Generative Adversarial Networks for Synthesizing New Data

In the previous chapter, we focused on recurrent neural networks for modeling sequences. In this chapter, we will explore generative adversarial networks (GANs) and see their application in synthesizing new data samples. GANs are considered to be one of the most important breakthroughs in deep learning, allowing computers to generate new data (such as new images).

In this chapter, we will cover the following topics:

• Introducing generative models for synthesizing new data
• Autoencoders, variational autoencoders, and their relationship to GANs
• Understanding the building blocks of GANs
• Implementing a simple GAN model to generate handwritten digits
• Understanding transposed convolution and batch normalization
• Improving GANs: deep convolutional GANs and GANs using the Wasserstein distance

Introducing generative adversarial networks

Let’s first look at the foundations of GAN models. The overall objective of a GAN is to synthesize new data that has the same distribution as its training dataset. Therefore, GANs, in their original form, are considered to be in the unsupervised learning category of machine learning tasks, since no labeled data is required. It is worth noting, however, that extensions made to the original GAN can lie in both the semi-supervised and supervised domains.

The general GAN concept was first proposed in 2014 by Ian Goodfellow and his colleagues as a method for synthesizing new images using deep neural networks (NNs) (Generative Adversarial Nets, in Advances in Neural Information Processing Systems by I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, pp. 2672-2680, 2014). While the initial GAN architecture proposed in this paper was based on fully connected layers, similar to multilayer...

Implementing a GAN from scratch

In this section, we will cover how to implement and train a GAN model to generate new images such as MNIST digits. Since the training on a normal central processing unit (CPU) may take a long time, in the following subsection, we will cover how to set up the Google Colab environment, which will allow us to run the computations on graphics processing units (GPUs).

Training GAN models on Google Colab

Some of the code examples in this chapter may require extensive computational resources that go beyond a conventional laptop or a workstation without a GPU. If you already have an NVIDIA GPU-enabled computing machine available, with CUDA and cuDNN libraries installed, you can use that to speed up the computations.

However, since many of us do not have access to high-performance computing resources, we will use the Google Colaboratory environment (often referred to as Google Colab), which is a free cloud computing service (available in most countries...

Improving the quality of synthesized images using a convolutional and Wasserstein GAN

In this section, we will implement a DCGAN, which will enable us to improve the performance we saw in the previous GAN example. Additionally, we will briefly talk about an extra key technique, Wasserstein GAN (WGAN).

The techniques that we will cover in this section will include the following:

• Transposed convolution
• Batch normalization (BatchNorm)
• WGAN

The DCGAN was proposed in 2016 by A. Radford, L. Metz, and S. Chintala in their article Unsupervised representation learning with deep convolutional generative adversarial networks, which is freely available at https://arxiv.org/pdf/1511.06434.pdf. In this article, the researchers proposed using convolutional layers for both the generator and discriminator networks. Starting from a random vector, z, the DCGAN first uses a fully connected layer to project z into a new vector with a proper size so that it can be reshaped...

Other GAN applications

In this chapter, we mainly focused on generating examples using GANs and looked at a few tricks and techniques to improve the quality of synthesized outputs. The applications of GANs are expanding rapidly, including in computer vision, machine learning, and even other domains of science and engineering. A nice list of different GAN models and application areas can be found at https://github.com/hindupuravinash/the-gan-zoo.

It is worth mentioning that we covered GANs in an unsupervised fashion; that is, no class label information was used in the models that were covered in this chapter. However, the GAN approach can be generalized to semi-supervised and supervised tasks, as well. For example, the conditional GAN (cGAN) proposed by Mehdi Mirza and Simon Osindero in the paper Conditional Generative Adversarial Nets, 2014 (https://arxiv.org/pdf/1411.1784.pdf) uses the class label information and learns to synthesize new images conditioned on the provided label...

Summary

In this chapter, you first learned about generative models in deep learning and their overall objective: synthesizing new data. We then covered how GAN models use a generator network and a discriminator network, which compete with each other in an adversarial training setting to improve each other. Next, we implemented a simple GAN model using only fully connected layers for both the generator and the discriminator.

We also covered how GAN models can be improved. First, you saw a DCGAN, which uses deep convolutional networks for both the generator and the discriminator. Along the way, you also learned about two new concepts: transposed convolution (for upsampling the spatial dimensionality of feature maps) and BatchNorm (for improving convergence during training).

We then looked at a WGAN, which uses the EM distance to measure the distance between the distributions of real and fake samples. Finally, we talked about the WGAN with GP to maintain the 1-Lipschitz property...