search

jorge-pessoa / pytorch-msssim

PyTorch differentiable Multi-Scale Structural Similarity (MS-SSIM) loss
Other
453 stars 70 forks source link

nan problem #11

Closed YuhuiMa closed 4 years ago

YuhuiMa commented 4 years ago

How to deal with the problem of nan loss during training?

chledowski commented 4 years ago

Use normalize=True

jorge-pessoa commented 4 years ago

Hello,

As @chledowski suggested you can test normalize=True at least for the initial iterations of the optimization process. I have pushed some changes to the code which should help significantly in the stability of the optimization in the initial iterations.

I would ask you to test those changes and re-open an issue if you are still facing problems with the optimization. Thank you for your feedback.