Towards Accessible Improved Generative Adversarial Networks

The first goal of this project is to implement GANs as desribed first in:

Generative Adversarial Networks, I. J. Goodfellow et al. , Jun 2014.

GANs were vastly improved since then, with architectural changes, or changing the loss function and our goal is to provide detailed notebooks about those main improvements. The notebooks themselves won't go too much in detail about the theory or our results, which is why we also wrote a guide summarizing everything, with first some theory and them some experiments, all in one place.

We also wanted to show some issued that can occur when training GANs and how to solve those common problems, such as mode collapse for instance.

Here is a list of present notebooks and what you will find in them:

Notebook name	Content
GAN	Implementation of the original GAN paper
Conditional GAN	Implementation of conditional GAN
Strong X	Example of having either a strong generator or a strong discriminator
DCGAN	Implementation of DCGAN
Conditional DCGAN	Implementation of conditional DCGAN
Kernel size of NxN	Use of various kernel sizes
Minibatch discrimination	Implementation of minibatch discrimination
Experience replay	Implementation of experience replay
Noisy labels	Implementation of noisy labels
Virtual Batch Normalization	Implementation of Virtual Batch Normalization
WGAN with gradient penalty	Implementation of Wasserstein GAN with gradient penalty

Technology used

For this project, we used the following libraries :

• Tensorflow 2