loss of GAN

[viridyu]

No description provided.

[viridyu]

Hi
I have trained cycleGan on my own dataset. During the training, I found that the GAN loss of Generator A cannot be convergent and becomes larger. At the same time, the GAN loss of Generator B, Discriminator A and B all go convergent. Have anyone met this problem before? Thank you.

[junyanz]

It is quite normal. It is fine as long as the losses do not explode. We are using LSGAN and there is no guarantee that the GAN loss will converge. You can get a convergent loss if you are using WGAN(or WGAN-GP).