Deep Learning with Python

Transcription

1 Deep Learning with Python A Hands-on Introduction Nikhil Ketkar

2 Deep Learning with Python: A Hands-on Introduction Nikhil Ketkar Bangalore, Karnataka, India ISBN-13 (pbk): ISBN-13 (electronic): DOI / Library of Congress Control Number: Copyright 2017 by Nikhil Ketkar This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. Trademarked names, logos, and images may appear in this book. Rather than use a trademark symbol with every occurrence of a trademarked name, logo, or image we use the names, logos, and images only in an editorial fashion and to the benefit of the trademark owner, with no intention of infringement of the trademark. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect to the material contained herein. Managing Director: Welmoed Spahr Editorial Director: Todd Green Acquisitions Editor: Celestin Suresh John Development Editor: Matthew Moodie and Anila Vincent Technical Reviewer: Jojo Moolayail Coordinating Editor: Prachi Mehta Copy Editor: Larissa Shmailo Compositor: SPi Global Indexer: SPi Global Artist: SPi Global Cover image designed by Freepik Distributed to the book trade worldwide by Springer Science+Business Media New York, 233 Spring Street, 6th Floor, New York, NY Phone SPRINGER, fax (201) , orders-ny@springer-sbm.com, or visit Apress Media, LLC is a California LLC and the sole member (owner) is Springer Science + Business Media Finance Inc (SSBM Finance Inc). SSBM Finance Inc is a Delaware corporation. For information on translations, please rights@apress.com, or visit rights-permissions. Apress titles may be purchased in bulk for academic, corporate, or promotional use. ebook versions and licenses are also available for most titles. For more information, reference our Print and ebook Bulk Sales web page at Any source code or other supplementary material referenced by the author in this book is available to readers on GitHub via the book s product page, located at For more detailed information, please visit Printed on acid-free paper

3 To Aditi.

4 Contents at a Glance About the Author... xiii About the Technical Reviewer...xv Acknowledgments...xvii Chapter 1: Introduction to Deep Learning... 1 Chapter 2: Machine Learning Fundamentals... 7 Chapter 3: Feed Forward Neural Networks Chapter 4: Introduction to Theano Chapter 5: Convolutional Neural Networks Chapter 6: Recurrent Neural Networks Chapter 7: Introduction to Keras Chapter 8: Stochastic Gradient Descent Chapter 9: Automatic Differentiation Chapter 10: Introduction to GPUs Chapter 11: Introduction to Tensorflow Chapter 12: Introduction to PyTorch Chapter 13: Regularization Techniques Chapter 14: Training Deep Learning Models Index v

5 Contents About the Author... xiii About the Technical Reviewer...xv Acknowledgments...xvii Chapter 1: Introduction to Deep Learning... 1 Historical Context... 1 Advances in Related Fields... 3 Prerequisites... 3 Overview of Subsequent Chapters... 4 Installing the Required Libraries... 5 Chapter 2: Machine Learning Fundamentals... 7 Intuition... 7 Binary Classification... 7 Regression... 8 Generalization... 9 Regularization Summary Chapter 3: Feed Forward Neural Networks Unit Overall Structure of a Neural Network Expressing the Neural Network in Vector Form Evaluating the output of the Neural Network Training the Neural Network vii

6 Contents Deriving Cost Functions using Maximum Likelihood Binary Cross Entropy Cross Entropy Squared Error Summary of Loss Functions Types of Units/Activation Functions/Layers Linear Unit Sigmoid Unit Softmax Layer Rectified Linear Unit (ReLU) Hyperbolic Tangent Neural Network Hands-on with AutoGrad Summary Chapter 4: Introduction to Theano What is Theano Theano Hands-On Summary Chapter 5: Convolutional Neural Networks Convolution Operation Pooling Operation Convolution-Detector-Pooling Building Block Convolution Variants Intuition behind CNNs Summary Chapter 6: Recurrent Neural Networks RNN Basics Training RNNs Bidirectional RNNs Gradient Explosion and Vanishing viii

7 Contents Gradient Clipping Long Short Term Memory Summary Chapter 7: Introduction to Keras Summary Chapter 8: Stochastic Gradient Descent Optimization Problems Method of Steepest Descent Batch, Stochastic (Single and Mini-batch) Descent Batch Stochastic Single Example Stochastic Mini-batch Batch vs. Stochastic Challenges with SGD Local Minima Saddle Points Selecting the Learning Rate Slow Progress in Narrow Valleys Algorithmic Variations on SGD Momentum Nesterov Accelerated Gradient (NAS) Annealing and Learning Rate Schedules Adagrad RMSProp Adadelta Adam Resilient Backpropagation Equilibrated SGD ix

8 Contents Tricks and Tips for using SGD Preprocessing Input Data Choice of Activation Function Preprocessing Target Value Initializing Parameters Shuffling Data Batch Normalization Early Stopping Gradient Noise Parallel and Distributed SGD Hogwild Downpour Hands-on SGD with Downhill Summary Chapter 9: Automatic Differentiation Numerical Differentiation Symbolic Differentiation Automatic Differentiation Fundamentals Forward/Tangent Linear Mode Reverse/Cotangent/Adjoint Linear Mode Implementation of Automatic Differentiation Hands-on Automatic Differentiation with Autograd Summary Chapter 10: Introduction to GPUs Summary Chapter 11: Introduction to Tensorflow Summary Chapter 12: Introduction to PyTorch Summary x

9 Contents Chapter 13: Regularization Techniques Model Capacity, Overfitting, and Underfitting Regularizing the Model Early Stopping Norm Penalties Dropout Summary Chapter 14: Training Deep Learning Models Performance Metrics Data Procurement Splitting Data for Training/Validation/Test Establishing Achievable Limits on the Error Rate Establishing the Baseline with Standard Choices Building an Automated, End-to-End Pipeline Orchestration for Visibility Analysis of Overfitting and Underfitting Hyper-Parameter Tuning Summary Index xi

10 About the Author Nikhil Ketkar currently leads the Machine Learning Platform team at Flipkart, India s largest e-commerce company. He received his PhD from Washington State University. Following that, he conducted postdoctoral research at University of North Carolina at Charotte, which was followed by a brief stint in high frequency trading at TransMarket in Chicago. More recently, he led the data mining team in Guavus, a startup doing big data analytics in the telecom domain and Indix, a startup doing data science in the e-commerce domain. His research interests include machine learning and graph theory. xiii

11 About the Technical Reviewer Jojo Moolayil is a data scientist and author of Smarter Decisions The Intersection of Internet of Things and Decision Science. With over four years of industrial experience in data science, decision science, and IoT, he has worked with industry leaders on high-impact and critical projects across multiple verticals. He is currently associated with General Electric, a pioneer and leader in data science for industrial IoT, and lives in Bengaluru, the Silicon Valley of India. He was born and raised in Pune, India and graduated from the University of Pune with a major in information technology engineering. He started his career with Mu Sigma, the world s largest pure play analytics provider, and worked with the leaders of many Fortune 50 clients. One of the early enthusiasts to venture into IoT analytics, he now focuses on solving decision science problems for industrial IoT use cases. As a part of his role at GE, he also develops data science and decision science products and platforms for industrial IoT. xv

12 Acknowledgments I would like to thank my colleagues at Flipkart and Indix, and the technical reviewers for their feedback and comments. I will also like to thank Charu Mudholkar for proofreading in the final stages. xvii