search

sanghyun-son / EDSR-PyTorch

PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)
MIT License
2.43k stars 669 forks source link

training problem with VDSR model #210

Open cscss opened 5 years ago

cscss commented 5 years ago

Dear author: Hello! I want to compare the difference between these models, but when I train VDSR model with this code. i found that the loss never descent and the PSNR of validation never changed. When I train EDSR with your code, it's fine. I can not find the reason. Can you help me?

cscss commented 5 years ago

ps: I train both model with DIV2K dataset.

Azriel5474 commented 4 years ago

hello,i have a problem in training VDSR too. I train vdsr in this code python main.py --model vdsr --scale 2 --patch_size 96 --block_feats 128 --save vdsr_DIV2K_x2 --data_train DIV2K --data_test DIV2K --pre_train ''

Then an runtimerrror occurs RuntimeError: The size of tensor a (48) must match the size of tensor b (96) at non-singleton dimension 3

I think that error is VDSR have to bicubic LR to the same resolution HR but how i bicubic image or change VDSR model?

DaxingZ commented 2 years ago

Dear author: Hello! I want to compare the difference between these models, but when I train VDSR model with this code. i found that the loss never descent and the PSNR of validation never changed. When I train EDSR with your code, it's fine. I can not find the reason. Can you help me?

Have you found the reason? The same thing happened to me when I used other datasets.

DaxingZ commented 2 years ago

hello,i have a problem in training VDSR too. I train vdsr in this code python main.py --model vdsr --scale 2 --patch_size 96 --block_feats 128 --save vdsr_DIV2K_x2 --data_train DIV2K --data_test DIV2K --pre_train ''

Then an runtimerrror occurs RuntimeError: The size of tensor a (48) must match the size of tensor b (96) at non-singleton dimension 3

I think that error is VDSR have to bicubic LR to the same resolution HR but how i bicubic image or change VDSR model?

You can use Matlab's bicubic operation to downsample HR images and up them to create VDSR datasets.

Zency-Sun commented 1 year ago

@DaxingZ @Azriel5474 ,Have the two seniors solved this problem yet? How was it solved?

Azriel5474 commented 1 year ago

已收到邮件，辛苦，祝好