Question about the G/Loss_cls and D/Loss_cls

[stillbetter]

I am modifing the StarGAN for image quality assessment. The dataset is not that large as CelebA, so I reduce the repeat number of ResBlock and conv layer in G and D. But I found the G/Loss_cls decreased rather quickly than D/Loss_cls. I change the g_lr&d_lr but helps little. Do you have the same question when training?

What's more, I find StarGAN change the image color . Sometimes even turns black to white. Should I add some extra loss fuction to control or just because the dataset I use is not that large, I need to change some parameters?

Thanks!