Trainig a conditional GAN - UNet

Hello everyone! Continuing my tests with UNet, I am trying to create a GAN similar to pix2pix. So, I am using a UNet as generator and a discriminator with the structure usual to conditional GANs. While training, I noticed that the losses remain around the same numbers. When I had a small dataset (20 samples) the generator loss would go down to about 0.7 and the results were surpsisingly not too bad. But now that I added more samples, it’s not even doing that.ganA_25_500 (this is with 20 samples and 500 epochs, the right being the predicted image).
40 samples_400 epoch (these are the losses with the new dataset)
I also tried to use a scheduler, in case it is a learning rate problem, but I ’ m not quite sure how to control it. I know my question is quite broad, but if you have any suggestions I would be happy to try them. Also this is a link to my notebook and samples, if you want to have a look.

1 Like