EDSR-PyTorch About test Set5 have some error

Hey, I download your predtrain model ,EDSRX2 to test Set5!. But have some pixel error

And the PSNR of Set5 is 39.14. It seem big than your paper?

May 08 '19 01:05 xiaobiaodu

Hello.

Would you provide the test script you used? I think I can give you an answer with #164

Thank you.

May 08 '19 19:05 sanghyun-son

code.zip

May 09 '19 01:05 xiaobiaodu

Hi.

I have send my code. If u want to run this code. U just modify the test.py line80 for your model path, and data_fortest.py line17 for your benchmark path(Data folder structure is same with yous).

I will appreciate you for your help! Thank you!

May 09 '19 01:05 xiaobiaodu

@echosum Sorry but for some personal reason, I cannot run your code right now.

I checked the code and thought that line 45, 47 and 48 of your ``test.py` may cause the problem.

This problem usually happens when the output image is not clipped well.

In other words, you need to check that pixel values of the quantized image are in [0, 255].

If not, you can simply resolve this problem by clipping them.

Please let me know if you cannot fix the problem after the solution :)

Thank you

May 09 '19 19:05 sanghyun-son

` hr = model(lr)

            hr = common.quantize(hr,255)

            print('range',torch.max(hr),torch.min(hr))

            hr_np = hr.cpu().data.numpy().astype('uint8').transpose(0, 2, 3, 1)

            print(np.max(hr_np),np.min(hr_np))

            img = Image.fromarray(hr_np[0],'RGB')

            img.save('results/%s'%batch['info'][0])`

May 09 '19 23:05 xiaobiaodu

Hi. In the code that I send to you. I have fix the problem that some piexl is error.

But the PSNR is always different with the result in your paper.

I have check the range of HR image is [0,255].

The PSNR is always different with the result in your paper.

I always can't fix the problem by myself. I will appreciate you.

Thanks for your help.

May 09 '19 23:05 xiaobiaodu

Hi I try many ways to find the problem that PSNR is wrong. I think maybe the key in the input of model.

May 10 '19 15:05 xiaobiaodu

Hi， I have fix these problem. Thanks for your help

May 11 '19 02:05 xiaobiaodu

@echosum hi, I downloaded the edsr_x4.pth and test it on the validation set of div2k and I got a pretty low psnr(19.7) and ssim(0.459). Then I test on set5 and the psnr still has a very low value(19.2). What could I miss?