GAN Loss and Discriminator Loss are abnormal. #383
Description
I tried to reproduce VAE on ImageNet, but during the training process, I found that the discriminator loss kept increasing. When the discriminator is added to the training, the overall loss continues to decrease and becomes a very small negative number (around -1e6).
What could be the possible reasons? All hyper-parameters are set according to the config provided.
The following is the training curve showing the changes in all losses, including disc_loss, gan_loss, logit_fake, and logit_real. Is this phenomenon normal？
