KAIR
KAIR copied to clipboard
Problem with SwinIR Urban100 & Manga109 replication
As I replicate SwinIR SR x3&x4 on DIV2K dataset, I encountered a drop on U100&M109 testsets compared to results reported in the paper:
All training are done with the original DIV2K dataset, without lmdb or patch preprocessing.
SR x3
Only the following configs in options/swinir/train_swinir_sr_classical.json are changed:
opt['scale']=3
opt['datasets']['train']['H_size']=144
opt['netG']['upscale']=3
Test script parsers are:
python main_test_swinir.py --task classical_sr --scale 3 --training_patch_size 48
The results are:
SR x3 (PSNR) | Set5 | Set14 | B100 | U100 | M109 |
---|---|---|---|---|---|
Paper | 34.89 | 30.77 | 29.37 | 29.29 | 34.74 |
Replication | 34.89 | 30.75 | 29.35 | 29.22 | 34.66 |
SRx4
Only the following configs in options/swinir/train_swinir_sr_classical.json are changed:
opt['scale']=4
opt['datasets']['train']['H_size']=192
opt['netG']['upscale']=4
Test script parsers are:
python main_test_swinir.py --task classical_sr --scale 4 --training_patch_size 48
The results are:
SR x4 (PSNR) | Set5 | Set14 | B100 | U100 | M109 |
---|---|---|---|---|---|
Paper | 32.72 | 28.94 | 27.83 | 27.07 | 31.67 |
Replication | 32.74 | 28.98 | 27.82 | 26.94 | 31.49 |
This is even more strange when I could successfully replicate DF2K experiments with almost the same configs (but in training patch size 64). What might be the problem? Thank you very much! Also, thanks for the open-sourcing of this wonderful repo!