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

Bad results on Set14 and Urban100 but normal on Set5 and B100 when testing #253

Open fengshenfeilian opened 4 years ago

fengshenfeilian commented 4 years ago

Thanks for your great job. I feel confused about the test results when train RCAN model. I use "CUDA_VISIBLE_DEVICES=1 python main.py --model RCAN --save RCAN_G10R20_X2 --scale 2 --save_results --epochs 1000 --reset --n_resgroups 10 --n_resblocks 20 --patch_size 96 --chop False --decay '200-400-600-800'" to train RCAN, and find that PSNR of Set5 and B100 is similar to the paper but Set14 and Urban100 has a large margin about 0.4. I have no idea about why this happened. Could you please help to analyze possible problems. Thanks a lot!

Senwang98 commented 3 years ago

Have you solved this problem?

Senwang98 commented 3 years ago

Which val set do you use?