I am using image to image synthesis and using GAN architecture. For discriminator network I am calculating real and fake loss individually and backpropagating them. Then I am evaluating the sum of the real and fake loss to check the network performance. But after 4/5 epochs the LossD goes to zero which is according to this GitHub - soumith/ganhacks: starter from "How to Train a GAN?" at NIPS2016 is a symptom of early failure detection. As he says to check the norm of the gradients how do I do that?
