Implementing Improved Training Techniques for GANs - Week 1

The aim of my summer project with mlpack is to add new features and training techniques to the GAN framework so, that it is ready for release. More specificially, I will be working on adding improved training techniques for GANs such as MinibatchDiscrimination and VirtualbatchNormalization. I will also be implementing Conditional GANs and other regularization methods if time remains.

This week I was able to have my pending PRs for GANs merged. I started working on MinibatchDiscrimination which we have decided to implement as a layer in the mlpack API. The major difficulty in implementing a layer is always in deriving the gradient and implementing it successfully however after some debugging I was able to have the numerical graident test to pass. I believe the major coding part of the layer is complete and more tests and optimizations should follow in the next week.

I also worked on Inception Score function that we require to test the performance of the layer. Hopefully testing in the coming week won't be too troublesome and we would be able to merge both.