a sliding window approach to get all the output elements. The following figure provides an example with x = [3 2 1 7 1 2 5 4] and 13 1 𝒘𝒘 = [ 1 ] so that the first three output elements are computed: 24 4

[ 522 ]

Chapter 15

You can see in the preceding example that the padding size is zero (p = 0). Notice that the rotated filter, 𝒘𝒘𝑟𝑟 , is shifted by two cells each time we shift. This shift is another hyperparameter of a convolution, the stride, s. In this example, the stride is two, s = 2. Note that the stride has to be a positive number smaller than the size of the input vector. We will talk more about padding and strides in the next section. Cross-correlation Cross-correlation (or simply correlation) between an input vector and a filter is denoted by 𝒚𝒚 = 𝒙𝒙 ⋆ 𝒘𝒘 and is very much like a sibling of a convolution, with a small difference: in cross-correlation, the multiplication is performed in the same direction. Therefore, it is not a requirement to rotate the filter matrix, w, in each dimension. Mathematically, cross-correlation is defined as follows: +∞

𝒚𝒚 = 𝒙𝒙 ⋆ 𝒘𝒘 → 𝑦𝑦[𝑖𝑖] = ∑ 𝑥𝑥[𝑖𝑖 + 𝑘𝑘]𝑤𝑤[𝑘𝑘] 𝑘𝑘=−∞

The same rules for padding and stride may be applied to crosscorrelation as well. Note that most deep learning frameworks (including TensorFlow) implement cross-correlation but refer to it as convolution, which is a common convention in the deep learning field.

Padding inputs to control the size of the output feature maps

So far, we've only used zero-padding in convolutions to compute finite-sized output vectors. Technically, padding can be applied with any 𝑝𝑝 ≥ 0. Depending on the choice of p, boundary cells may be treated differently than the cells located in the middle of x. Now, consider an example where n = 5 and m = 3. Then, with p=0, x[0] is only used in computing one output element (for instance, y[0]), while x[1] is used in the computation of two output elements (for instance, y[0] and y[1]). So, you can see that this different treatment of elements of x can artificially put more emphasis on the middle element, x[2], since it has appeared in most computations. We can avoid this issue if we choose p = 2, in which case, each element of x will be involved in computing three elements of y. Furthermore, the size of the output, y, also depends on the choice of the padding strategy we use.

[ 523 ]

Classifying Images with Deep Convolutional Neural Networks

There are three modes of padding that are commonly used in practice: full, same, and valid: • In full mode, the padding parameter, p, is set to p = m – 1. Full padding increases the dimensions of the output; thus, it is rarely used in CNN architectures. • Same padding is usually used to ensure that the output vector has the same size as the input vector, x. In this case, the padding parameter, p, is computed according to the filter size, along with the requirement that the input size and output size are the same. • Finally, computing a convolution in the valid mode refers to the case where p = 0 (no padding). The following figure illustrates the three different padding modes for a simple 5 × 5 pixel input with a kernel size of 3 × 3 and a stride of 1:

The most commonly used padding mode in CNNs is same padding. One of its advantages over the other padding modes is that same padding preserves the size of the vector—or the height and width of the input images when we are working on image-related tasks in computer vision—which makes designing a network architecture more convenient. One big disadvantage of valid padding versus full and same padding, for example, is that the volume of the tensors will decrease substantially in NNs with many layers, which can be detrimental to the network performance. In practice, it is recommended that you preserve the spatial size using same padding for the convolutional layers and decrease the spatial size via pooling layers instead. As for full padding, its size results in an output larger than the input size. Full padding is usually used in signal processing applications where it is important to minimize boundary effects. However, in the deep learning context, boundary effects are usually not an issue, so we rarely see full padding being used in practice.

[ 524 ]

Chapter 15

Determining the size of the convolution output

The output size of a convolution is determined by the total number of times that we shift the filter, w, along the input vector. Let's assume that the input vector is of size n and the filter is of size m. Then, the size of the output resulting from 𝒚𝒚 = 𝒙𝒙 ∗ 𝒘𝒘 , with padding, p, and stride, s, would be determined as follows:

𝑛𝑛 + 2𝑝𝑝 − 𝑚𝑚 𝑜𝑜 = ⌊ ⌋+1 𝑠𝑠

Here, ⌊∙⌋ denotes the floor operation.

The floor operation The floor operation returns the largest integer that is equal to or smaller than the input, for example:

Consider the following two cases:

𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓(1.77) = ⌊1.77⌋ = 1

• Compute the output size for an input vector of size 10 with a convolution kernel of size 5, padding 2, and stride 1:

𝑛𝑛 = 10, 𝑚𝑚 = 5,

𝑝𝑝 = 2,

10 + 2 × 2 − 5 𝑠𝑠 = 1 → 𝑜𝑜 = ⌊ ⌋ + 1 = 10 1

(Note that in this case, the output size turns out to be the same as the input; therefore, we can conclude this to be the same-padding mode.) • How does the output size change for the same input vector when we have a kernel of size 3 and stride 2?

𝑛𝑛 = 10, 𝑚𝑚 = 3,

𝑝𝑝 = 2,

10 + 2 × 2 − 3 𝑠𝑠 = 2 → 𝑜𝑜 = ⌊ ⌋+1 = 6 2

If you are interested in learning more about the size of the convolution output, we recommend the manuscript A guide to convolution arithmetic for deep learning, by Vincent Dumoulin and Francesco Visin, which is freely available at https://arxiv.org/abs/1603.07285. Finally, in order to learn how to compute convolutions in one dimension, a naive implementation is shown in the following code block, and the results are compared with the numpy.convolve function. The code is as follows: >>> import numpy as np [ 525 ]

Classifying Images with Deep Convolutional Neural Networks >>> def conv1d(x, w, p=0, s=1): ... w_rot = np.array(w[::-1]) ... x_padded = np.array(x) ... if p > 0: ... zero_pad = np.zeros(shape=p) ... x_padded = np.concatenate([zero_pad, ... x_padded, ... zero_pad]) ... res = [] ... for i in range(0, int(len(x)/s),s): ... res.append(np.sum(x_padded[i:i+w_rot.shape[0]] * ... w_rot)) ... return np.array(res) >>> ## Testing: >>> x = [1, 3, 2, 4, 5, 6, 1, 3] >>> w = [1, 0, 3, 1, 2] >>> print('Conv1d Implementation:', ... conv1d(x, w, p=2, s=1)) Conv1d Implementation: [ 5. 14. 16. 26. 24. 34. 19. 22.] >>> print('NumPy Results:', ... np.convolve(x, w, mode='same')) NumPy Results: [ 5 14 16 26 24 34 19 22]

So far, we have mostly focused on convolutions for vectors (1D convolutions). We started with the 1D case to make the concepts easier to understand. In the next section, we will cover 2D convolutions in more detail, which are the building blocks of CNNs for image-related tasks.

Performing a discrete convolution in 2D

The concepts you learned in the previous sections are easily extendible to 2D. When we deal with 2D inputs, such as a matrix, 𝑿𝑿𝑛𝑛1×𝑛𝑛2, and the filter matrix, 𝑾𝑾𝑚𝑚1×𝑚𝑚2, where 𝑚𝑚1 ≤ 𝑛𝑛1 and 𝑚𝑚2 ≤ 𝑛𝑛2, then the matrix 𝒀𝒀 = 𝑿𝑿 ∗ 𝑾𝑾 is the result of a 2D convolution between X and W. This is defined mathematically as follows: +∞

𝒀𝒀 = 𝑿𝑿 ∗ 𝑾𝑾 → 𝑌𝑌[𝑖𝑖, 𝑗𝑗] = ∑

+∞

∑ 𝑋𝑋[𝑖𝑖 − 𝑘𝑘1 , 𝑗𝑗 − 𝑘𝑘2 ] 𝑊𝑊[𝑘𝑘1, 𝑘𝑘2 ]

𝑘𝑘1 =−∞ 𝑘𝑘2 =−∞

[ 526 ]

Chapter 15

Notice that if you omit one of the dimensions, the remaining formula is exactly the same as the one we used previously to compute the convolution in 1D. In fact, all the previously mentioned techniques, such as zero-padding, rotating the filter matrix, and the use of strides, are also applicable to 2D convolutions, provided that they are extended to both dimensions independently. The following figure demonstrates 2D convolution of an input matrix of size 8 × 8, using a kernel of size 3 × 3. The input matrix is padded with zeros with p = 1. As a result, the output of the 2D convolution will have a size of 8 × 8:

The following example illustrates the computation of a 2D convolution between an input matrix, 𝑿𝑿3×3, and a kernel matrix, 𝑾𝑾3×3 , using padding p = (1, 1) and stride s = (2, 2). According to the specified padding, one layer of zeros is added on each padded side of the input matrix, which results in the padded matrix 𝑿𝑿5×5 , as follows:

[ 527 ]

Classifying Images with Deep Convolutional Neural Networks

With the preceding filter, the rotated filter will be:

0.5 1 0.5 𝑾𝑾𝑟𝑟 = [0.1 0.4 0.3] 0.4 0.7 0.5

Note that this rotation is not the same as the transpose matrix. To get the rotated filter in NumPy, we can write W_rot=W[::-1,::-1]. Next, we can shift the rotated filter matrix along the padded input matrix, 𝑿𝑿padded , like a sliding window and compute the sum of the element-wise product, which is denoted by the ⊙ operator in the following figure:

The result will be the 2 × 2 matrix, Y.

Let's also implement the 2D convolution according to the naive algorithm described. The scipy.signal package provides a way to compute 2D convolution via the scipy.signal.convolve2d function: >>> import numpy as np >>> import scipy.signal >>> def conv2d(X, W, p=(0, 0), s=(1, 1)): ... W_rot = np.array(W)[::-1,::-1] ... X_orig = np.array(X) ... n1 = X_orig.shape[0] + 2*p[0] ... n2 = X_orig.shape[1] + 2*p[1] ... X_padded = np.zeros(shape=(n1, n2)) ... X_padded[p[0]:p[0]+X_orig.shape[0], ... p[1]:p[1]+X_orig.shape[1]] = X_orig ... [ 528 ]

Chapter 15 ... ... ... ... ... ... ... ... ... ... >>> X = >>> W =

res = [] for i in range(0, int((X_padded.shape[0] - \ W_rot.shape[0])/s[0])+1, s[0]): res.append([]) for j in range(0, int((X_padded.shape[1] - \ W_rot.shape[1])/s[1])+1, s[1]): X_sub = X_padded[i:i+W_rot.shape[0], j:j+W_rot.shape[1]] res[-1].append(np.sum(X_sub * W_rot)) return(np.array(res)) [[1, 3, 2, 4], [5, 6, 1, 3], [1, 2, 0, 2], [3, 4, 3, 2]] [[1, 0, 3], [1, 2, 1], [0, 1, 1]]

>>> print('Conv2d Implementation:\n', ... conv2d(X, W, p=(1, 1), s=(1, 1))) Conv2d Implementation: [[ 11. 25. 32. 13.] [ 19. 25. 24. 13.] [ 13. 28. 25. 17.] [ 11. 17. 14. 9.]]

>>> print('SciPy Results:\n', ... scipy.signal.convolve2d(X, W, mode='same')) SciPy Results: [[11 25 32 13] [19 25 24 13] [13 28 25 17] [11 17 14 9]]

Efficient algorithms for computing convolution We provided a naive implementation to compute a 2D convolution for the purpose of understanding the concepts. However, this implementation is very inefficient in terms of memory requirements and computational complexity. Therefore, it should not be used in real-world NN applications. One aspect is that the filter matrix is actually not rotated in most tools like TensorFlow. Moreover, in recent years, much more efficient algorithms have been developed that use the Fourier transform to compute convolutions. It is also important to note that in the context of NNs, the size of a convolution kernel is usually much smaller than the size of the input image.

[ 529 ]

Classifying Images with Deep Convolutional Neural Networks

For example, modern CNNs usually use kernel sizes such as 1 × 1, 3 × 3, or 5 × 5 , for which efficient algorithms have been designed that can carry out the convolutional operations much more efficiently, such as Winograd's minimal filtering algorithm. These algorithms are beyond the scope of this book, but if you are interested in learning more, you can read the manuscript Fast Algorithms for Convolutional Neural Networks, by Andrew Lavin and Scott Gray, 2015, which is freely available at https://arxiv.org/abs/1509.09308.

In the next section, we will discuss subsampling or pooling, which is another important operation often used in CNNs.

Subsampling layers

Subsampling is typically applied in two forms of pooling operations in CNNs: max-pooling and mean-pooling (also known as average-pooling). The pooling layer is usually denoted by 𝑃𝑃𝑛𝑛1 ×𝑛𝑛2. Here, the subscript determines the size of the neighborhood (the number of adjacent pixels in each dimension) where the max or mean operation is performed. We refer to such a neighborhood as the pooling size. The operation is described in the following figure. Here, max-pooling takes the maximum value from a neighborhood of pixels, and mean-pooling computes their average:

The advantage of pooling is twofold: • Pooling (max-pooling) introduces a local invariance. This means that small changes in a local neighborhood do not change the result of max-pooling. [ 530 ]

Chapter 15

Therefore, it helps with generating features that are more robust to noise in the input data. Refer to the following example, which shows that the maxpooling of two different input matrices, 𝑿𝑿1 and 𝑿𝑿2, results in the same output:

10 70 255 𝑿𝑿1 = 0 70 [ 35

100 95 80 𝑿𝑿2 = 255 30 [ 70

255 125 255 105 0 150 255 10 200 15 100 25 100 255 40 30 30 30

100 100 10 150 150 100

170 100 0 25 70 25 10 10 0 150 20 10 100 95 0 20 0 60 ] 50 125 10 20 100 200

100 125 125 120 70 70

50 170 150 125 70 95 ]}

max pooling 𝑃𝑃2×2

→

255 125 170 [255 150 150] 70 200 95

• Pooling decreases the size of features, which results in higher computational efficiency. Furthermore, reducing the number of features may reduce the degree of overfitting as well. Overlapping versus non-overlapping pooling Traditionally, pooling is assumed to be nonoverlapping. Pooling is typically performed on nonoverlapping neighborhoods, which can be done by setting the stride parameter equal to the pooling size. For example, a non-overlapping pooling layer, 𝑃𝑃𝑛𝑛1 ×𝑛𝑛2, requires a stride parameter 𝑠𝑠 = (𝑛𝑛1 , 𝑛𝑛2 ). On the other hand, overlapping pooling occurs if the stride is smaller than the pooling size. An example where overlapping pooling is used in a convolutional network is described in ImageNet Classification with Deep Convolutional Neural Networks, by A. Krizhevsky, I. Sutskever, and G. Hinton, 2012, which is freely available as a manuscript at https://papers.nips.cc/paper/4824-imagenetclassification-with-deep-convolutionalneural-networks.

While pooling is still an essential part of many CNN architectures, several CNN architectures have also been developed without using pooling layers. Instead of using pooling layers to reduce the feature size, researchers use convolutional layers with a stride of 2. [ 531 ]

Classifying Images with Deep Convolutional Neural Networks

In a sense, you can think of a convolutional layer with stride 2 as a pooling layer with learnable weights. If you are interested in an empirical comparison of different CNN architectures developed with and without pooling layers, we recommend reading the research article Striving for Simplicity: The All Convolutional Net, by Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. This article is freely available at https://arxiv.org/abs/1412.6806.

Putting everything together – implementing a CNN

So far, you have learned about the basic building blocks of CNNs. The concepts illustrated in this chapter are not really more difficult than traditional multilayer NNs. We can say that the most important operation in a traditional NN is matrix multiplication. For instance, we use matrix multiplications to compute the preactivations (or net inputs), as in z = Wx + b. Here, x is a column vector ( ℝ𝑛𝑛×1 matrix) representing pixels, and W is the weight matrix connecting the pixel inputs to each hidden unit. In a CNN, this operation is replaced by a convolution operation, as in 𝒁𝒁 = 𝑾𝑾 ∗ 𝑿𝑿 + 𝒃𝒃, where X is a matrix representing the pixels in a ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡 × 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ arrangement. In both cases, the pre-activations are passed to an activation function to obtain the activation of a hidden unit, 𝑨𝑨 = 𝜙𝜙(𝒁𝒁), where 𝜙𝜙 is the activation function. Furthermore, you will recall that subsampling is another building block of a CNN, which may appear in the form of pooling, as was described in the previous section.

Working with multiple input or color channels

An input to a convolutional layer may contain one or more 2D arrays or matrices with dimensions 𝑁𝑁1 × 𝑁𝑁2 (for example, the image height and width in pixels). These 𝑁𝑁1 × 𝑁𝑁2 matrices are called channels. Conventional implementations of convolutional layers expect a rank-3 tensor representation as an input, for example a three-dimensional array, 𝑿𝑿𝑁𝑁1×𝑁𝑁2×𝐶𝐶𝑖𝑖𝑖𝑖 , where 𝐶𝐶𝑖𝑖𝑖𝑖 is the number of input channels. For example, let's consider images as input to the first layer of a CNN. If the image is colored and uses the RGB color mode, then 𝐶𝐶𝑖𝑖𝑖𝑖 = 3 (for the red, green, and blue color channels in RGB). However, if the image is in grayscale, then we have 𝐶𝐶𝑖𝑖𝑖𝑖 = 1, because there is only one channel with the grayscale pixel intensity values.

[ 532 ]

Chapter 15

Reading an image file When we work with images, we can read images into NumPy arrays using the uint8 (unsigned 8-bit integer) data type to reduce memory usage compared to 16-bit, 32-bit, or 64-bit integer types, for example. Unsigned 8-bit integers take values in the range [0, 255], which are sufficient to store the pixel information in RGB images, which also take values in the same range. In Chapter 13, Parallelizing Neural Network Training with TensorFlow, you saw that TensorFlow provides a module for loading/storing and manipulating images via tf.io and tf.image submodules. Let's recap how to read an image (this example RGB image is located in the code bundle folder that is provided with this chapter at https://github.com/rasbt/pythonmachine-learning-book-3rd-edition/tree/master/code/ch15): >>> import tensorflow as tf >>> img_raw = tf.io.read_file('example-image.png') >>> img = tf.image.decode_image(img_raw) >>> print('Image shape:', img.shape) Image shape: (252, 221, 3)

When you build models and data loaders in TensorFlow, it is recommended that you use tf.image as well to read in the input images. Now, let's also look at an example of how we can read in an image into our Python session using the imageio package. We can install imageio either via conda or pip from the command-line terminal: > conda install imageio

or > pip install imageio

Once imageio is installed, we can use the imread function to read in the same image we used previously by using the imageio package: >>> import imageio >>> img = imageio.imread('example-image.png') >>> print('Image shape:', img.shape) Image shape: (252, 221, 3) >>> print('Number of channels:', img.shape[2]) Number of channels: 3 >>> print('Image data type:', img.dtype) Image data type: uint8 >>> print(img[100:102, 100:102, :]) [[[179 134 110] [182 136 112]] [[180 135 11] [182 137 113]]] [ 533 ]

Classifying Images with Deep Convolutional Neural Networks

Now that you are familiar with the structure of input data, the next question is, how can we incorporate multiple input channels in the convolution operation that we discussed in the previous sections? The answer is very simple: we perform the convolution operation for each channel separately and then add the results together using the matrix summation. The convolution associated with each channel (c) has its own kernel matrix as W[:, :, c]. The total pre-activation result is computed in the following formula:

Given an example 𝑿𝑿𝑛𝑛1×𝑛𝑛2 ×𝐶𝐶𝑖𝑖𝑖𝑖 , a kernel matrix 𝑾𝑾𝑚𝑚1 ×𝑚𝑚2×𝐶𝐶𝑖𝑖𝑖𝑖 , and bias value 𝑏𝑏

⇒

𝒁𝒁

𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶

𝐶𝐶𝑖𝑖𝑖𝑖

= ∑ 𝑾𝑾[: , : , 𝑐𝑐] ∗ 𝑿𝑿[: , : , 𝑐𝑐] 𝑐𝑐=1

Pre-activation: { Feature map:

𝒁𝒁 = 𝒁𝒁𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 + 𝑏𝑏𝐶𝐶 𝑨𝑨 = 𝜙𝜙(𝒁𝒁)

The final result, A, is a feature map. Usually, a convolutional layer of a CNN has more than one feature map. If we use multiple feature maps, the kernel tensor becomes four-dimensional: 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ × ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡 × 𝐶𝐶𝑖𝑖𝑖𝑖 × 𝐶𝐶𝑜𝑜𝑜𝑜𝑜𝑜 . Here, 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ × ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡 is the kernel size, 𝐶𝐶𝑖𝑖𝑖𝑖 is the number of input channels, and 𝐶𝐶𝑜𝑜𝑜𝑜𝑜𝑜 is the number of output feature maps. So, now let's include the number of output feature maps in the preceding formula and update it, as follows:

Given an example 𝑿𝑿𝑛𝑛1×𝑛𝑛2 ×𝐶𝐶𝑖𝑖𝑖𝑖 , a kernel matrix 𝑾𝑾𝑚𝑚1×𝑚𝑚2×𝐶𝐶𝑖𝑖𝑖𝑖 ×𝐶𝐶𝑜𝑜𝑜𝑜𝑜𝑜 , ⇒ and bias vector 𝒃𝒃𝐶𝐶𝑜𝑜𝑜𝑜𝑜𝑜

𝐶𝐶𝑖𝑖𝑖𝑖

𝒁𝒁𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 [: , : , 𝑘𝑘] = ∑ 𝑾𝑾[: , : , 𝑐𝑐, 𝑘𝑘] ∗ 𝑿𝑿[: , : , 𝑐𝑐] 𝑐𝑐=1

𝒁𝒁[: , : , 𝑘𝑘] = 𝒁𝒁𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 [: , : , 𝑘𝑘] + 𝑏𝑏[𝑘𝑘] { 𝑨𝑨[: , : , 𝑘𝑘] = 𝜙𝜙(𝒁𝒁[: , : , 𝑘𝑘])

To conclude our discussion of computing convolutions in the context of NNs, let's look at the example in the following figure, which shows a convolutional layer, followed by a pooling layer. In this example, there are three input channels. The kernel tensor is four-dimensional. Each kernel matrix is denoted as 𝑚𝑚1 × 𝑚𝑚2, and there are three of them, one for each input channel. Furthermore, there are five such kernels, accounting for five output feature maps. Finally, there is a pooling layer for subsampling the feature maps:

[ 534 ]

Chapter 15

How many trainable parameters exist in the preceding example? To illustrate the advantages of convolution, parametersharing, and sparse connectivity, let's work through an example. The convolutional layer in the network shown in the preceding figure is a four-dimensional tensor. So, there are 𝑚𝑚1 × 𝑚𝑚2 × 3 × 5 parameters associated with the kernel. Furthermore, there is a bias vector for each output feature map of the convolutional layer. Thus, the size of the bias vector is 5. Pooling layers do not have any (trainable) parameters; therefore, we can write the following:

𝑚𝑚1 × 𝑚𝑚2 × 3 × 5 + 5

If the input tensor is of size 𝑛𝑛1 × 𝑛𝑛2 × 3 , assuming that the convolution is performed with the same-padding mode, then the output feature maps would be of size 𝑛𝑛1 × 𝑛𝑛2 × 5 .

Note that if we use a fully connected layer instead of a convolutional layer, this number will be much larger. In the case of a fully connected layer, the number of parameters for the weight matrix to reach the same number of output units would have been as follows:

(𝑛𝑛1 × 𝑛𝑛2 × 3) × (𝑛𝑛1 × 𝑛𝑛2 × 5) = (𝑛𝑛1 × 𝑛𝑛2 )2 × 3 × 5

In addition, the size of the bias vector is 𝑛𝑛1 × 𝑛𝑛2 × 5 (one bias element for each output unit). Given that 𝑚𝑚1 < 𝑛𝑛1 and 𝑚𝑚2 < 𝑛𝑛2, we can see that the difference in the number of trainable parameters is significant.

[ 535 ]

Classifying Images with Deep Convolutional Neural Networks

Lastly, as was already mentioned, typically, the convolution operations are carried out by treating an input image with multiple color channels as a stack of matrices; that is, we perform the convolution on each matrix separately and then add the results, as was illustrated in the previous figure. However, convolutions can also be extended to 3D volumes if you are working with 3D datasets, for example, as shown in the paper VoxNet: A 3D Convolutional Neural Network for Real-Time Object Recognition (2015), by Daniel Maturana and Sebastian Scherer, which can be accessed at https://www.ri.cmu.edu/pub_files/2015/9/voxnet_maturana_scherer_ iros15.pdf. In the next section, we will talk about how to regularize an NN.

Regularizing an NN with dropout

Choosing the size of a network, whether we are dealing with a traditional (fully connected) NN or a CNN, has always been a challenging problem. For instance, the size of a weight matrix and the number of layers need to be tuned to achieve a reasonably good performance. You will recall from Chapter 14, Going Deeper – The Mechanics of TensorFlow, that a simple network without any hidden layer could only capture a linear decision boundary, which is not sufficient for dealing with an exclusive or (XOR) or similar problem. The capacity of a network refers to the level of complexity of the function that it can learn to approximate. Small networks, or networks with a relatively small number of parameters, have a low capacity and are therefore likely to underfit, resulting in poor performance, since they cannot learn the underlying structure of complex datasets. However, very large networks may result in overfitting, where the network will memorize the training data and do extremely well on the training dataset while achieving a poor performance on the held-out test dataset. When we deal with real-world machine learning problems, we do not know how large the network should be a priori. One way to address this problem is to build a network with a relatively large capacity (in practice, we want to choose a capacity that is slightly larger than necessary) to do well on the training dataset. Then, to prevent overfitting, we can apply one or multiple regularization schemes to achieve a good generalization performance on new data, such as the held-out test dataset.

[ 536 ]

Chapter 15

In Chapter 3, A Tour of Machine Learning Classifiers Using scikit-learn, we covered L1 and L2 regularization. In the section Tackling overfitting via regularization, you saw that both techniques, L1 and L2 regularization, can prevent or reduce the effect of overfitting by adding a penalty to the loss that results in shrinking the weight parameters during training. While both L1 and L2 regularization can be used for NNs as well, with L2 being the more common choice of the two, there are other methods for regularizing NNs, such as dropout, which we discuss in this section. But before we move on to discussing dropout, to use L2 regularization within a convolutional or fully connected (dense) network, you can simply add the L2 penalty to the loss function by setting the kernel_regularizer of a particular layer when using the Keras API, as follows (it will then automatically modify the loss function accordingly): >>> from tensorflow import keras >>> conv_layer = keras.layers.Conv2D( ... filters=16, ... kernel_size=(3,3), ... kernel_regularizer=keras.regularizers.l2(0.001)) >>> fc_layer = keras.layers.Dense( ... units=16, ... kernel_regularizer=keras.regularizers.l2(0.001))

In recent years, dropout has emerged as a popular technique for regularizing (deep) NNs to avoid overfitting, thus improving the generalization performance (Dropout: a simple way to prevent neural networks from overfitting, by N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, Journal of Machine Learning Research 15.1, pages 1929-1958, 2014, http://www.jmlr.org/papers/volume15/srivastava14a/ srivastava14a.pdf). Dropout is usually applied to the hidden units of higher layers and works as follows: during the training phase of an NN, a fraction of the hidden units is randomly dropped at every iteration with probability 𝑝𝑝drop (or keep probability 𝑝𝑝keep = 1 − 𝑝𝑝drop). This dropout probability is determined by the user and the common choice is p = 0.5, as discussed in the previously mentioned article by Nitish Srivastava and others, 2014. When dropping a certain fraction of input neurons, the weights associated with the remaining neurons are rescaled to account for the missing (dropped) neurons. The effect of this random dropout is that the network is forced to learn a redundant representation of the data. Therefore, the network cannot rely on an activation of any set of hidden units, since they may be turned off at any time during training, and is forced to learn more general and robust patterns from the data.

[ 537 ]

Classifying Images with Deep Convolutional Neural Networks

This random dropout can effectively prevent overfitting. The following figure shows an example of applying dropout with probability p = 0.5 during the training phase, whereby half of the neurons will become inactive randomly (dropped units are selected randomly in each forward pass of training). However, during prediction, all neurons will contribute to computing the pre-activations of the next layer.

As shown here, one important point to remember is that units may drop randomly during training only, whereas for the evaluation (inference) phase, all the hidden units must be active (for instance, 𝑝𝑝drop = 0 or 𝑝𝑝keep = 1). To ensure that the overall activations are on the same scale during training and prediction, the activations of the active neurons have to be scaled appropriately (for example, by halving the activation if the dropout probability was set to p = 0.5). However, since it is inconvenient to always scale activations when making predictions, TensorFlow and other tools scale the activations during training (for example, by doubling the activations if the dropout probability was set to p = 0.5). This approach is commonly referred to as inverse dropout. While the relationship is not immediately obvious, dropout can be interpreted as the consensus (averaging) of an ensemble of models. As discussed in Chapter 7, Combining Different Models for Ensemble Learning, in ensemble learning, we train several models independently. During prediction, we then use the consensus of all the trained models. We already know that model ensembles are known to perform better than single models. In deep learning, however, both training several models and collecting and averaging the output of multiple models is computationally expensive. Here, dropout offers a workaround, with an efficient way to train many models at once and compute their average predictions at test or prediction time. As mentioned previously, the relationship between model ensembles and dropout is not immediately obvious. However, consider that in dropout, we have a different model for each mini-batch (due to setting the weights to zero randomly during each forward pass). [ 538 ]

Chapter 15

Then, via iterating over the mini-batches, we essentially sample over 𝑀𝑀 = 2ℎ models, where h is the number of hidden units. The restriction and aspect that distinguishes dropout from regular ensembling, however, is that we share the weights over these "different models", which can be seen as a form of regularization. Then, during "inference" (for instance, predicting the labels in the test dataset), we can average over all these different models that we sampled over during training. This is very expensive, though. Then, averaging the models, that is, computing the geometric mean of the classmembership probability that is returned by a model, i, can be computed as follows: 𝑀𝑀

𝑝𝑝𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸 = [∏ 𝑝𝑝{𝑖𝑖} ]

1 𝑀𝑀

𝑗𝑗=1

Now, the trick behind dropout is that this geometric mean of the model ensembles (here, M models) can be approximated by scaling the predictions of the last (or final) model sampled during training by a factor of 1/(1 – p), which is much cheaper than computing the geometric mean explicitly using the previous equation. (In fact, the approximation is exactly equivalent to the true geometric mean if we consider linear models.)

Loss functions for classification

In Chapter 13, Parallelizing Neural Network Training with TensorFlow, we saw different activation functions, such as ReLU, sigmoid, and tanh. Some of these activation functions, like ReLU, are mainly used in the intermediate (hidden) layers of an NN to add non-linearities to our model. But others, like sigmoid (for binary) and softmax (for multiclass), are added at the last (output) layer, which results in classmembership probabilities as the output of the model. If the sigmoid or softmax activations are not included at the output layer, then the model will compute the logits instead of the class-membership probabilities. Focusing on classification problems here, depending on the type of problem (binary versus multiclass) and the type of output (logits versus probabilities), we should choose the appropriate loss function to train our model. Binary cross-entropy is the loss function for a binary classification (with a single output unit), and categorical cross-entropy is the loss function for multiclass classification. In the Keras API, two options for categorical cross-entropy loss are provided, depending on whether the ground truth labels are in a one-hot encoded format (for example, [0, 0, 1, 0]), or provided as integer labels (for example, y=2), which is also known as "sparse" representation in the context of Keras. [ 539 ]

Classifying Images with Deep Convolutional Neural Networks

The following table describes three loss functions available in Keras for dealing with all three cases: binary classification, multiclass with one-hot encoded ground truth labels, and multiclass with integer (sparse) labels. Each one of these three loss functions also has the option to receive the predictions in the form of logits or classmembership probabilities:

Please note that computing the cross-entropy loss by providing the logits, and not the class-membership probabilities, is usually preferred due to numerical stability reasons. If we provide logits as inputs to the loss function and set from_logits=True, the respective TensorFlow function uses a more efficient implementation to compute the loss and derivative of the loss with respect to the weights. This is possible since certain mathematical terms cancel and thus don't have to be computed explicitly when providing logits as inputs. The following code will show you how to use these three loss functions with two different formats, where either the logits or class-membership probabilities are given as inputs to the loss functions: >>> import tensorflow_datasets as tfds >>> ####### Binary Crossentropy >>> bce_probas = tf.keras.losses.BinaryCrossentropy(from_logits=False) >>> bce_logits = tf.keras.losses.BinaryCrossentropy(from_logits=True) >>> logits = tf.constant([0.8]) >>> probas = tf.keras.activations.sigmoid(logits) >>> tf.print( ... 'BCE (w Probas): {:.4f}'.format( ... bce_probas(y_true=[1], y_pred=probas)), [ 540 ]

Chapter 15 ... '(w Logits): {:.4f}'.format( ... bce_logits(y_true=[1], y_pred=logits))) BCE (w Probas): 0.3711 (w Logits): 0.3711 >>> ####### Categorical Crossentropy >>> cce_probas = tf.keras.losses.CategoricalCrossentropy( ... from_logits=False) >>> cce_logits = tf.keras.losses.CategoricalCrossentropy( ... from_logits=True) >>> logits = tf.constant([[1.5, 0.8, 2.1]]) >>> probas = tf.keras.activations.softmax(logits) >>> tf.print( ... 'CCE (w Probas): {:.4f}'.format( ... cce_probas(y_true=[0, 0, 1], y_pred=probas)), ... '(w Logits): {:.4f}'.format( ... cce_logits(y_true=[0, 0, 1], y_pred=logits))) CCE (w Probas): 0.5996 (w Logits): 0.5996 >>> ####### Sparse Categorical Crossentropy >>> sp_cce_probas = tf.keras.losses.SparseCategoricalCrossentropy( ... from_logits=False) >>> sp_cce_logits = tf.keras.losses.SparseCategoricalCrossentropy( ... from_logits=True) >>> tf.print( ... 'Sparse CCE (w Probas): {:.4f}'.format( ... sp_cce_probas(y_true=[2], y_pred=probas)), ... '(w Logits): {:.4f}'.format( ... sp_cce_logits(y_true=[2], y_pred=logits))) Sparse CCE (w Probas): 0.5996 (w Logits): 0.5996

Note that sometimes, you may come across an implementation where a categorical cross-entropy loss is used for binary classification. Typically, when we have a binary classification task, the model returns a single output value for each example. We interpret this single model output as the probability of the positive class (for example, class 1), P[class = 1]. In a binary classification problem, it is implied that P[class = 0] = 1 – P[class = 1]; hence, we do not need a second output unit in order to obtain the probability of the negative class. However, sometimes practitioners choose to return two outputs for each training example and interpret them as probabilities of each class: P[class = 0] versus P[class = 1]. Then, in such a case, using a softmax function (instead of the logistic sigmoid) to normalize the outputs (so that they sum to 1) is recommended, and categorical cross-entropy is the appropriate loss function.

[ 541 ]

Classifying Images with Deep Convolutional Neural Networks

Implementing a deep CNN using TensorFlow

In Chapter 14, Going Deeper – The Mechanics of TensorFlow, you may recall that we used TensorFlow Estimators for handwritten digit recognition problems, using different API levels of TensorFlow. You may also recall that we achieved about 89 percent accuracy using the DNNClassifier Estimator with two hidden layers. Now, let's implement a CNN and see whether it can achieve a better predictive performance compared to the MLP (DNNClassifier) for classifying handwritten digits. Note that the fully connected layers that we saw in Chapter 14, Going Deeper – The Mechanics of TensorFlow, were able to perform well on this problem. However, in some applications, such as reading bank account numbers from handwritten digits, even tiny mistakes can be very costly. Therefore, it is crucial to reduce this error as much as possible.

The multilayer CNN architecture

The architecture of the network that we are going to implement is shown in the following figure. The inputs are 28 × 28 grayscale images. Considering the number of channels (which is 1 for grayscale images) and a batch of input images, the input tensor's dimensions will be 𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 28 × 28 × 1.

The input data goes through two convolutional layers that have a kernel size of 5 × 5. The first convolution has 32 output feature maps, and the second one has 64 output feature maps. Each convolution layer is followed by a subsampling layer in the form of a max-pooling operation, 𝑃𝑃2×2. Then a fully connected layer passes the output to a second fully connected layer, which acts as the final softmax output layer. The architecture of the network that we are going to implement is shown in the following figure:

[ 542 ]

Chapter 15

The dimensions of the tensors in each layer are as follows: • Input: [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 28 × 28 × 1]

• Conv_1: [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 28 × 28 × 32]

• Pooling_1: [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 14 × 14 × 32] • Conv_2: [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 14 × 14 × 64] • Pooling_2: [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 7 × 7 × 64] • FC_1: [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 1024]

• FC_2 and softmax layer: [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 10]

For the convolutional kernels, we are using strides=1 such that the input dimensions are preserved in the resulting feature maps. For the pooling layers, we are using strides=2 to subsample the image and shrink the size of the output feature maps. We will implement this network using the TensorFlow Keras API.

Loading and preprocessing the data

You will recall that in Chapter 13, Parallelizing Neural Network Training with TensorFlow, you learned two ways of loading available datasets from the tensorflow_datasets module. One approach is based on a three-step process, and a simpler method uses a function called load, which wraps those three steps. Here, we will use the first method. The three steps for loading the MNIST dataset are as follows: >>> >>> >>> >>> >>> >>> >>>

import tensorflow_datasets as tfds ## Loading the data mnist_bldr = tfds.builder('mnist') mnist_bldr.download_and_prepare() datasets = mnist_bldr.as_dataset(shuffle_files=False) mnist_train_orig = datasets['train'] mnist_test_orig = datasets['test']

The MNIST dataset comes with a pre-specified training and test dataset partitioning scheme, but we also want to create a validation split from the train partition. Notice that in the third step, we used an optional argument, shuffle_files=False, in the .as_dataset() method. This prevented initial shuffling, which is necessary for us since we want to split the training dataset into two parts: a smaller training dataset and a validation dataset. (Note: if the initial shuffling was not turned off, it would incur reshuffling of the dataset every time we fetched a mini-batch of data.

[ 543 ]

Classifying Images with Deep Convolutional Neural Networks

An example of this behavior is shown in the online contents of this chapter, where you can see that the number of labels in the validation datasets changes due to reshuffling of the train/validation splits. This can cause false performance estimation of the model, since the train/validation datasets are indeed mixed.) We can split the train/validation datasets as follows: >>> BUFFER_SIZE = 10000 >>> BATCH_SIZE = 64 >>> NUM_EPOCHS = 20 >>> mnist_train = mnist_train_orig.map( ... lambda item: (tf.cast(item['image'], tf.float32)/255.0, ... tf.cast(item['label'], tf.int32))) >>> mnist_test = mnist_test_orig.map( ... lambda item: (tf.cast(item['image'], tf.float32)/255.0, ... tf.cast(item['label'], tf.int32))) >>> tf.random.set_seed(1) >>> mnist_train = mnist_train.shuffle(buffer_size=BUFFER_SIZE, ... reshuffle_each_iteration=False) >>> mnist_valid = mnist_train.take(10000).batch(BATCH_SIZE) >>> mnist_train = mnist_train.skip(10000).batch(BATCH_SIZE)

Now, after preparing the dataset, we are ready to implement the CNN we just described.

Implementing a CNN using the TensorFlow Keras API

For implementing a CNN in TensorFlow, we use the Keras Sequential class to stack different layers, such as convolution, pooling, and dropout, as well as the fully connected (dense) layers. The Keras layers API provides classes for each one: tf.keras.layers.Conv2D for a two-dimensional convolution layer; tf.keras. layers.MaxPool2D and tf.keras.layers.AvgPool2D for subsampling (maxpooling and average-pooling); and tf.keras.layers.Dropout for regularization using dropout. We will go over each of these classes in more detail.

Configuring CNN layers in Keras

Constructing a layer with the Conv2D class requires us to specify the number of output filters (which is equivalent to the number of output feature maps) and kernel sizes. [ 544 ]

Chapter 15

In addition, there are optional parameters that we can use to configure a convolutional layer. The most commonly used ones are the strides (with a default value of 1 in both x, y dimensions) and padding, which could be same or valid. Additional configuration parameters are listed in the official documentation: https://www.tensorflow.org/ versions/r2.0/api_docs/python/tf/keras/layers/Conv2D. It is worth mentioning that usually, when we read an image, the default dimension for the channels is the last dimension of the tensor array. This is called the "NHWC" format, where N stands for the number of images within the batch, H and W stand for height and width, respectively, and C stands for channels. Note that the Conv2D class assumes that inputs are in the NHWC format by default. (Other tools, such as PyTorch, use an NCHW format.) However, if you come across some data whose channels are placed at the first dimension (the first dimension after the batch dimension, or second dimension considering the batch dimension), you would need to swap the axes in your data to move the channels to the last dimension. Or, an alternative way to work with an NCHW-formatted input is to set data_format="channels_first". After the layer is constructed, it can be called by providing a four-dimensional tensor, with the first dimension reserved for a batch of examples; depending on the data_format argument, either the second or the fourth dimension corresponds to the channel; and the other two dimensions are the spatial dimensions. As shown in the architecture of the CNN model that we want to build, each convolution layer is followed by a pooling layer for subsampling (reducing the size of feature maps). The MaxPool2D and AvgPool2D classes construct the max-pooling and average-pooling layers, respectively. The argument pool_size determines the size of the window (or neighborhood) that will be used to compute the max or mean operations. Furthermore, the strides parameter can be used to configure the pooling layer, as we discussed earlier. Finally, the Dropout class will construct the dropout layer for regularization, with the argument rate used to determine the probability of dropping the input units during the training. When calling this layer, its behavior can be controlled via an argument named training, to specify whether this call will be made during training or during the inference.

Constructing a CNN in Keras

Now that you have learned about these classes, we can construct the CNN model that was shown in the previous figure. In the following code, we will use the Sequential class and add the convolution and pooling layers: >>> model = tf.keras.Sequential() >>> model.add(tf.keras.layers.Conv2D( [ 545 ]

Classifying Images with Deep Convolutional Neural Networks ... filters=32, kernel_size=(5, 5), ... strides=(1, 1), padding='same', ... data_format='channels_last', ... name='conv_1', activation='relu')) >>> model.add(tf.keras.layers.MaxPool2D( ... pool_size=(2, 2), name='pool_1')) >>> model.add(tf.keras.layers.Conv2D( ... filters=64, kernel_size=(5, 5), ... strides=(1, 1), padding='same', ... name='conv_2', activation='relu')) >>> model.add(tf.keras.layers.MaxPool2D( ... pool_size=(2, 2), name='pool_2'))

So far, we have added two convolution layers to the model. For each convolutional layer, we used a kernel of size 5 × 5 and 'same' padding. As discussed earlier, using padding='same' preserves the spatial dimensions (vertical and horizontal dimensions) of the feature maps such that the inputs and outputs have the same height and width (and the number of channels may only differ in terms of the number of filters used). The max-pooling layers with pooling size 2 × 2 and strides of 2 will reduce the spatial dimensions by half. (Note that if the strides parameter is not specified in MaxPool2D, by default, it is set equal to the pooling size.) While we can calculate the size of the feature maps at this stage manually, the Keras API provides a convenient method to compute this for us: >>> model.compute_output_shape(input_shape=(16, 28, 28, 1)) TensorShape([16, 7, 7, 64])

By providing the input shape as a tuple specified in this example, the method compute_output_shape calculated the output to have a shape (16, 7, 7, 64), indicating feature maps with 64 channels and a spatial size of 7 × 7. The first dimension corresponds to the batch dimension, for which we used 16 arbitrarily. We could have used None instead, that is, input_shape=(None, 28, 28, 1). The next layer that we want to add is a dense (or fully connected) layer for implementing a classifier on top of our convolutional and pooling layers. The input to this layer must have rank 2, that is, shape [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖_𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢]. Thus, we need to flatten the output of the previous layers to meet this requirement for the dense layer: >>> model.add(tf.keras.layers.Flatten()) >>> model.compute_output_shape(input_shape=(16, 28, 28, 1)) TensorShape([16, 3136])

[ 546 ]

Chapter 15

As the result of compute_output_shape indicates, the input dimensions for the dense layer are correctly set up. Next, we will add two dense layers with a dropout layer in between: >>> model.add(tf.keras.layers.Dense( ... units=1024, name='fc_1', ... activation='relu')) >>> model.add(tf.keras.layers.Dropout( ... rate=0.5)) >>> model.add(tf.keras.layers.Dense( ... units=10, name='fc_2', ... activation='softmax'))

The last fully connected layer, named 'fc_2', has 10 output units for the 10 class labels in the MNIST dataset. Also, we use the softmax activation to obtain the classmembership probabilities of each input example, assuming that the classes are mutually exclusive, so the probabilities for each example sum to 1. (This means that a training example can belong to only one class.) Based on what we discussed in the section Loss functions for classification, which loss should we use here? Remember that for a multiclass classification with integer (sparse) labels (as opposed to one-hot encoded labels), we use SparseCategoricalCrossentropy. The following code will call the build() method for late variable creation and compile the model: >>> tf.random.set_seed(1) >>> model.build(input_shape=(None, 28, 28, 1)) >>> model.compile( ... optimizer=tf.keras.optimizers.Adam(), ... loss=tf.keras.losses.SparseCategoricalCrossentropy(), ... metrics=['accuracy'])

The Adam optimizer Note that in this implementation, we used the tf.keras. optimizers.Adam() class for training the CNN model. The

Adam optimizer is a robust, gradient-based optimization method suited to nonconvex optimization and machine learning problems. Two popular optimization methods inspired Adam: RMSProp and AdaGrad.

The key advantage of Adam is in the choice of update step size derived from the running average of gradient moments. Please feel free to read more about the Adam optimizer in the manuscript, Adam: A Method for Stochastic Optimization, Diederik P. Kingma and Jimmy Lei Ba, 2014. The article is freely available at https://arxiv.org/abs/1412.6980. [ 547 ]

Classifying Images with Deep Convolutional Neural Networks

As you already know, we can train the model by calling the fit() method. Note that using the designated methods for training and evaluation (like evaluate() and predict()) will automatically set the mode for the dropout layer and rescale the hidden units appropriately so that we do not have to worry about that at all. Next, we will train this CNN model and use the validation dataset that we created for monitoring the learning progress: >>> history = model.fit(mnist_train, epochs=NUM_EPOCHS, ... validation_data=mnist_valid, ... shuffle=True) Epoch 1/20 782/782 [==============================] - 35s 45ms/step - loss: 0.1450 - accuracy: 0.8882 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00 Epoch 2/20 782/782 [==============================] - 34s 43ms/step - loss: 0.0472 - accuracy: 0.9833 - val_loss: 0.0507 - val_accuracy: 0.9839 .. Epoch 20/20 782/782 [==============================] - 34s 44ms/step - loss: 0.0047 - accuracy: 0.9985 - val_loss: 0.0488 - val_accuracy: 0.9920

Once the 20 epochs of training are finished, we can visualize the learning curves: >>> import matplotlib.pyplot as plt >>> hist = history.history >>> x_arr = np.arange(len(hist['loss'])) + 1 >>> >>> >>> >>> >>> >>> >>> >>> ... >>> >>>

fig = plt.figure(figsize=(12, 4)) ax = fig.add_subplot(1, 2, 1) ax.plot(x_arr, hist['loss'], '-o', label='Train loss') ax.plot(x_arr, hist['val_loss'], '-->> >>> >>> >>> >>>

import tensorflow as tf import tensorflow_datasets as tfds celeba_bldr = tfds.builder('celeb_a') celeba_bldr.download_and_prepare() celeba = celeba_bldr.as_dataset(shuffle_files=False)

>>> >>> >>> >>> >>> ... ... ... ...

celeba_train = celeba['train'] celeba_valid = celeba['validation'] celeba_test = celeba['test'] def count_items(ds): n = 0 for _ in ds: n += 1 return n

>>> print('Train set: Train set: 162770

{}'.format(count_items(celeba_train)))

>>> print('Validation: {}'.format(count_items(celeba_valid))) Validation: 19867 >>> print('Test set: Test set: 19962

{}'.format(count_items(celeba_test)))

So, instead of using all the available training and validation data, we will take a subset of 16,000 training examples and 1,000 examples for validation, as follows: >>> celeba_train = celeba_train.take(16000) >>> celeba_valid = celeba_valid.take(1000) >>> print('Train set: {}'.format(count_items(celeba_train))) Train set: 16000 >>> print('Validation: {}'.format(count_items(celeba_valid))) Validation: 1000

[ 551 ]

Classifying Images with Deep Convolutional Neural Networks

It is important to note that if the shuffle_files argument in celeba_bldr.as_dataset() was not set to False, we would still see 16,000 examples in the training dataset and 1,000 examples for the validation dataset. However, at each iteration, it would reshuffle the training data and take a new set of 16,000 examples. This would defeat the purpose, as our goal here is to intentionally train our model with a small dataset. Next, we will discuss data augmentation as a technique for boosting the performance of deep NNs.

Image transformation and data augmentation

Data augmentation summarizes a broad set of techniques for dealing with cases where the training data is limited. For instance, certain data augmentation techniques allow us to modify or even artificially synthesize more data and thereby boost the performance of a machine or deep learning model by reducing overfitting. While data augmentation is not only for image data, there are a set of transformations uniquely applicable to image data, such as cropping parts of an image, flipping, changing the contrast, brightness, and saturation. Let's see some of these transformations that are available via the tf.image module. In the following code block, we will first get five examples from the celeba_train dataset and apply five different types of transformation: 1) cropping an image to a bounding box, 2) flipping an image horizontally, 3) adjusting the contrast, 4) adjusting the brightness, and 5) center-cropping an image and resizing the resulting image back to its original size, (218, 178). In the following code, we will visualize the results of these transformations, showing each one in a separate column for comparison: >>> >>> >>> >>> ...

import matplotlib.pyplot as plt # take 5 examples examples = [] for example in celeba_train.take(5): examples.append(example['image'])

>>> fig = plt.figure(figsize=(16, 8.5)) >>> >>> >>> >>> >>> >>>

## Column 1: cropping to a bounding-box ax = fig.add_subplot(2, 5, 1) ax.set_title('Crop to a \nbounding-box', size=15) ax.imshow(examples[0]) ax = fig.add_subplot(2, 5, 6) img_cropped = tf.image.crop_to_bounding_box(

[ 552 ]

Chapter 15 ... examples[0], 50, 20, 128, 128) >>> ax.imshow(img_cropped) >>> >>> >>> >>> >>> >>> >>>

## Column 2: flipping (horizontally) ax = fig.add_subplot(2, 5, 2) ax.set_title('Flip (horizontal)', size=15) ax.imshow(examples[1]) ax = fig.add_subplot(2, 5, 7) img_flipped = tf.image.flip_left_right(examples[1]) ax.imshow(img_flipped)

>>> >>> >>> >>> >>> >>> ... >>>

## Column 3: adjust contrast ax = fig.add_subplot(2, 5, 3) ax.set_title('Adjust constrast', size=15) ax.imshow(examples[2]) ax = fig.add_subplot(2, 5, 8) img_adj_contrast = tf.image.adjust_contrast( examples[2], contrast_factor=2) ax.imshow(img_adj_contrast)

>>> >>> >>> >>> >>> >>> ... >>>

## Column 4: adjust brightness ax = fig.add_subplot(2, 5, 4) ax.set_title('Adjust brightness', size=15) ax.imshow(examples[3]) ax = fig.add_subplot(2, 5, 9) img_adj_brightness = tf.image.adjust_brightness( examples[3], delta=0.3) ax.imshow(img_adj_brightness)

>>> >>> >>> >>> >>> >>> ... >>> ... >>>

## Column 5: cropping from image center ax = fig.add_subplot(2, 5, 5) ax.set_title('Centeral crop\nand resize', size=15) ax.imshow(examples[4]) ax = fig.add_subplot(2, 5, 10) img_center_crop = tf.image.central_crop( examples[4], 0.7) img_resized = tf.image.resize( img_center_crop, size=(218, 178)) ax.imshow(img_resized.numpy().astype('uint8'))

>>> plt.show()

[ 553 ]

Classifying Images with Deep Convolutional Neural Networks

The following figure shows the results:

In the previous figure, the original images are shown in the first row and their transformed version in the second row. Note that for the first transformation (leftmost column), the bounding box is specified by four numbers: the coordinate of the upper-left corner of the bounding box (here x=20, y=50), and the width and height of the box (width=128, height=128). Also note that the origin (the coordinates at the location denoted as (0, 0)) for images loaded by TensorFlow (as well as other packages such as imageio) is the upper-left corner of the image. The transformations in the previous code block are deterministic. However, all such transformations can also be randomized, which is recommended for data augmentation during model training. For example, a random bounding box (where the coordinates of the upper-left corner are selected randomly) can be cropped from an image, an image can be randomly flipped along either the horizontal or vertical axes with a probability of 0.5, or the contrast of an image can be changed randomly, where the contrast_factor is selected at random, but with uniform distribution, from a range of values. In addition, we can create a pipeline of these transformations.

[ 554 ]

Chapter 15

For example, we can first randomly crop an image, then flip it randomly, and finally, resize it to the desired size. The code is as follows (since we have random elements, we set the random seed for reproducibility): >>> tf.random.set_seed(1) >>> fig = plt.figure(figsize=(14, 12)) >>> for i,example in enumerate(celeba_train.take(3)): ... image = example['image'] ... ... ax = fig.add_subplot(3, 4, i*4+1) ... ax.imshow(image) ... if i == 0: ... ax.set_title('Orig', size=15) ... ... ax = fig.add_subplot(3, 4, i*4+2) ... img_crop = tf.image.random_crop(image, size=(178, 178, 3)) ... ax.imshow(img_crop) ... if i == 0: ... ax.set_title('Step 1: Random crop', size=15) ... ... ax = fig.add_subplot(3, 4, i*4+3) ... img_flip = tf.image.random_flip_left_right(img_crop) ... ax.imshow(tf.cast(img_flip, tf.uint8)) ... if i == 0: ... ax.set_title('Step 2: Random flip', size=15) ... ... ax = fig.add_subplot(3, 4, i*4+4) ... img_resize = tf.image.resize(img_flip, size=(128, 128)) ... ax.imshow(tf.cast(img_resize, tf.uint8)) ... if i == 0: ... ax.set_title('Step 3: Resize', size=15) >>> plt.show()

[ 555 ]

Classifying Images with Deep Convolutional Neural Networks

The following figure shows random transformations on three example images:

Note that each time we iterate through these three examples, we get slightly different images due to random transformations. For convenience, we can define a wrapper function to use this pipeline for data augmentation during model training. In the following code, we will define the function preprocess(), which will receive a dictionary containing the keys 'image' and 'attributes'. The function will return a tuple containing the transformed image and the label extracted from the dictionary of attributes.

[ 556 ]

Chapter 15

We will only apply data augmentation to the training examples, however, and not to the validation or test images. The code is as follows: >>> def preprocess(example, size=(64, 64), mode='train'): ... image = example['image'] ... label = example['attributes']['Male'] ... if mode == 'train': ... image_cropped = tf.image.random_crop( ... image, size=(178, 178, 3)) ... image_resized = tf.image.resize( ... image_cropped, size=size) ... image_flip = tf.image.random_flip_left_right( ... image_resized) ... return image_flip/255.0, tf.cast(label, tf.int32) ... else: # use center- instead of ... # random-crops for non-training data ... image_cropped = tf.image.crop_to_bounding_box( ... image, offset_height=20, offset_width=0, ... target_height=178, target_width=178) ... image_resized = tf.image.resize( ... image_cropped, size=size) ... return image_resized/255.0, tf.cast(label, tf.int32)

Now, to see data augmentation in action, let's create a small subset of the training dataset, apply this function to it, and iterate over the dataset five times: >>> tf.random.set_seed(1) >>> ds = celeba_train.shuffle(1000, reshuffle_each_iteration=False) >>> ds = ds.take(2).repeat(5) >>> ds = ds.map(lambda x:preprocess(x, size=(178, 178), mode='train')) >>> fig = plt.figure(figsize=(15, 6)) >>> for j,example in enumerate(ds): ... ax = fig.add_subplot(2, 5, j//2+(j%2)*5+1) ... ax.set_xticks([]) ... ax.set_yticks([]) ... ax.imshow(example[0]) >>> plt.show()

[ 557 ]

Classifying Images with Deep Convolutional Neural Networks

This figure shows the five resulting transformations for data augmentation on two example images:

Next, we will apply this preprocessing function to our training and validation datasets. We will use an image size of (64, 64). Furthermore, we will specify mode='train' when working with training data and use mode='eval' for the validation data so that the random elements of the data augmentation pipeline will be applied only to the training data: >>> import numpy as np >>> >>> >>> >>>

BATCH_SIZE = 32 BUFFER_SIZE = 1000 IMAGE_SIZE = (64, 64) steps_per_epoch = np.ceil(16000/BATCH_SIZE)

>>> ds_train = celeba_train.map( ... lambda x: preprocess(x, size=IMAGE_SIZE, mode='train')) >>> ds_train = ds_train.shuffle(buffer_size=BUFFER_SIZE).repeat() >>> ds_train = ds_train.batch(BATCH_SIZE) >>> ds_valid = celeba_valid.map( ... lambda x: preprocess(x, size=IMAGE_SIZE, mode='eval')) >>> ds_valid = ds_valid.batch(BATCH_SIZE)

Training a CNN gender classifier

By now, building a model with TensorFlow's Keras API and training it should be straightforward. The design of our CNN is as follows: the CNN model receives input images of size 64 × 64 × 3 (the images have three color channels, using 'channels_last'). [ 558 ]

Chapter 15

The input data goes through four convolutional layers to make 32, 64, 128, and 256 feature maps using filters with kernel a size of 3 × 3. The first three convolution layers are followed by max-pooling, 𝑃𝑃2×2. Two dropout layers are also included for regularization: >>> model = tf.keras.Sequential([ ... tf.keras.layers.Conv2D( ... 32, (3, 3), padding='same', activation='relu'), ... tf.keras.layers.MaxPooling2D((2, 2)), ... tf.keras.layers.Dropout(rate=0.5), ... ... tf.keras.layers.Conv2D( ... 64, (3, 3), padding='same', activation='relu'), ... tf.keras.layers.MaxPooling2D((2, 2)), ... tf.keras.layers.Dropout(rate=0.5), ... ... tf.keras.layers.Conv2D( ... 128, (3, 3), padding='same', activation='relu'), ... tf.keras.layers.MaxPooling2D((2, 2)), ... ... tf.keras.layers.Conv2D( ... 256, (3, 3), padding='same', activation='relu') >>> ])

Let's see the shape of the output feature maps after applying these layers: >>> model.compute_output_shape(input_shape=(None, 64, 64, 3)) TensorShape([None, 8, 8, 256])

There are 256 feature maps (or channels) of size 8 × 8. Now, we can add a fully connected layer to get to the output layer with a single unit. If we reshape (flatten) the feature maps, the number of input units to this fully connected layer will be 8 × 8 × 256 = 16,384. Alternatively, let's consider a new layer, called global averagepooling, which computes the average of each feature map separately, thereby reducing the hidden units to 256. We can then add a fully connected layer. Although we have not discussed global average-pooling explicitly, it is conceptually very similar to other pooling layers. Global average-pooling can be viewed, in fact, as a special case of average-pooling when the pooling size is equal to the size of the input feature maps. To understand this, consider the following figure showing an example of input feature maps of shape [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 64 × 64 × 8]. The channels are numbered 𝑘𝑘 = 0, 1, … , 7. The global average-pooling operation calculates the average of each channel so that the output will have shape [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 8]. (Note: GlobalAveragePooling2D in the Keras API will automatically squeeze the output. [ 559 ]

Classifying Images with Deep Convolutional Neural Networks

Without squeezing the output, the shape would be [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 1 × 1 × 8], as the global average-pooling would reduce the spatial dimension of 64 × 64 to a 1 × 1):

Therefore, given that, in our case, the shape of the feature maps prior to this layer is [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 8 × 8 × 256], we expect to get 256 units as output, that is, the shape of the output will be [𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏ℎ𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 × 256]. Let's add this layer and recompute the output shape to verify that this is true: >>> model.add(tf.keras.layers.GlobalAveragePooling2D()) >>> model.compute_output_shape(input_shape=(None, 64, 64, 3)) TensorShape([None, 256])

Finally, we can add a fully connected (dense) layer to get a single output unit. In this case, we can specify the activation function to be 'sigmoid' or just use activation=None, so that the model will output the logits (instead of classmembership probabilities), which is preferred for model training in TensorFlow and Keras due to numerical stability, as discussed earlier: >>> model.add(tf.keras.layers.Dense(1, activation=None)) >>> tf.random.set_seed(1) >>> model.build(input_shape=(None, 64, 64, 3)) >>> model.summary() Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d (Conv2D) multiple 896 _________________________________________________________________ max_pooling2d (MaxPooling2D) multiple 0 _________________________________________________________________ dropout (Dropout) multiple 0 [ 560 ]

Chapter 15 _________________________________________________________________ conv2d_1 (Conv2D) multiple 18496 _________________________________________________________________ max_pooling2d_1 (MaxPooling2 multiple 0 _________________________________________________________________ dropout_1 (Dropout) multiple 0 _________________________________________________________________ conv2d_2 (Conv2D) multiple 73856 _________________________________________________________________ max_pooling2d_2 (MaxPooling2 multiple 0 _________________________________________________________________ conv2d_3 (Conv2D) multiple 295168 _________________________________________________________________ global_average_pooling2d (Gl multiple 0 _________________________________________________________________ dense (Dense) multiple 257 ================================================================= Total params: 388,673 Trainable params: 388,673 Non-trainable params: 0 _________________________________________________________________

The next step is to compile the model and, at this time, we have to decide what loss function to use. We have a binary classification with a single output unit, so that means we should use BinaryCrossentropy. In addition, since our last layer does not apply sigmoid activation (we used activation=None), the outputs of our model are logits, not probabilities. Therefore, we will also specify from_logits=True in BinaryCrossentropy so that the loss function applies the sigmoid function internally, which is, due to the underlying code, more efficient than doing it manually. The code for compiling and training the model is as follows: >>> model.compile(optimizer=tf.keras.optimizers.Adam(), ... loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), ... metrics=['accuracy']) >>> history = model.fit(ds_train, validation_data=ds_valid, ... epochs=20, ... steps_per_epoch=steps_per_epoch)

Let's now visualize the learning curve and compare the training and validation loss and accuracies after each epoch: >>> hist = history.history >>> x_arr = np.arange(len(hist['loss'])) + 1

[ 561 ]

Classifying Images with Deep Convolutional Neural Networks >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> ... >>> >>> >>> >>>

fig = plt.figure(figsize=(12, 4)) ax = fig.add_subplot(1, 2, 1) ax.plot(x_arr, hist['loss'], '-o', label='Train loss') ax.plot(x_arr, hist['val_loss'], '-->> import tensorflow as tf >>> tf.random.set_seed(1) >>> rnn_layer = tf.keras.layers.SimpleRNN( ... units=2, use_bias=True, ... return_sequences=True) >>> rnn_layer.build(input_shape=(None, None, 5)) >>> w_xh, w_oo, b_h = rnn_layer.weights >>> print('W_xh shape:', w_xh.shape) >>> print('W_oo shape:', w_oo.shape) >>> print('b_h shape:', b_h.shape) W_xh shape: (5, 2) W_oo shape: (2, 2) b_h shape: (2,)

The input shape for this layer is (None, None, 5), where the first dimension is the batch dimension (using None for variable batch size), the second dimension corresponds to the sequence (using None for the variable sequence length), and the last dimension corresponds to the features. Notice that we set return_sequences=True, which, for an input sequence of length 3, will result in the output sequence ⟨𝒐𝒐(0) , 𝒐𝒐(1) , 𝒐𝒐(2) ⟩ . Otherwise, it would only return the final output, 𝒐𝒐(2).

Now, we will call the forward pass on the rnn_layer and manually compute the outputs at each time step and compare them: >>> x_seq = tf.convert_to_tensor( ... [[1.0]*5, [2.0]*5, [3.0]*5], ... dtype=tf.float32) >>> ## output of SimepleRNN: >>> output = rnn_layer(tf.reshape(x_seq, shape=(1, 3, 5))) >>> ## manually computing the output: >>> out_man = [] >>> for t in range(len(x_seq)): ... xt = tf.reshape(x_seq[t], (1, 5)) ... print('Time step {} =>'.format(t)) ... print(' Input :', xt.numpy()) ... ... ht = tf.matmul(xt, w_xh) + b_h ... print(' Hidden :', ht.numpy()) ... ... if t>0: [ 579 ]

Modeling Sequential Data Using Recurrent Neural Networks ... prev_o = out_man[t-1] ... else: ... prev_o = tf.zeros(shape=(ht.shape)) ... ot = ht + tf.matmul(prev_o, w_oo) ... ot = tf.math.tanh(ot) ... out_man.append(ot) ... print(' Output (manual) :', ot.numpy()) ... print(' SimpleRNN output:'.format(t), ... output[0][t].numpy()) ... print() Time step 0 => Input : [[1. 1. 1. 1. 1.]] Hidden : [[0.41464037 0.96012145]] Output (manual) : [[0.39240566 0.74433106]] SimpleRNN output: [0.39240566 0.74433106] Time step 1 => Input : Hidden : Output (manual) : SimpleRNN output:

[[2. 2. 2. 2. 2.]] [[0.82928073 1.9202429 ]] [[0.80116504 0.9912947 ]] [0.80116504 0.9912947 ]

Time step 2 => Input : Hidden : Output (manual) : SimpleRNN output:

[[3. 3. 3. 3. 3.]] [[1.243921 2.8803642]] [[0.95468265 0.9993069 ]] [0.95468265 0.9993069 ]

In our manual forward computation, we used the hyperbolic tangent (tanh) activation function, since it is also used in SimpleRNN (the default activation). As you can see from the printed results, the outputs from the manual forward computations exactly match the output of the SimpleRNN layer at each time step. Hopefully, this hands-on task has enlightened you on the mysteries of recurrent networks.

The challenges of learning long-range interactions

BPTT, which was briefly mentioned earlier, introduces some new challenges. 𝜕𝜕𝒉𝒉(𝑡𝑡) , in computing the gradients of a loss Because of the multiplicative factor, 𝜕𝜕𝒉𝒉(𝑘𝑘) function, the so-called vanishing and exploding gradient problems arise. These problems are explained by the examples in the following figure, which shows an RNN with only one hidden unit for simplicity: [ 580 ]

Chapter 16

𝜕𝜕𝒉𝒉(𝑡𝑡) has t – k multiplications; therefore, multiplying the weight, w, by 𝜕𝜕𝒉𝒉(𝑘𝑘) itself t – k times results in a factor, 𝑤𝑤 𝑡𝑡−𝑘𝑘. As a result, if |𝑤𝑤| < 1, this factor becomes very small when t – k is large. On the other hand, if the weight of the recurrent edge is |𝑤𝑤| > 1, then 𝑤𝑤 𝑡𝑡−𝑘𝑘 becomes very large when t – k is large. Note that large t – k refers to long-range dependencies. We can see that a naive solution to avoid vanishing or exploding gradients can be reached by ensuring |𝑤𝑤| = 1. If you are interested and would like to investigate this in more detail, read On the difficulty of training recurrent neural networks, by R. Pascanu, T. Mikolov, and Y. Bengio, 2012 (https://arxiv.org/pdf/1211.5063.pdf). Basically,

In practice, there are at least three solutions to this problem: • Gradient clipping • TBPTT • LSTM Using gradient clipping, we specify a cut-off or threshold value for the gradients, and we assign this cut-off value to gradient values that exceed this value. In contrast, TBPTT simply limits the number of time steps that the signal can backpropagate after each forward pass. For example, even if the sequence has 100 elements or steps, we may only backpropagate the most recent 20 time steps. While both gradient clipping and TBPTT can solve the exploding gradient problem, the truncation limits the number of steps that the gradient can effectively flow back and properly update the weights. On the other hand, LSTM, designed in 1997 by Sepp Hochreiter and Jürgen Schmidhuber, has been more successful in vanishing and exploding gradient problems while modeling long-range dependencies through the use of memory cells. Let's discuss LSTM in more detail. [ 581 ]

Modeling Sequential Data Using Recurrent Neural Networks

Long short-term memory cells

As stated previously, LSTMs were first introduced to overcome the vanishing gradient problem (Long Short-Term Memory, S. Hochreiter and J. Schmidhuber, Neural Computation, 9(8): 1735-1780, 1997). The building block of an LSTM is a memory cell, which essentially represents or replaces the hidden layer of standard RNNs. In each memory cell, there is a recurrent edge that has the desirable weight, w = 1, as we discussed, to overcome the vanishing and exploding gradient problems. The values associated with this recurrent edge are collectively called the cell state. The unfolded structure of a modern LSTM cell is shown in the following figure:

Notice that the cell state from the previous time step, 𝑪𝑪(𝑡𝑡−1) , is modified to get the cell state at the current time step, 𝑪𝑪(𝑡𝑡) , without being multiplied directly with any weight factor. The flow of information in this memory cell is controlled by several computation units (often called gates) that will be described here. In the previous figure, ⊙ refers to the element-wise product (element-wise multiplication) and ⨁ means element-wise summation (element-wise addition). Furthermore, 𝒙𝒙(𝑡𝑡) refers to the input data at time t, and 𝒉𝒉(𝑡𝑡−1) indicates the hidden units at time t – 1. Four boxes are indicated with an activation function, either the sigmoid function (𝜎𝜎) or tanh, and a set of weights; these boxes apply a linear combination by performing matrix-vector multiplications on their inputs (which are 𝒉𝒉(𝑡𝑡−1) and 𝒙𝒙(𝑡𝑡) ). These units of computation with sigmoid activation functions, whose output units are passed through ⊙, are called gates.

[ 582 ]

Chapter 16

In an LSTM cell, there are three different types of gates, which are known as the forget gate, the input gate, and the output gate: • The forget gate ( 𝒇𝒇𝑡𝑡) allows the memory cell to reset the cell state without growing indefinitely. In fact, the forget gate decides which information is allowed to go through and which information to suppress. Now, 𝒇𝒇𝑡𝑡 is computed as follows:

𝒇𝒇𝑡𝑡 = 𝜎𝜎(𝑾𝑾𝑥𝑥𝑥𝑥 𝒙𝒙(𝑡𝑡) + 𝑾𝑾ℎ𝑓𝑓 𝒉𝒉(𝑡𝑡−1) + 𝒃𝒃𝑓𝑓 )

Note that the forget gate was not part of the original LSTM cell; it was added a few years later to improve the original model (Learning to Forget: Continual Prediction with LSTM, F. Gers, J. Schmidhuber, and F. Cummins, Neural Computation 12, 2451-2471, 2000).

̃ 𝑡𝑡 ) are responsible for updating the • The input gate (𝒊𝒊𝑡𝑡 ) and candidate value ( 𝑪𝑪 cell state. They are computed as follows:

𝒊𝒊𝑡𝑡 = 𝜎𝜎(𝑾𝑾𝑥𝑥𝑥𝑥 𝒙𝒙(𝑡𝑡) + 𝑾𝑾ℎ𝑖𝑖 𝒉𝒉(𝑡𝑡−1) + 𝒃𝒃𝑖𝑖 )

̃ 𝑡𝑡 = tanh(𝑾𝑾𝑥𝑥𝑥𝑥 𝒙𝒙(𝑡𝑡) + 𝑾𝑾ℎ𝑐𝑐 𝒉𝒉(𝑡𝑡−1) + 𝒃𝒃𝑐𝑐 ) 𝑪𝑪

The cell state at time t is computed as follows:

̃ 𝑡𝑡 ) 𝑪𝑪(𝑡𝑡) = (𝑪𝑪(𝑡𝑡−1) ⊙ 𝒇𝒇𝑡𝑡 ) ⨁ (𝒊𝒊𝑡𝑡 ⊙ 𝑪𝑪

• The output gate (𝒐𝒐𝑡𝑡 ) decides how to update the values of hidden units:

𝒐𝒐𝑡𝑡 = 𝜎𝜎(𝑾𝑾𝑥𝑥𝑥𝑥 𝒙𝒙(𝑡𝑡) + 𝑾𝑾ℎ𝑜𝑜 𝒉𝒉(𝑡𝑡−1) + 𝒃𝒃𝑜𝑜 )

Given this, the hidden units at the current time step are computed as follows:

𝒉𝒉(𝑡𝑡) = 𝒐𝒐𝑡𝑡 ⊙ tanh(𝑪𝑪(𝑡𝑡) )

The structure of an LSTM cell and its underlying computations might seem very complex and hard to implement. However, the good news is that TensorFlow has already implemented everything in optimized wrapper functions, which allows us to define our LSTM cells easily and efficiently. We will apply RNNs and LSTMs to realworld datasets later in this chapter. [ 583 ]

Modeling Sequential Data Using Recurrent Neural Networks

Other advanced RNN models LSTMs provide a basic approach for modeling long-range dependencies in sequences. Yet, it is important to note that there are many variations of LSTMs described in literature (An Empirical Exploration of Recurrent Network Architectures, Rafal Jozefowicz, Wojciech Zaremba, and Ilya Sutskever, Proceedings of ICML, 2342-2350, 2015). Also worth noting is a more recent approach, Gated Recurrent Unit (GRU), which was proposed in 2014. GRUs have a simpler architecture than LSTMs; therefore, they are computationally more efficient, while their performance in some tasks, such as polyphonic music modeling, is comparable to LSTMs. If you are interested in learning more about these modern RNN architectures, refer to the paper, Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling, by Junyoung Chung and others, 2014 (https://arxiv.org/pdf/1412.3555v1.pdf).

Implementing RNNs for sequence modeling in TensorFlow

Now that we have covered the underlying theory behind RNNs, we are ready to move on to the more practical portion of this chapter: implementing RNNs in TensorFlow. During the rest of this chapter, we will apply RNNs to two common problem tasks: 1. Sentiment analysis 2. Language modeling These two projects, which we will walk through together in the following pages, are both fascinating but also quite involved. Thus, instead of providing the code all at once, we will break the implementation up into several steps and discuss the code in detail. If you like to have a big picture overview and want to see all the code at once before diving into the discussion, take a look at the code implementation first, which you can view at https://github.com/rasbt/python-machine-learning-book3rd-edition/tree/master/ch16.

[ 584 ]

Chapter 16

Project one – predicting the sentiment of IMDb movie reviews

You may recall from Chapter 8, Applying Machine Learning to Sentiment Analysis, that sentiment analysis is concerned with analyzing the expressed opinion of a sentence or a text document. In this section and the following subsections, we will implement a multilayer RNN for sentiment analysis using a many-to-one architecture. In the next section, we will implement a many-to-many RNN for an application of language modeling. While the chosen examples are purposefully simple to introduce the main concepts of RNNs, language modeling has a wide range of interesting applications, such as building chatbots—giving computers the ability to directly talk and interact with humans.

Preparing the movie review data

In the preprocessing steps in Chapter 8, we created a clean dataset named movie_ data.csv, which we will use again now. First, we will import the necessary modules and read the data into a pandas DataFrame, as follows: >>> >>> >>> >>>

import import import import

tensorflow as tf tensorflow_datasets as tfds numpy as np pandas as pd

>>> df = pd.read_csv('movie_data.csv', encoding='utf-8')

Remember that this data frame, df, consists of two columns, namely 'review' and 'sentiment', where 'review' contains the text of movie reviews (the input features), and 'sentiment' represents the target label we want to predict (0 refers to negative sentiment and 1 refers to positive sentiment). The text component of these movie reviews is sequences of words, and the RNN model classifies each sequence as a positive (1) or negative (0) review. However, before we can feed the data into an RNN model, we need to apply several preprocessing steps: 1. Create a TensorFlow dataset object and split it into separate training, testing, and validation partitions. 2. Identify the unique words in the training dataset.

[ 585 ]

Modeling Sequential Data Using Recurrent Neural Networks

3. Map each unique word to a unique integer and encode the review text into encoded integers (an index of each unique word). 4. Divide the dataset into mini-batches as input to the model. Let's proceed with the first step: creating a TensorFlow dataset from this data frame: >>> ## Step 1: create a dataset >>> target = df.pop('sentiment') >>> ds_raw = tf.data.Dataset.from_tensor_slices( ... (df.values, target.values)) >>> ## inspection: >>> for ex in ds_raw.take(3): ... tf.print(ex[0].numpy()[0][ :50], ex[1]) b'In 1974, the teenager Martha Moxley (Maggie Grace)' 1 b'OK... so... I really like Kris Kristofferson and h' 0 b'***SPOILER*** Do not read this, if you think about' 0

Now, we can split it into training, testing, and validation datasets. The entire dataset contains 50,000 examples. We will keep the first 25,000 examples for evaluation (hold-out testing dataset), and then 20,000 examples will be used for training and 5,000 for validation. The code is as follows: >>> tf.random.set_seed(1) >>> ds_raw = ds_raw.shuffle( ... 50000, reshuffle_each_iteration=False) >>> >>> >>> >>>

ds_raw_test = ds_raw.take(25000) ds_raw_train_valid = ds_raw.skip(25000) ds_raw_train = ds_raw_train_valid.take(20000) ds_raw_valid = ds_raw_train_valid.skip(20000)

To prepare the data for input to a NN, we need to encode it into numeric values, as was mentioned in steps 2 and 3. To do this, we will first find the unique words (tokens) in the training dataset. While finding unique tokens is a process for which we can use Python datasets, it can be more efficient to use the Counter class from the collections package, which is part of Python's standard library. In the following code, we will instantiate a new Counter object (token_counts) that will collect the unique word frequencies. Note that in this particular application (and in contrast to the bag-of-words model), we are only interested in the set of unique words and won't require the word counts, which are created as a side product. To split the text into words (or tokens), the tensorflow_datasets package provides a Tokenizer class. [ 586 ]

Chapter 16

The code for collecting unique tokens is as follows: >>> ## Step 2: find unique tokens (words) >>> from collections import Counter >>> tokenizer = tfds.features.text.Tokenizer() >>> token_counts = Counter() >>> for example in ds_raw_train: ... tokens = tokenizer.tokenize(example[0].numpy()[0]) ... token_counts.update(tokens) >>> print('Vocab-size:', len(token_counts)) Vocab-size: 87007

If you want to learn more about Counter, refer to its documentation at https:// docs.python.org/3/library/collections.html#collections.Counter. Next, we are going to map each unique word to a unique integer. This can be done manually using a Python dictionary, where the keys are the unique tokens (words) and the value associated with each key is a unique integer. However, the tensorflow_datasets package already provides a class, TokenTextEncoder, which we can use to create such a mapping and encode the entire dataset. First, we will create an encoder object from the TokenTextEncoder class by passing the unique tokens (token_counts contains the tokens and their counts, although here, their counts are not needed, so they will be ignored). Calling the encoder.encode() method will then convert its input text into a list of integer values: >>> ## Step 3: encoding unique tokens to integers >>> encoder = tfds.features.text.TokenTextEncoder(token_counts) >>> example_str = 'This is an example!' >>> print(encoder.encode(example_str)) [232, 9, 270, 1123]

Note that there might be some tokens in the validation or testing data that are not present in the training data and are thus not included in the mapping. If we have q tokens (that is the size of token_counts passed to the TokenTextEncoder, which in this case is 87,007), then all tokens that haven't been seen before, and are thus not included in token_counts, will be assigned