This repository contains my first code in PyTorch: a GAN implemented from scratch (well, not really) and trained to generate MNIST like digits. Minibatch discrimination, was also implemented to avoid mode-collapse, which is a common phenomenon observed in trained GANS. (Link to paper: https://arxiv.org/abs/1606.03498).
Here is comparison of the generated outputs of GAN, with and without any minibatch discrimination. The networks employed were very simple (only one hidden layer), trained for 20 epochs each with a batch size of 20 and learning rate = 1e-4. Note that the code for minibatch discrimination is not really optimal at this stage becuase of a for-loop. I am yet to find a way to fix that issue using PyTorch.
(Left) Vanilla GAN outputs (Right) GAN outputs using Minibatch Discrimination Layer
Clearly Minibatch discrimination has improved the output digits, but they still don't look realistic enough. Possible ways we can improve this:
(a) Adding more layers. As of now the there is only a single hidden layer which is too simplistic in my opinion.
(b) Some other improved techniques for training GANs such Wesserstein/Unrolled GANS.