GANs Explained

This text is a reference for learning and mastering GANs (applied perspective). In these pages, we provide links to official code and documentation, discuss how the GAN networks function, present code examples in editable sandboxes with explanations.

Cover image

Copyright © 2023 Benjamin Kenwright.

The examples for this book were developed by Benjamin Kenwright and is licensed under the MIT license.

The cover art was created by Benjamin Kenwright and is a fun illustration depicting robots learning to paint through trial and error..

Special thanks to reviewers, colleagues and friends for input and discussion while developing the initial versions of the text.

Contents

Preface

Introduction to Generative Adversarial Networks (GANs)

The Magic of GANs

What are GANs?

A Brief History of GANs

How GANs Compare to Other Generative Models

Examples (Generating Faces)

Getting Ready for the Adventure

Setting Up the Python Environment

Data Preprocessing: The Canvas for Creativity

GPU Acceleration for Faster Training

Checking if PyTorch has GPU Support

Understanding the GAN

The Intuition Behind GANs

Training (Learning)

Common Types of GANs

Vanilla GAN

Deep Convolutional GANs (DCGANS)

Conditional GAN (cGAN)

Wasserstein GAN (WGAN)

Progressive GAN (PGAN)

StyleGAN and StyleGAN2

GAN Landscape

Common Problems with GANs

Vanishing Gradients

Mode Collapse

GANS are very sensitive to hyperparameters

Building Your First GAN

Vanilla GANs

Creating the Generator

Implementing the Generator in Python

Understanding the Generator Architecture

Unveiling the Discriminator

Coding the Discriminator in Python

Decoding the Discriminator Architecture

Testing Discriminator

Loading Real Data (Test Images MNIST)

The Dance Begins: Training Your GAN

Vanilla GAN

How it all works

Results

Training GANs: Challenges and Solutions

Improving GAN Training Stability

Evaluating GAN Performance

Things to Try

DCGAN: Deep Convolutional GANs

Deep Convolutional Networks

CelebA Image Dataset

Custom Dataset Downloader

Generator

Testing Generator (RGB Output)

Discriminator

Training

Setup

Training Loop

Output

Conditional GANs: Steering Creativity with Labels

Implementing Conditional GANs in Python

Generating Images with Specific Attributes

Fashion MNIST

Implementation (Beyond Purely Random)

Output

Taking it Further

Image-to-Image Translation with GANs

Image-to-Image Translation with Paired Data

Pix2Pix (Paired Images)

CycleGan (Unpaired Images)

Progressive GANS

Generate Large Images by Progressively Adding Layers

How to Progressively Grow a GAN

Growing the Generator

Growing the Discriminator

Examples

Synthetic Photographs of Celebrity Faces

Synthetic Photographs of Objects

Further Reading

Super-Resolution with GANs

Enhancing Image Quality with SRGAN

Single-Image Super-Resolution: ESRGAN

GAN Art and Style Transfer

Creating Art with StyleGAN

Artistic Style Transfer with GANs

Neural Style Transfer: VGG19 and Gram Matrix

Music Generation with GANs

GANs in the World of Music

MIDI Data Representation for Music GANs

Melody Generation with MIDI GAN

Text-to-Image Synthesis

Bridging Text and Images with StackGAN

Attention Mechanisms for Improved Text-to-Image GANs

Semantic Image Synthesis with DALL-E

GANs for Anomaly Detection

Identifying Outliers with AnoGAN

One-Class GAN for Novelty Detection

GANs for Fraud Detection and Cybersecurity

Simpson Face Generator

The Ethics of AI and GANs

Addressing Bias and Fairness in GANs

Responsible Data Collection and Use

Implications of Deepfakes and Misinformation

Appendix

Data Sets

List of the Torchvision Datasets

MNIST Data Set

CIFAR10

CIFAR-100

Fashion MNIST

SVHN (Street View House Numbers)

STL-10

CelebA