How to backpropagate combined loss over generator and only classification loss over discriminator?

autograd
2

You are likely running into this issue causing the second backward call to fail as the parameters were updated resulting in stale forward activations.