Generative Deep Learning, 2nd Edition

Chapter 4. Generative Adversarial Networks

Chapter Goals

In this chapter you will:

• Learn about the architectural design of a generative adversarial network (GAN).

• Build and train a deep convolutional GAN (DCGAN) from scratch using Keras.

• Use the DCGAN to generate new images.

• Understand some of the common problems faced when training a DCGAN.

• Learn how the Wasserstein GAN (WGAN) architecture addresses these problems.

• Understand additional enhancements that can be made to the WGAN, such as incorporating a gradient penalty (GP) term into the loss function.

• Build a WGAN-GP from scratch using Keras.

• Use the WGAN-GP to generate faces.

• Learn how a conditional GAN (CGAN) gives you the ability to condition generated output on a given label.

• Build and train a CGAN in Keras and use it to manipulate a generated image.

In 2014, Ian Goodfellow et al. presented a paper entitled “Generative Adversarial Nets”¹ at the Neural Information Processing Systems conference (NeurIPS) in Montreal. The introduction of generative adversarial networks (or GANs, as they are more commonly known) is now regarded as a key turning point in the history of generative modeling, as the core ideas presented in this paper have spawned some of the most successful and impressive generative models ever created.

This chapter will first lay out the theoretical underpinning of GANs, then we will see how to build our own GAN using Keras.

Introduction

Let’s start with a short story to illustrate some ...