About the resize problem of real20

[Rachel-kk]

Hi，I've resized the short edge of images in real20 dataset to 512(used torchvision.transforms.Resize).. I tested your released model but it got psnr of 21.89, ssim of 0.804.

I don't understand why the psnr I tested is not the same as your paper(the psnr of 22.89, the ssim of 0.803 in your paper) , I hope to get your answer.

[Vandermode]

Hi, I just uploaded my testing data into GoogleDrive from the local computer. Could you please check whether you can reproduce our results

[Rachel-kk]

Hi, @Vandermode Thank you for providing your testing data. I downloaded it and retested on your released model, the result is as follows:

	Real(20)	Postcard(179)	Wild(52)	Soilobject(200)
your paper	22.89 / 0.803	22.04 / 0.876	24.25 / 0.853	24.87 / 0.896
retest results	22.91 / 0.806	22.07 / 0.877	24.21 / 0.857	24.85 / 0.898

I observed that the psnr error is within 0.03, and it's right. Next, I will retrain your released model and reproduce your results.

[Rachel-kk]

Hi, I run your code on GitHub. However, I didn't reproduce the same performance as the paper. For VOC2012 dataset, following the way you said, I crop the center region with size 224 x 224. Is there anywhere else that would affect performance?

[Vandermode]

Could you please tell me more information about your environment? for example, the PyTorch version, the GPU version, etc..

[Vandermode]

Note this work was done many years ago, and actually, I tried to reproduce its results several months ago, but I found a very weird issue---the reproduction was successful in old GPU architectures (e.g., 1080Ti, V100), but often failed in SoTA GPUs (e.g., 3090Ti, A100)... I don't figure out how this could be possible, maybe you encountered the same issue

[Rachel-kk]

Hi, I train on two environments separately, as follow. the Pytorch version is 1.7.1+cuda10.1; the GPU version is TITAN V. the Pytorch version is 1.10.1+cuda11.1; the GPU version is RTX 2080