the code for computing PSNR is wrong

[Delicious-Bitter-Melon]

Your code computing PSNR as follows is wrong:

DisCo/tool/metrics/ssim_l1_lpips_psnr.py

Line 13 in 8538889

mse = np.mean((original - compressed) ** 2)

.

The accurate code is mse = np.mean((original/1.0 - compressed/1.0) ** 2), because original and compressed images are uint8 in your code, and (original - compressed) * * 2 will cause numerical overflow.

[Delicious-Bitter-Melon]

Please update your results in the paper.

[Wangt-CN]

Could you please check the formal implementatoin here: https://github.com/bonlime/pytorch-tools/blob/master/pytorch_tools/metrics/psnr.py#L11
The input range is [0,255]

[Delicious-Bitter-Melon]

Could you please check the formal implementatoin here: https://github.com/bonlime/pytorch-tools/blob/master/pytorch_tools/metrics/psnr.py#L11 The input range is [0,255]

The author claims in the issues that they have not used and even tested PSNR and SSIM.

[Delicious-Bitter-Melon]

https://github.com/bonlime/pytorch-tools/blob/01f40fe75c2a11683c247d8fcf3d3645a5783441/tests/metrics/test_metric.py#L105

And their images are float32, but in your codes images are uint8.