I’m looking for advice with my CycleGAN. I’m finding that if I update the discriminator loss every batch then my generator’s loss is terrible and my discriminator loss is 0. Naturally I’ve tried variations in how frequently I update the discriminator which yielded much better results. However the model still isn’t good enough so I’ve been playing with various parameters with varying success. Out of curiosity, I just tried never updating the discriminator which has led to my best generator loss yet. This surprised me as I would have thought the generator isn’t getting any useful feedback from the discriminator so wouldn’t improve. The former also led to decent identity loss, while the latter was fairly average.

I’m starting to wonder if the discriminator is doing anything at all or if I’ve basically created an autoencoder? Any debugging tips would be massively appreciated. For what it’s worth, eyeballing my results they’re ok. They’re not at a satisfactory level yet, but I can see the model is learning something. I’d really appreciate any tips