This book opens with an introduction to deep learning and generative models, and their applications in artificial intelligence (AI). You will then learn how to build, evaluate, and improve your first GAN with the help of easy-to-follow examples. The next few chapters will guide you through training a GAN model to produce and improve high-resolution images. You will also learn how to implement conditional GANs that give you the ability to control characteristics of GAN outputs. You will build on your knowledge further by exploring a new training methodology for progressive growing of GANs. Moving on, you'll gain insights into state-of-the-art models in image synthesis, speech enhancement, and natural language generation using GANs. In addition to this, you'll be able to identify GAN samples with TequilaGAN.
Author(s): Rafael Valle
Publisher: Packt
Year: 2019
Language: English
Pages: 426
Title Page
Copyright and Credits
Hands-On Generative Adversarial Networks with Keras
About Packt
Why subscribe?
Packt.com
Foreword
Contributors
About the author
About the reviewer
Packt is searching for authors like you
Preface
Who this book is for
What this book covers
To get the most out of this book
Download the example code files
Download the color images
Conventions used
Get in touch
Reviews
Section 1: Introduction and Environment Setup
Deep Learning Basics and Environment Setup
Deep learning basics
Artificial Neural Networks (ANNs)
The parameter estimation
Backpropagation
Loss functions
L1 loss
L2 loss
Categorical crossentropy loss
Non-linearities
Sigmoid
Tanh
ReLU
A fully connected layer
The convolution layer
The max pooling layer
Deep learning environment setup
Installing Anaconda and Python
Setting up a virtual environment in Anaconda
Installing TensorFlow
Installing Keras
Installing data visualization and machine learning libraries
The matplotlib library
The Jupyter library
The scikit-learn library
NVIDIA's CUDA Toolkit and cuDNN
The deep learning environment test
Summary
Introduction to Generative Models
Discriminative and generative models compared
Comparing discriminative and generative models
Generative models
Autoregressive models
Variational autoencoders
Reversible flows
Generative adversarial networks
GANs – building blocks
The discriminator
The generator
Real and fake data
Random noise
Discriminator and generator loss
GANs – strengths and weaknesses
Summary
Section 2: Training GANs
Implementing Your First GAN
Technical requirements
Imports
Implementing a Generator and Discriminator
Generator
Discriminator
Auxiliary functions
Training your GAN
Summary
Further reading
Evaluating Your First GAN
The evaluation of GANs
Image quality
Image variety
Domain specifications
Qualitative methods
k-nearest neighbors
Mode analysis
Other methods
Quantitative methods
The Inception score
The Frechét Inception Distance
Precision, Recall, and the F1 Score
GANs and the birthday paradox
Summary
Improving Your First GAN
Technical requirements
Challenges in training GANs
Mode collapse and mode drop
Training instability
Sensitivity to hyperparameter initialization
Vanishing gradients
Tricks of the trade
Tracking failure
Working with labels
Working with discrete inputs
Adding noise
Input normalization
Modified objective function
Distribute latent vector
Weight normalization
Avoid sparse gradients
Use a different optimizer
Learning rate schedule
GAN model architectures
ResNet GAN
GAN algorithms and loss functions
Least Squares GAN
Wasserstein GAN
Wasserstein GAN with gradient penalty
Relativistic GAN
Summary
Section 3: Application of GANs in Computer Vision, Natural Language Processing, and Audio
Progressive Growing of GANs
Technical requirements
Progressive Growing of GANs
Increasing variation using minibatch standard deviation
Normalization in the generator and the discriminator
Pixelwise feature vector normalization in the generator
Experimental setup
Training
Helper functions
Initializations
Training loops
Model implementation
Custom layers
The discriminator
The generator
GANs
Summary
Generation of Discrete Sequences Using GANs
Technical requirements
Natural language generation with GANs
Experimental setup
Data
Auxiliary training functions
Training
Imports and global variables
Initializations
Training loop
Logging
Model implementation
Helper functions
Discriminator
Generator
Inference
Model trained on words
Model trained on characters
Summary
Text-to-Image Synthesis with GANs
Technical Requirements
Text-to-image synthesis
Experimental setup
Data utils
Logging utils
Training
Initial setup
The training loop
Model implementation
Wrapper
Discriminator
Generator
Improving the baseline model
Training
Inference
Sampling the generator
Interpolation in the Latent Space
Interpolation in the text-embedding space
Inferencing with arithmetic in the text-embedding space
Summary
TequilaGAN - Identifying GAN Samples
Technical requirements
Identifying GAN samples
Related work
Feature extraction
Centroid
Slope
Metrics
Jensen-Shannon divergence
Kolgomorov-Smirnov Two-Sample test
Experiments
MNIST
Summary
References
Whats next in GANs
What we've GANed so far
Generative models
Architectures
Loss functions
Tricks of the trade
Implementations
Unanswered questions in GANs
Are some losses better than others?
Do GANs do distribution learning?
All about that inductive bias
How can you kill a GAN?
Artistic GANs
Visual arts
GANGogh
Image inpainting
Vid2Vid
GauGAN
Sonic arts
MuseGAN
GANSynth
Recent and yet-to-be-explored GAN topics
Summary
Closing remarks
Further reading
