I’ve implemented a minimal WGAN-GP on MNIST (code here).it kinda works and outputs some digits but loss/gradient magnitudes are so huge, for example:
-
GradientPenalty \approx 10^{19}
-
D(\hat{x})- D(x) \approx -10^{10}
-
-D(\hat{x}) \approx 10^{10}
There is definitely something wrong with these numbers, but I’ve got no clues, have you any ideas?