Customizing Training Procedure

[romankouz]

Hi Author!

This was a super interesting work and I wanted to explore the architecture of this model a bit. However, I'm getting a little confused about the specifics of the codebase. I can't seem to figure out:

1. Where are the recognition and generator networks are trained?
2. If I wanted to slightly change any of the recognition/generator networks, where can I do that?
3. Why is the training time so high? I can't seem to fully dissect what computations are being done and I'd like to try a few approaches that maybe don't perform data augmentation.

Sorry if these questions are somewhat novice; I just couldn't figure this out entirely by myself :(