liyunsheng13 / BDL

MIT License
222 stars 30 forks source link

CycleGAN training problem #6

Closed liyunsheng13 closed 3 years ago

liyunsheng13 commented 5 years ago

hi,When I train cyclegan, GA can easily turn the sky in GTAV into a tree. Do you know what the problem is?thank you @liyunsheng13

Originally posted by @PJ1yang in https://github.com/liyunsheng13/BDL/issues/5#issuecomment-499867201

liyunsheng13 commented 5 years ago

You are right. This problem is hard to be solved. I think it is caused by the difference of the spatial layout between GTA5 and Cityscapes dataset. For GTA5, there is not a lot of trees in the sky part but the Cityscapes is different. When CycleGAN is trained, it realize that spatial layout difference and tries to convert it from source domain to the target domain. If you train CycleGAN with datasets that have larger domain gap, you can find this phenomenon more obvious. Right now, there is not a way to solve this problem. I suggest you to make the identity loss or the reconstruction perception loss more important. They might help.

ghost commented 5 years ago

Thanks for your advice