search

andreas128 / SRFlow

Official SRFlow training code: Super-Resolution using Normalizing Flow in PyTorch
Other
824 stars 112 forks source link

meet NAN when I use realSR dataset to train SrFlow #25

Closed ccJia closed 3 years ago

ccJia commented 3 years ago

At the early beginning, the loss like this image But after 3 k epoch, the loss meet NaN problem ,like this: image I use pretrain model "SRFlow_DF2K_4X.pth" as initial weights. Could you give me some suggestion to solve this problem?

ccJia commented 3 years ago

without pretrain model get the same problem. image