GANs in Action: Deep learning with Generative Adversarial Networks

Description

GANs in Action teaches you how to build and train your own Generative Adversarial Networks, one of the most important innovations in deep learning. In this book, you'll learn how to start building your own simple adversarial system as you explore the foundation of GAN architecture: the generator and discriminator networks. About the Technology Generative Adversarial Networks, GANs, are an incredible AI technology capable of creating images, sound, and videos that are indistinguishable from the "real thing." By pitting two neural networks against each other—one to generate fakes and one to spot them—GANs rapidly learn to produce photo-realistic faces and other media objects. With the potential to produce stunningly realistic animations or shocking deepfakes, GANs are a huge step forward in deep learning systems. About the Book GANs in Action teaches you to build and train your own Generative Adversarial Networks. You'll start by creating simple generator and discriminator networks that are the foundation of GAN architecture. Then, following numerous hands-on examples, you'll train GANs to generate high-resolution images, image-to-image translation, and targeted data generation. Along the way, you'll find pro tips for making your system smart, effective, and fast. What's inside • Building your first GAN • Handling the progressive growing of GANs • Practical applications of GANs • Troubleshooting your system About the Reader For data professionals with intermediate Python skills, and the basics of deep learning-based image processing. About the Author Jakub Langr is a Computer Vision Cofounder at Founders Factory (YEPIC.AI). Vladimir Bok is a Senior Product Manager overseeing machine learning infrastructure and research teams at a New York-based startup. GANs in Action brief contents contents preface Jakub Langr Vladimir Bok acknowledgments Jakub Langr Vladimir Bok about this book Who should read this book About the code liveBook discussion forum Other online resources How this book is organized: a roadmap About the authors about the cover illustration Saint-Sauveur Part 1—Introduction to GANs and generative modeling 1 Introduction to GANs 1.1 What are Generative Adversarial Networks? 1.2 How do GANs work? 1.3 GANs in action 1.3.1 GAN training 1.3.2 Reaching equilibrium 1.4 Why study GANs? Summary 2 Intro to generative modeling with autoencoders 2.1 Introduction to generative modeling 2.2 How do autoencoders function on a high level? 2.3 What are autoencoders to GANs? 2.4 What is an autoencoder made of? 2.5 Usage of autoencoders 2.6 Unsupervised learning 2.6.1 New take on an old idea 2.6.2 Generation using an autoencoder 2.6.3 Variational autoencoder 2.7 Code is life 2.8 Why did we try aGAN? Summary 3 Your first GAN: Generating handwritten digits 3.1 Foundations of GANs: Adversarial training 3.1.1 Cost functions 3.1.2 Training process 3.2 The Generator and the Discriminator 3.2.1 Conflicting objectives 3.2.2 Confusion matrix 3.3 GAN training algorithm 3.4 Tutorial: Generating handwritten digits 3.4.1 Importing modules and specifying model input dimensions 3.4.2 Implementing the Generator 3.4.3 Implementing the Discriminator 3.4.4 Building the model 3.4.5 Training 3.4.6 Outputting sample images 3.4.7 Running the model 3.4.8 Inspecting the results 3.5 Conclusion Summary 4 Deep Convolutional GAN 4.1 Convolutional neural networks 4.1.1 Convolutional filters 4.1.2 Parameter sharing 4.1.3 ConvNets visualized 4.2 Brief history of the DCGAN 4.3 Batch normalization 4.3.1 Understanding normalization 4.3.2 Computing batch normalization 4.4 Tutorial: Generating handwritten digits with DCGAN 4.4.1 Importing modules and specifying model input dimensions 4.4.2 Implementing the Generator 4.4.3 Implementing the Discriminator 4.4.4 Building and running the DCGAN 4.4.5 Model output 4.5 Conclusion Summary Part 2—Advanced topics in GANs 5 Training and common challenges: GANing for success 5.1 Evaluation 5.1.1 Evaluation framework 5.1.2 Inception score 5.1.3 Fréchet inception distance 5.2 Training challenges 5.2.1 Adding network depth 5.2.2 Game setups 5.2.3 Min-Max GAN 5.2.4 Non-Saturating GAN 5.2.5 When to stop training 5.2.6 Wasserstein GAN 5.3 Summary of game setups 5.4 Training hacks 5.4.1 Normalizations of inputs 5.4.2 Batch normalization 5.4.3 Gradient penalties 5.4.4 Train the Discriminator more 5.4.5 Avoid sparse gradients 5.4.6 Soft and noisy labels Summary 6 Progressing with GANs 6.1 Latent space interpolation 6.2 They grow up so fast 6.2.1 Progressive growing and smoothing of higher-resolution layers 6.2.2 Example implementation 6.2.3 Mini-batch standard deviation 6.2.4 Equalized learning rate 6.2.5 Pixel-wise feature normalization in the generator 6.3 Summary of key innovations 6.4 TensorFlow Hub and hands-on 6.5 Practical applications Summary 7 Semi-Supervised GAN 7.1 Introducing the Semi-Supervised GAN 7.1.1 What is a Semi-Supervised GAN? 7.1.2 Architecture 7.1.3 Training process 7.1.4 Training objective 7.2 Tutorial: Implementing a Semi-Supervised GAN 7.2.1 Architecture diagram 7.2.2 Implementation 7.2.3 Setup 7.2.4 The dataset 7.2.5 The Generator 7.2.6 The Discriminator 7.2.7 Building the model 7.2.8 Training 7.3 Comparison to a fully supervised classifier 7.4 Conclusion Summary 8 Conditional GAN 8.1 Motivation 8.2 What is Conditional GAN? 8.2.1 CGAN Generator 8.2.2 CGAN Discriminator 8.2.3 Summary table 8.2.4 Architecture diagram 8.3 Tutorial: Implementing a Conditional GAN 8.3.1 Implementation 8.3.2 Setup 8.3.3 CGAN Generator 8.3.4 CGAN Discriminator 8.3.5 Building the model 8.3.6 Training 8.3.7 Outputting sample images 8.3.8 Training the model 8.3.9 Inspecting the output: Targeted data generation 8.4 Conclusion Summary 9 CycleGAN 9.1 Image-to-image translation 9.2 Cycle-consistency loss: There and back aGAN 9.3 Adversarial loss 9.4 Identity loss 9.5 Architecture 9.5.1 CycleGAN architecture: building the network 9.5.2 Generator architecture 9.5.3 Discriminator architecture 9.6 Object-oriented design of GANs 9.7 Tutorial: CycleGAN 9.7.1 Building the network 9.7.2 Building the Generator 9.7.3 Building the Discriminator 9.7.4 Training the CycleGAN 9.7.5 Running CycleGAN 9.8 Expansions, augmentations, and applications 9.8.1 Augmented CycleGAN 9.8.2 Applications Summary Part 3—Where to go from here 10 Adversarial examples 10.1 Context of adversarial examples 10.2 Lies, damned lies, and distributions 10.3 Use and abuse of training 10.4 Signal and the noise 10.5 Not all hope is lost 10.6 Adversaries to GANs 10.7 Conclusion Summary 11 Practical applications of GANs 11.1 GANs in medicine 11.1.1 Using GANs to improve diagnostic accuracy 11.1.2 Methodology 11.1.3 Results 11.2 GANs in fashion 11.2.1 Using GANs to design fashion 11.2.2 Methodology 11.2.3 Creating new items matching individual preferences 11.2.4 Adjusting existing items to better match individual preferences 11.3 Conclusion Summary 12 Looking ahead 12.1 Ethics 12.2 GAN innovations 12.2.1 Relativistic GAN 12.2.2 Self-Attention GAN 12.2.3 BigGAN 12.3 Further reading 12.4 Looking back and closing thoughts Summary index