anoff opened this issue 7 years ago · 0 comments
seems like the network is learning very complex constraints with individual pixels, try to learn very simple shapes with a simple network?!