Cannot get the objective quality（PSNR/SSIM）

[Berlin0610]

Hi, Longguang,

Thank your for providing the resource code and the pre-trained model. It is a wonderful and elegant work. However, when I use the pre-trained model or the retrained model according to your training code to test the Set5 dataset, there are some problems as follows:

1. Compared with the bucubic method, the pre-trained or retrained model cannot get good the objective quality in terms of PSNR and SSIM. It is X2 super-resolution. The related data is attached. I don't know why this result is so strange.

2. It should be mentioned that the subjective quality of the super-resolution image from pretrained or retrained model is obviously better than that from bicubic.

3. In addition, when I have this problem, I tested these model in some other images( not public). And the related results are similar as above mentioned. The traditional quality measures cannot have good performance, while the subjective quality is better than bicubic.

As a result, I am confused why this happened. Sincerely hope that I can get your help at your convenience.

Regards,
Bolin

[hahazh]

hi, I think I got the reason!
UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
you may specify align_corners=True .
And if you use a custom dataset, you need to recalculate RGB_mean.

[Alethiea-Cheng]

hi, I think I got the reason! UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details. you may specify align_corners=True . And if you use a custom dataset, you need to recalculate RGB_mean.

cool, that is a really detail I might digged into so long.