CycleGAN training

I have a tricky problem training cycleGAN. If I do a small number of epochs it progresses well, but the result is not good enough. If I train for a larger number of epochs, I find that discriminator B gets stuck and doesn’t improve, while everything else continues to improve. I’m assuming this means that generator B is tricking discriminator B too easily. Correct? But I’m not sure what strategies would make the most sense to help discriminator B. Do I perhaps just need more data?
I’m doing pix2pix—something along the lines of sketches to cats—and using paired data.

Any thoughts appreciated.

1 Like

Yes D is finding it too easy. First steps would be to use (more) label smoothing (fake/noisy labels) and reduce D lr. Play with bs. Change weight initialisation, etc. Gans are hard to train and regularise, hyper params matter a lot.

Maybe take a look at the just finished Kaggle GAN challenge (noise to image) which now probably has as many Gan tips as anywhere else on the internet.

1 Like