ctmakro / dcgan-keras Goto Github PK

Is it some dcgan problem or just an architectural mistake? Net seems to return images not in [-1; 1], lower bound is lifted up in -0.6, -0.5 or so. So normalization like in this code makes greyed images. I tried to calculate min value and use it, but this makes images look a bit weird and have unrealistic colors sometimes. Any ideas?

I'm using this with another dataset, and I want to run on a larger (64x64) resolution, but it's hard to figure out what I have to change in the code, can you give me a hint of what to change? Thanks.

I was wondering if there is a particular reason why the gradients of the generator and discriminator are computed and applied in two steps instead of using minimize?

See: https://github.com/ctmakro/DCGAN-Keras/blob/master/lets_gan.py#L163

Thanks for posting this code. Any chance you could also post a license?