I had some issues in the past implementing CycleGAN (see my repo on document denoising. After spending more time with Pytorch, I think I understood some of the mistakes I made previously and I hope to fix those issues in this implementation.
Hopefully I will be able to do that, but I'd like to get something working and off the ground first (using hard coded sizes in this implementation to simplify things and focus on the CycleGAN training loop rather than the network architectures).