search

dvlab-research / MASA-SR

MASA-SR: Matching Acceleration and Spatial Adaptation for Reference-Based Image Super-Resolution (CVPR2021)
161 stars 20 forks source link

PSNR of CUFED ? #8

Open Lkinyuu opened 3 years ago

Lkinyuu commented 3 years ago

train as python train.py --use_tb_logger --data_augmentation --max_iter 250 --loss_l1 --name train_masa_rec

but the best PSNR of CUFED only 27.31

Lkinyuu commented 3 years ago

but the PSNR/SSIM on paper is 27.54/0.814 although i havn't test the model, the psnr in test phase should be lower than train?

SkyeLu commented 3 years ago

Hi, did you stitch 4 references to one image while testing? If not, please refer to this issue: #2 You should try the following command: python test.py --resume './pretrained_weights/masa_rec.pth' --testset TestSet_multi --name masa_TestSet_multi

Lkinyuu commented 3 years ago

Yeah,I have seen this issue. Actually,I haven’t done the test steps yet.But,as I said, the best PSNR I got in the eval phase during train is only 27.31,even lower than the PSNR in test on paper. In my cognition, the test PSNR should be lower than the eval PSNR.

SkyeLu commented 3 years ago

In the eval phase during training, we do not use four reference images, only one reference image is used for fast evaluation.

wdmwhh commented 2 years ago

Hi, I also try to reproduce the result. The best PSNR of CUFED is only 27.294 (for L1 and 27.466 for all references) in the 250 epochs.

SkyeLu commented 2 years ago

Hi, I also try to reproduce the result. The best PSNR of CUFED is only 27.294 (for L1 and 27.466 for all references) in the 250 epochs.

Hi, Different random seeds and experiment environments might lead to a slight difference between the replicated result and our reported result. Checkpoints from different epochs could also lead to different results (0.0x dB difference on PSNR is acceptable). You may try to run the code for multiple times using different random seeds, and also test with checkpoints from different epochs to get better results.