Explore GANs by building one that can generate high quality image of deep space (akin to images taken by the Hubble Telescope).
The data consists of images sampled from the larger Hubble eXtreme Deep Field (XDF).
I was able to replicate the results of the Pro-GAN architecture described in this paper by Nvidia. With this I was able to generate relatively high quality images resembling the Hubble eXtreme deep field.
The main idea behind the Pro-GAN architecture is to progressively grow the size of the GAN incrementally, in order to produce faster and more stable training results.
Due to hardware constraints, I modified the architecture described in the paper by scaling it down from having a maximum of 512 channels per convolution operator to a maximum of 256 channels. Despite this scaling down, the results appear to be fairly good.
A number of tricks that are mentioned in the paper were employed including an equalized learning rate, a minibatch standard deviation layer, and pixelwise normalization. All of these tricks helped to stablize the training and reduced the chances of encountering mode collapse significantly.