h3lio5 / gan-pytorch Goto Github PK

Hi,

Thanks much for creating this repo, it's really useful. I'm curious about this quoted line below, why does D(G(z)) that is close to 0 mean that generator successfully fools the Discriminator? Is it that what we want is maximizing D(G(z)) to be as close as 1, i.e., generator successfully generates fake image (label =0) and the discriminator predicts it as real image (label=1, by having D(G(z)) close to 1)?

Hi,

As the title says, replacing k=4 by k=1 and keeping the rest of the code untouched allows the quality of the generated handwritten digits to be two or three times better.

By the way it seems that k=1 is the value used in the original paper experiments :

The number of steps to apply to the discriminator, k, is a hyperparameter. We used k = 1, the least expensive option, in our experiments.