Still Life with Woodpecker

Through a series of recent breakthroughs, deep learning has boosted the entire field of machine learning. Now, even programmers who know close to nothing about this technology can use simple, efficient tools to implement programs capable of learning from data. This practical book shows you how.By using concrete examples, minimal theory, and two production-ready Python frameworks—Scikit-Learn and TensorFlow—author Aurélien Géron helps you gain an intuitive understanding of the concepts and tools for building intelligent systems. You'll learn a range of techniques, starting with simple linear regression and progressing to deep neural networks. With exercises in each chapter to help you apply what you've learned, all you need is programming experience to get started.Explore the machine learning landscape, particularly neural netsUse Scikit-Learn to track an example machine-learning project end-to-endExplore several training models, including support vector machines, decision trees, random forests, and ensemble methodsUse the TensorFlow library to build and train neural netsDive into neural net architectures, including convolutional nets, recurrent nets, and deep reinforcement learningLearn techniques for training and scaling deep neural nets Machine Learning in Your Projects......Page 5 Objective and Approach......Page 6 Prerequisites......Page 7 Roadmap......Page 8 Changes in the Second Edition......Page 9 Other Resources......Page 10 Conventions Used in This Book......Page 12 Code Examples......Page 13 O’Reilly Online Learning......Page 14 How to Contact Us......Page 15 Acknowledgments......Page 16 I. The Fundamentals of Machine Learning......Page 20 1. The Machine Learning Landscape......Page 21 What Is Machine Learning?......Page 22 Why Use Machine Learning?......Page 23 Examples of Applications......Page 27 Types of Machine Learning Systems......Page 29 Supervised/Unsupervised Learning......Page 30 Batch and Online Learning......Page 39 Instance-Based Versus Model-Based Learning......Page 42 Insufficient Quantity of Training Data......Page 50 Nonrepresentative Training Data......Page 52 Poor-Quality Data......Page 54 Irrelevant Features......Page 55 Overfitting the Training Data......Page 56 Underfitting the Training Data......Page 58 Stepping Back......Page 59 Hyperparameter Tuning and Model Selection......Page 60 Data Mismatch......Page 62 Exercises......Page 63 Working with Real Data......Page 66 Look at the Big Picture......Page 68 Frame the Problem......Page 69 Select a Performance Measure......Page 71 Check the Assumptions......Page 74 Create the Workspace......Page 75 Download the Data......Page 79 Take a Quick Look at the Data Structure......Page 81 Create a Test Set......Page 85 Discover and Visualize the Data to Gain Insights......Page 91 Visualizing Geographical Data......Page 92 Looking for Correlations......Page 94 Experimenting with Attribute Combinations......Page 98 Data Cleaning......Page 100 Handling Text and Categorical Attributes......Page 104 Custom Transformers......Page 107 Feature Scaling......Page 108 Transformation Pipelines......Page 109 Training and Evaluating on the Training Set......Page 112 Better Evaluation Using Cross-Validation......Page 114 Fine-Tune Your Model......Page 116 Grid Search......Page 117 Randomized Search......Page 119 Analyze the Best Models and Their Errors......Page 120 Evaluate Your System on the Test Set......Page 121 Launch, Monitor, and Maintain Your System......Page 123 Try It Out!......Page 126 Exercises......Page 127 MNIST......Page 130 Training a Binary Classifier......Page 134 Measuring Accuracy Using Cross-Validation......Page 135 Confusion Matrix......Page 137 Precision and Recall......Page 140 Precision/Recall Trade-off......Page 141 The ROC Curve......Page 146 Multiclass Classification......Page 150 Error Analysis......Page 154 Multilabel Classification......Page 159 Multioutput Classification......Page 160 Exercises......Page 163 4. Training Models......Page 166 Linear Regression......Page 167 The Normal Equation......Page 169 Gradient Descent......Page 173 Batch Gradient Descent......Page 178 Stochastic Gradient Descent......Page 182 Mini-batch Gradient Descent......Page 186 Polynomial Regression......Page 188 Learning Curves......Page 191 Regularized Linear Models......Page 196 Ridge Regression......Page 197 Lasso Regression......Page 199 Elastic Net......Page 202 Early Stopping......Page 203 Estimating Probabilities......Page 205 Training and Cost Function......Page 207 Decision Boundaries......Page 208 Softmax Regression......Page 211 Exercises......Page 216 Linear SVM Classification......Page 219 Soft Margin Classification......Page 220 Nonlinear SVM Classification......Page 222 Polynomial Kernel......Page 224 Similarity Features......Page 225 Gaussian RBF Kernel......Page 226 Computational Complexity......Page 227 SVM Regression......Page 228 Decision Function and Predictions......Page 230 Training Objective......Page 231 Quadratic Programming......Page 233 Kernelized SVMs......Page 234 Online SVMs......Page 237 Exercises......Page 238 Training and Visualizing a Decision Tree......Page 240 Making Predictions......Page 241 Estimating Class Probabilities......Page 243 Computational Complexity......Page 244 Regularization Hyperparameters......Page 245 Regression......Page 247 Instability......Page 249 Exercises......Page 250 Voting Classifiers......Page 253 Bagging and Pasting......Page 257 Bagging and Pasting in Scikit-Learn......Page 258 Out-of-Bag Evaluation......Page 259 Random Patches and Random Subspaces......Page 260 Extra-Trees......Page 261 Feature Importance......Page 262 AdaBoost......Page 263 Gradient Boosting......Page 267 Stacking......Page 272 Exercises......Page 276 8. Dimensionality Reduction......Page 279 The Curse of Dimensionality......Page 280 Main Approaches for Dimensionality Reduction......Page 281 Projection......Page 282 Manifold Learning......Page 285 PCA......Page 286 Principal Components......Page 287 Projecting Down to d Dimensions......Page 289 Explained Variance Ratio......Page 290 Choosing the Right Number of Dimensions......Page 291 PCA for Compression......Page 292 Incremental PCA......Page 294 Kernel PCA......Page 295 Selecting a Kernel and Tuning Hyperparameters......Page 296 LLE......Page 299 Other Dimensionality Reduction Techniques......Page 301 Exercises......Page 303 9. Unsupervised Learning Techniques......Page 306 Clustering......Page 307 K-Means......Page 310 Limits of K-Means......Page 323 Using Clustering for Image Segmentation......Page 324 Using Clustering for Preprocessing......Page 326 Using Clustering for Semi-Supervised Learning......Page 328 DBSCAN......Page 332 Other Clustering Algorithms......Page 336 Gaussian Mixtures......Page 338 Anomaly Detection Using Gaussian Mixtures......Page 346 Selecting the Number of Clusters......Page 348 Bayesian Gaussian Mixture Models......Page 353 Other Algorithms for Anomaly and Novelty Detection......Page 358 Exercises......Page 360 II. Neural Networks and Deep Learning......Page 363 10. Introduction to Artificial Neural Networks with Keras......Page 364 From Biological to Artificial Neurons......Page 365 Biological Neurons......Page 366 Logical Computations with Neurons......Page 368 The Perceptron......Page 370 The Multilayer Perceptron and Backpropagation......Page 376 Regression MLPs......Page 380 Classification MLPs......Page 382 Implementing MLPs with Keras......Page 384 Installing TensorFlow 2......Page 386 Building an Image Classifier Using the Sequential API......Page 387 Building a Regression MLP Using the Sequential API......Page 401 Building Complex Models Using the Functional API......Page 402 Using the Subclassing API to Build Dynamic Models......Page 408 Saving and Restoring a Model......Page 410 Using Callbacks......Page 411 Using TensorBoard for Visualization......Page 414 Fine-Tuning Neural Network Hyperparameters......Page 418 Number of Hidden Layers......Page 423 Number of Neurons per Hidden Layer......Page 424 Learning Rate, Batch Size, and Other Hyperparameters......Page 425 Exercises......Page 428 11. Training Deep Neural Networks......Page 434 The Vanishing/Exploding Gradients Problems......Page 435 Glorot and He Initialization......Page 436 Nonsaturating Activation Functions......Page 438 Batch Normalization......Page 443 Gradient Clipping......Page 451 Reusing Pretrained Layers......Page 452 Transfer Learning with Keras......Page 455 Unsupervised Pretraining......Page 457 Pretraining on an Auxiliary Task......Page 459 Momentum Optimization......Page 460 Nesterov Accelerated Gradient......Page 462 AdaGrad......Page 464 Adam and Nadam Optimization......Page 466 Learning Rate Scheduling......Page 471 Avoiding Overfitting Through Regularization......Page 477 l1 and l2 Regularization......Page 478 Dropout......Page 479 Monte Carlo (MC) Dropout......Page 483 Max-Norm Regularization......Page 486 Summary and Practical Guidelines......Page 487 Exercises......Page 489 A Quick Tour of TensorFlow......Page 493 Using TensorFlow like NumPy......Page 497 Tensors and Operations......Page 498 Tensors and NumPy......Page 500 Type Conversions......Page 501 Variables......Page 502 Other Data Structures......Page 503 Custom Loss Functions......Page 504 Saving and Loading Models That Contain Custom Components......Page 506 Custom Activation Functions, Initializers, Regularizers, and Constraints......Page 508 Custom Metrics......Page 510 Custom Layers......Page 514 Custom Models......Page 518 Losses and Metrics Based on Model Internals......Page 521 Computing Gradients Using Autodiff......Page 524 Custom Training Loops......Page 529 TensorFlow Functions and Graphs......Page 533 AutoGraph and Tracing......Page 535 TF Function Rules......Page 537 Exercises......Page 539 13. Loading and Preprocessing Data with TensorFlow......Page 543 The Data API......Page 544 Chaining Transformations......Page 545 Shuffling the Data......Page 547 Preprocessing the Data......Page 551 Putting Everything Together......Page 553 Prefetching......Page 554 Using the Dataset with tf.keras......Page 556 The TFRecord Format......Page 558 A Brief Introduction to Protocol Buffers......Page 559 TensorFlow Protobufs......Page 561 Loading and Parsing Examples......Page 563 Handling Lists of Lists Using the SequenceExample Protobuf......Page 565 Preprocessing the Input Features......Page 566 Encoding Categorical Features Using One-Hot Vectors......Page 567 Encoding Categorical Features Using Embeddings......Page 570 Keras Preprocessing Layers......Page 576 TF Transform......Page 579 The TensorFlow Datasets (TFDS) Project......Page 581 Exercises......Page 582 14. Deep Computer Vision Using Convolutional Neural Networks......Page 587 The Architecture of the Visual Cortex......Page 588 Convolutional Layers......Page 589 Filters......Page 592 Stacking Multiple Feature Maps......Page 593 TensorFlow Implementation......Page 596 Memory Requirements......Page 600 Pooling Layers......Page 601 TensorFlow Implementation......Page 603 CNN Architectures......Page 606 LeNet-5......Page 609 AlexNet......Page 611 GoogLeNet......Page 615 ResNet......Page 620 Xception......Page 625 SENet......Page 627 Implementing a ResNet-34 CNN Using Keras......Page 630 Using Pretrained Models from Keras......Page 632 Pretrained Models for Transfer Learning......Page 634 Classification and Localization......Page 638 Object Detection......Page 640 Fully Convolutional Networks......Page 642 You Only Look Once (YOLO)......Page 645 Semantic Segmentation......Page 648 Exercises......Page 653 15. Processing Sequences Using RNNs and CNNs......Page 658 Recurrent Neurons and Layers......Page 659 Memory Cells......Page 662 Input and Output Sequences......Page 663 Training RNNs......Page 665 Forecasting a Time Series......Page 666 Baseline Metrics......Page 668 Implementing a Simple RNN......Page 669 Deep RNNs......Page 671 Forecasting Several Time Steps Ahead......Page 673 Fighting the Unstable Gradients Problem......Page 679 Tackling the Short-Term Memory Problem......Page 682 Exercises......Page 692 16. Natural Language Processing with RNNs and Attention......Page 695 Generating Shakespearean Text Using a Character RNN......Page 696 Creating the Training Dataset......Page 697 How to Split a Sequential Dataset......Page 698 Chopping the Sequential Dataset into Multiple Windows......Page 699 Building and Training the Char-RNN Model......Page 702 Using the Char-RNN Model......Page 703 Generating Fake Shakespearean Text......Page 704 Stateful RNN......Page 705 Sentiment Analysis......Page 708 Masking......Page 713 Reusing Pretrained Embeddings......Page 716 An Encoder–Decoder Network for Neural Machine Translation......Page 718 Bidirectional RNNs......Page 722 Beam Search......Page 724 Attention Mechanisms......Page 726 Visual Attention......Page 730 Attention Is All You Need: The Transformer Architecture......Page 732 Recent Innovations in Language Models......Page 745 Exercises......Page 747 17. Representation Learning and Generative Learning Using Autoencoders and GANs......Page 752 Efficient Data Representations......Page 754 Performing PCA with an Undercomplete Linear Autoencoder......Page 756 Stacked Autoencoders......Page 758 Implementing a Stacked Autoencoder Using Keras......Page 759 Visualizing the Reconstructions......Page 760 Visualizing the Fashion MNIST Dataset......Page 762 Unsupervised Pretraining Using Stacked Autoencoders......Page 763 Tying Weights......Page 765 Training One Autoencoder at a Time......Page 766 Convolutional Autoencoders......Page 768 Recurrent Autoencoders......Page 769 Denoising Autoencoders......Page 770 Sparse Autoencoders......Page 772 Variational Autoencoders......Page 776 Generating Fashion MNIST Images......Page 782 Generative Adversarial Networks......Page 784 The Difficulties of Training GANs......Page 789 Deep Convolutional GANs......Page 792 Progressive Growing of GANs......Page 796 StyleGANs......Page 799 Exercises......Page 803 Learning to Optimize Rewards......Page 806 Policy Search......Page 808 Introduction to OpenAI Gym......Page 810 Neural Network Policies......Page 815 Evaluating Actions: The Credit Assignment Problem......Page 817 Policy Gradients......Page 819 Markov Decision Processes......Page 825 Temporal Difference Learning......Page 830 Q-Learning......Page 831 Exploration Policies......Page 833 Approximate Q-Learning and Deep Q-Learning......Page 834 Implementing Deep Q-Learning......Page 835 Fixed Q-Value Targets......Page 841 Double DQN......Page 842 Dueling DQN......Page 843 The TF-Agents Library......Page 844 Installing TF-Agents......Page 846 TF-Agents Environments......Page 847 Environment Specifications......Page 848 Environment Wrappers and Atari Preprocessing......Page 849 Training Architecture......Page 854 Creating the Deep Q-Network......Page 856 Creating the DQN Agent......Page 858 Creating the Replay Buffer and the Corresponding Observer......Page 860 Creating Training Metrics......Page 862 Creating the Collect Driver......Page 863 Creating the Dataset......Page 865 Creating the Training Loop......Page 868 Overview of Some Popular RL Algorithms......Page 870 Exercises......Page 872 19. Training and Deploying TensorFlow Models at Scale......Page 875 Using TensorFlow Serving......Page 876 Creating a Prediction Service on GCP AI Platform......Page 888 Using the Prediction Service......Page 894 Deploying a Model to a Mobile or Embedded Device......Page 898 Using GPUs to Speed Up Computations......Page 904 Getting Your Own GPU......Page 906 Using a GPU-Equipped Virtual Machine......Page 909 Colaboratory......Page 910 Managing the GPU RAM......Page 912 Placing Operations and Variables on Devices......Page 915 Parallel Execution Across Multiple Devices......Page 918 Model Parallelism......Page 922 Data Parallelism......Page 925 Training at Scale Using the Distribution Strategies API......Page 931 Training a Model on a TensorFlow Cluster......Page 933 Running Large Training Jobs on Google Cloud AI Platform......Page 938 Black Box Hyperparameter Tuning on AI Platform......Page 940 Exercises......Page 942 Thank You!......Page 943 Chapter 1: The Machine Learning Landscape......Page 946 Chapter 2: End-to-End Machine Learning Project......Page 949 Chapter 3: Classification......Page 950 Chapter 4: Training Models......Page 951 Chapter 5: Support Vector Machines......Page 954 Chapter 6: Decision Trees......Page 957 Chapter 7: Ensemble Learning and Random Forests......Page 959 Chapter 8: Dimensionality Reduction......Page 961 Chapter 9: Unsupervised Learning Techniques......Page 964 Chapter 10: Introduction to Artificial Neural Networks with Keras......Page 967 Chapter 11: Training Deep Neural Networks......Page 971 Chapter 12: Custom Models and Training with TensorFlow......Page 974 Chapter 13: Loading and Preprocessing Data with TensorFlow......Page 978 Chapter 14: Deep Computer Vision Using Convolutional Neural Networks......Page 983 Chapter 15: Processing Sequences Using RNNs and CNNs......Page 988 Chapter 16: Natural Language Processing with RNNs and Attention......Page 992 Chapter 17: Representation Learning and Generative Learning Using Autoencoders and GANs......Page 995 Chapter 18: Reinforcement Learning......Page 998 Chapter 19: Training and Deploying TensorFlow Models at Scale......Page 1002 Frame the Problem and Look at the Big Picture......Page 1008 Get the Data......Page 1009 Explore the Data......Page 1010 Prepare the Data......Page 1011 Shortlist Promising Models......Page 1012 Fine-Tune the System......Page 1013 Launch!......Page 1014 C. SVM Dual Problem......Page 1016 Manual Differentiation......Page 1020 Finite Difference Approximation......Page 1021 Forward-Mode Autodiff......Page 1022 Reverse-Mode Autodiff......Page 1026 Hopfield Networks......Page 1030 Boltzmann Machines......Page 1032 Restricted Boltzmann Machines......Page 1035 Deep Belief Nets......Page 1036 Self-Organizing Maps......Page 1039 Strings......Page 1042 Ragged Tensors......Page 1044 Sparse Tensors......Page 1045 Tensor Arrays......Page 1046 Sets......Page 1047 Queues......Page 1048 TF Functions and Concrete Functions......Page 1051 Exploring Function Definitions and Graphs......Page 1053 A Closer Look at Tracing......Page 1055 Using AutoGraph to Capture Control Flow......Page 1057 Handling Variables and Other Resources in TF Functions......Page 1059 Using TF Functions with tf.keras (or Not)......Page 1061 Index......Page 1063 Through a series of recent breakthroughs, deep learning has boosted the entire field of machine learning. Now, even programmers who know close to nothing about this technology can use simple, efficient tools to implement programs capable of learning from data. The updated edition of this best-selling book uses concrete examples, minimal theory, and two production-ready Python frameworks-Scikit-Learn and TensorFlow 2-to help you gain an intuitive understanding of the concepts and tools for building intelligent systems. Practitioners will learn a range of techniques that they can quickly put to use on the job. Part 1 employs Scikit-Learn to introduce fundamental machine learning tasks, such as simple linear regression. Part 2, which has been significantly updated, employs Keras and TensorFlow 2 to guide the reader through more advanced machine learning methods using deep neural networks. With exercises in each chapter to help you apply what you've learned, all you need is programming experience to get started. NEW FOR THE SECOND EDITION:Updated all code to TensorFlow 2 ; Introduced the high-level Keras API ; New and expanded coverage including TensorFlow's Data API, Eager Execution, Estimators API, deploying on Google Cloud ML, handling time series, embeddings and more With Early Release ebooks, you get books in their earliest form-the author's raw and unedited content as he or she writes-so you can take advantage of these technologies long before the official release of these titles. You'll also receive updates when significant changes are made, new chapters are available, and the final ebook bundle is released Through a series of recent breakthroughs, deep learning has boosted the entire field of machine learning. Now, even programmers who know close to nothing about this technology can use simple, efficient tools to implement programs capable of learning from data. The updated edition of this best-selling book uses concrete examples, minimal theory, and two production-ready Python frameworks--Scikit-Learn and TensorFlow 2--to help you gain an intuitive understanding of the concepts and tools for building intelligent systems. Practitioners will learn a range of techniques that they can quickly put to use on the job. Part 1 employs Scikit-Learn to introduce fundamental machine learning tasks, such as simple linear regression. Part 2, which has been significantly updated, employs Keras and TensorFlow 2 to guide the reader through more advanced machine learning methods using deep neural networks. With exercises in each chapter to help you apply what you've learned, all you need is programming experience to get started. NEW FOR THE SECOND EDITION: Updated all code to TensorFlow 2Introduced the high-level Keras APINew and expanded coverage including TensorFlow's Data API, Eager Execution, Estimators API, deploying on Google Cloud ML, handling time series, embeddings and more.