sanghyun-son / EDSR-PyTorch

PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)
MIT License
2.44k stars 671 forks source link

About the PSNR and L1 loss value #224

Open momoda-mht opened 5 years ago

momoda-mht commented 5 years ago

Hi, when I run some models in the code, including but not limited to RCAN, the PSNR value seems to be lower and lower after 25 epochs, and the L1 loss vary a lot between different epochs as the below picture shown, is it normal? PSNR train

AbdulMoqeet commented 4 years ago

Set5 contains only 5 images, but in your case, it shows 40 images. Please make sure, you have correctly, set-up the training and test data.

momoda-mht commented 4 years ago

thank you for your reply. My Set5 contain 40 images because I add some other images, and the cause for getting lower PSNR value may be the way I add to the Set5 is not correct. But I still have the problem that the L1 loss vary a lot between epochs. By the way, when I use the model 'rcan.py' in the 'model' profile for training, after several epochs, the L1 loss suddenly increase, and the PSNR value drop a lot, is there any bug in data loading? QQ图片20200106162206

LY54 commented 4 years ago

@momoda-mht Hello, I have the same problem. Have you solved it? What's your suggestion

Senwang98 commented 3 years ago

I guess you use RCAN pre-trained model? It seems to use lower learning rate to load pre-trained RCAN.pt model.