GAN — GAN Series (from the beginning to the end)

Summary

The website provides a comprehensive series on Generative Adversarial Networks (GANs), covering their applications, issues, and solutions, as well as showcasing various GAN architectures and improvements.

Abstract

The website content offers an in-depth exploration of GANs, beginning with an overview of their applications and moving on to discuss common issues encountered during training, such as mode collapse and training instability. It also presents a variety of GAN architectures, including DCGAN, CycleGAN, and StyleGAN, and delves into techniques for improving GAN performance, such as spectral normalization and progressive growing. The series examines different cost functions and their impact on GAN training, as well as optimization strategies to mitigate problems like mode collapse. Each article in the series is designed to provide insights into the complexities of GANs and to guide readers through the advancements and nuances of this deep learning domain.

Opinions

The series suggests that training GANs is inherently challenging due to issues like mode collapse and training instability.
It posits that the choice of cost function significantly affects GAN performance, with alternatives like LSGAN, WGAN, and EBGAN being explored for better results.
The articles convey that architectural improvements, such as those seen in DCGAN, SAGAN, and StyleGAN2, are crucial for enhancing the quality of generated images.
The use of labels and additional information in models like CGAN and InfoGAN is presented as a beneficial approach to improving GAN outputs.
The series implies that optimization techniques, including Unrolled GAN and DRAGAN, are essential tools for addressing mode collapse and improving training dynamics.
It is suggested that the development of GANs is an iterative process, with each new architecture or technique building upon the lessons learned from previous models.
The content indicates a consensus that despite the challenges, GANs have a wide range of cool applications, demonstrating their practical utility in various domains.

GAN — GAN Series (from the beginning to the end)

GAN — A comprehensive review into the gangsters of GANs (Part 1)

Are we there yet? In this GAN series, we examine how GAN is applied to deep learning problems and look into its…

GAN — A comprehensive review into the gangsters of GANs (Part 2)

This article studies the motivation and the direction of the GAN research in improving GANs. By reviewing them in a…

Basic

GAN — What’s GAN Generative Adversary Networks?

To create something from nothing is one of the greatest feelings, … It’s heaven.

GAN — Some cool applications of GANs.

In this section, we will sample some of the GAN’s applications and display some of their results.

GAN — How to measure GAN performance?

Inception Score (IS)

GAN — CycleGAN

Let’s say we collect three set of pictures: one for real pictures, one for Monet paintings and the last one for Van…

GAN — Super Resolution GAN (SRGAN)

Super-resolution GAN applies a deep network in combination with an adversary network to produce higher resolution…

GAN Issues

GAN — Why it is so hard to train Generative Advisory Networks!

It is easier to recognize a Monet’s painting than drawing one. Generative models (creating data) are considered much…

General GAN Improvements

GAN — Ways to improve GAN performance

GAN models can suffer badly in the following areas comparing to other deep networks.

Improving the Network design

GAN — DCGAN (deep convolutional generative adversarial networks)

Replace all maxpooling with convolutional stride.

GAN — CGAN & InfoGAN (using labels to improve GAN)

In discriminative models, like classification, we may handcraft features to make models to perform better. This…

GAN — Stacked Generative Adversarial Networks (SGAN)

Stacked Generative Adversarial Networks (SGAN) composes of

GAN — Progressive growing of GANs

The progressive growing of GANs trains the GAN network in multiple phases. In the first phase, it takes in a latent…

GAN — Self-Attention Generative Adversarial Networks (SAGAN)

Motivation

GAN — Spectral Normalization

GAN is vulnerable to mode collapse and training instability. In this article, we look into WGAN and Spectral…

GAN — StyleGAN & StyleGAN2

Do you know your style? Most GAN models don’t. In the vanilla GAN, we generate an image from a latent factor z.

Improving the cost function

GAN — LSGAN (How to be a good helper?)

The GAN discriminator distinguishes real images from the generated one. It usually performs reasonably well. But as a…

GAN — Wasserstein GAN & WGAN-GP

Training GAN is known to be hard. We regularly deal with the performance, model stability, model convergence and mode…

GAN — Energy based GAN (EBGAN) & Boundary Equilibrium GAN (BEGAN)

GAN is a minimax game in which opponents take action to undercut the other. This makes GAN much harder to converge, if…

GAN — RSGAN & RaGAN (A new generation of cost function.)

How can we train GAN faster with better image quality? We have many new cost functions already. But Relativistic GAN…

GAN — What is wrong with the GAN cost function?

We work hard to produce mathematical models for deep learning. But often, we are not successful and fall back to the…

GAN- Does LSGAN, WGAN, WGAN-GP or BEGAN matter?

A 2017 Google Brain paper “Are GANs Created Equal?” claims that

Proof (GAN optimal point)

We have mentioned optimizing GAN is optimizing JS-divergence. This is not obvious from the cost function:

GAN — DRAGAN

Mode collapse is one major concern in training GAN. In the figure below, we use an BEGAN to generate faces. For each…

Optimization

GAN — Unrolled GAN (how to reduce mode collapse)

Intuition