Open rakesh-reddy95 opened 2 months ago
Sorry, it is unnecessary. I forgot to delete it.
But I see it's been used in multiple places in image_restoration2_model.py
image_restoration2_model.py is just useful for Learning Degradation Representations for Image Deblurring (another paper).
Okay after fixing I see this grid over the output. And is random everytime I generate the output.
I have commented out state_dict_pth_classifier and loading 'ckpts/RevD-B/net_g_GoPro.pth'
It may be the problem like https://github.com/megvii-research/NAFNet/issues/25 , could you test the model trained on RealBlur? The images from GoPro dataset are almost token by day.
This is indeed giving complete grid by running using RealBlur model.
Alright, I will try to provide a version that cancel the TLSC in the encoder tomorrow.
Thank you.
Hi, PatchAdaRevIDSlideV2 in basicsr.arch.AdaRevID_arch is the version cancel the TLSC, while I am not sure it can solve the problem.
I see that the issue is with the model. I have used RevD-L then the issue is not there.
Thank you very much. I will check the model when I am free.
Hello, Thanks for your work! I have used PatchAdaRevIDSlideV2 with weights of RevD-S, and still some results are broken in Gopro dataset. This scene have much more broken images.
Sorry. I see the results are not broken in my public results 百度网盘(提取码:27ex). Probably something went wrong while I was sorting through the code. I will check it next week.
Hi @Brandon-git-hub @rakesh-reddy95 , I found that I accidentally deleted AdaRevIDSlideV2 for testing in basicsr.arch.AdaRevID_arch. I have re-uploaded it. The broken results on GoPro do not appear when I test with AdaRevIDSlideV2. Maybe it is the reason cause the grid over your image @rakesh-reddy95 .
@INVOKERer Thanks for your help! I think my issue might be related to the classifier. I noticed that even I didn't load the classifier weights into model, the classifier is still being used (each time I run, the number of decoders is different). How can I run RevD-L without the adaptive function?
You can turn the hyper-param pretrain from false to true for RevD.
I run validation.py on the Gopro dataset and get a PSNR of 34.656160. This is better than the result reported in the paper. I'm not sure why, but it's a good result!
When I tested on AdaRevD-B, I also encountered the above problem. How should I solve it?And The results show that the number of num_decoders is 0.
Hi @Brandon-git-hub @rakesh-reddy95 @luciferincc , the reason for this error was found to be that I set up the AdaRevD-L network by default in yml, and the model in val.py is the pth of AdaRevD-B. This issue has been fixed.
Hi @Brandon-git-hub @rakesh-reddy95 @luciferincc , the reason for this error was found to be that I set up the AdaRevD-L network by default in yml, and the model in val.py is the pth of AdaRevD-B. This issue has been fixed.
Thanks, I have downloaded the new yml file and it works.
Hi @INVOKERer
I don't see the initialisation for this function. from basicsr.models.archs.discriminator import MultiscaleDiscriminator