Open JKANG94 opened 3 years ago
The random color shift algorithm also is not the same as the paper described.
@JKANG94 I am quite sure that the random color shift is the same as the paper described, it was implemented in another format, but these are mathmatically equaivalent. For adaptive coloring, as I have mentioned, it was not stable and easily causes pixel overflow. Still, you can try to use it by using pixel normalization and adding a tanh layer in the generator. I believe these have been implemented though unused.
clipping and tanh may help, but they don't solve the issue at all. You can make it solid though, you just have to plug in a couple work arounds to the math functions that will break due to inputs.
In train.py, I found that the structure loss is not built with adaptive coloring algorithm. It is same as the content loss. So why didn't use adaptive coloring algorithm?