Implementing Essential Deep Learning Modules - Week 3

We're into the final week of Phase I and, thanks to Marcus' suggestions and reviews, right on track! A number of key decisions were taken up by us this week.

The GAN code is finally ready to get merged, and the DCGAN code has been pushed, which should be ready to get merged by the end of this week. As discussed with Marcus, the DCGAN code uses the same class as the standard GAN, which allows us the freedom to just implement the test cases, as the required layers are already available in mlpack.

This week we got some good results with the GAN code on a 500-image subset of MNIST dataset, and I'm currently in process of running the code on the full 70,000-image MNIST dataset, which would be posted in the PR itself. Also, we decided to not run the test on Travis, and instead, push the code to the mlpack/models repository, whch would be done soon.

Also, we decided to try out optimizer separation and implementing batch support for the GAN implementation after Phase I, which would allow us flexibility to compare and contrast our single optimizer output as well, and provide full user flexibility to try out different variants of the GAN infrastructure.

That's all for now! See you next week!