GAN Training Instability

Hi guys,

While I was training DCGAN, I got a loss curve as follows.

It seems that after a certain number of (~100) iterations, generator loss will increase tremendously.

Any thoughts on why this will happen?

Can you provide more details? How you iterate between G and D? Dataset and its size? LR? Does the discriminator loss converge (from the plot it looks like a straight line)?

GAN is difficult to train, there are a lot of tips and tricks that you can try.