Discriminator Training Improvement

[dthul]

Hi Conor,

I trained your GAN model several times and noticed that it would always start to converge but then it would diverge and produce completely broken images. Sometimes it was able to catch itself again, sometimes not. But it always happened in all of my test runs.

In the current code the discriminator training works as follows: one batch of real data followed by one batch of fake data, then switch to training the generator. If you instead mix up the batches of real and fake data such that each of the two batches consists of 50% real and 50% fake, training seems to become much more stable. It never diverged on me yet after the change.

I see that the current discriminator training protocol (training on one batch of only real data followed by one batch of only fake data) is also used for the other models your code implements, not only for the GAN. I didn't try whether mixing the batches also has a comparable impact on the other models.