Struggle with PyTorch version code

[onwn]

I'm afraid I might bother you, but it seems that the PyTorch implementation cannot reproduce the expected result. When I tried to do so, I met following issues......:

1. When I run train/test code(without touching any setting),

• DnCNN-S for noise level=25 made PSNR=29.24 in <30 epoch
• DnCNN-S for noise level=15 got stuck at PSNR=23.xx. I tried this many times but didn't work.
• DnCNN-S for noise level=50 made 25.9x in <50 epoch, but couldn't reach near 26.23 at all. There's multi_step scheduling, however I forced it to use smaller learning rate after the convergence, and it didn't help. +) I tried some modifications on the model, and still the problem remained. Just <0.8dB changed. So think it is not about model. +) I also tried to change the optimizer and lr scheduling exactly same as it is mentioned on the original paper, but that made it even worse.

2. Is it right to add Gaussian noise and don't clip it? I thought it must be clipped according to [0., 255.] (uint8) or [0., 1.] (float32) because in real case(of course it is not completely 'real' 'cause it's AWGN) the corrupted image would be in [0., 255.] range. How much does the clipping lowers its performance, or is it better? I wanted to test it myself but... like I mentioned in 1., the baseline model isn't working correctly, so... If you have already tried it, could you let me know?

[cszn]

1. Please make sure you have changed the noise level in the training and testing for noise level 15 and 50.

2. You may refer to FFDNet. I added a discussion on Un-Clipping vs. Clipping of Noisy Images for Training in Section III. G. I also released the models for the two settings. Their PSNRs are almost the same.

[onwn]

Could you please provide the pretrained models for noise level=15 and level=50? Actually I'm sure that I set the noise level and other settings properly, so if the problem still happens even when I use your pretrained model then I can check that the trouble is from elsewhere.

[cszn]

@onwn https://github.com/Liyong8490/DnCNN-pytorch

[onwn]

It seems working fine and thanks. closing it.

[CodeR57]

@onwn Can you please help me with reproducing the mentioned PSNR scores. I ran the Main_test.py, but the PSNR scores are much lower(~1dB). PFA