search

zudi-lin / rcan-it

Revisiting RCAN: Improved Training for Image Super-Resolution
MIT License
90 stars 15 forks source link

Some issues on MyOwn data #14

Open iamweiweishi opened 1 year ago

iamweiweishi commented 1 year ago

Hi, I have collected about 20000 images. And the HR-resolution is around 600*600. When the training starts, everything looks fine. After about 10000 iterations, the loss become larger and larger. Finally, the loss become nan. Any suggestions on this kind of issue?

Best regards, Thank you.

zudi-lin commented 1 year ago

hello @iamweiweishi, we described in our paper that the training is sensitive to the learning rate (see 4.1 in https://arxiv.org/pdf/2201.11279.pdf). I would suggest you decrease the learning rate by 8x and try the training again.