passerer / hpinet Goto Github PK

Appreciate the extraordinary work of coding, I encountered some issues might need your help.

I used the HPINet-M-x4.pth provided in the checkpoints folder, the PSNR result is only 31.23 db on Set5 for x4 Super Resolution with model M, while the paper claims it should be around 32.60 db of PSNR.

The result on other test datasets is listed as follows:
Set5, Mean PSNR: 31.23, SSIM: 0.8878
Set14, Mean PSNR: 28.19, SSIM: 0.7855
B100, Mean PSNR: 27.35, SSIM: 0.7435
Urban100, Mean PSNR: 26.10, SSIM: 0.7986

Here is my conda environment setup:
python        3.6.13        h12debd9_1
pytorch        1.8.0        py3.6_cuda10.2_cudnn7.6.5_0        pytorch
torchvision   0.9.0       py36_cu102       pytorch
einops             0.3.2        pypi_0       pypi
scikit-image   0.17.2        pypi_0       pypi

Which should be exactly the recommended settings in the readme part of the repository
Executed with command python test.py --model M --scale 4

.png extension is adopted on Set5 instead of .bmp which might cause some minor issues, but the performance gap to what stated on the paper is still not neglectable on other datasets.

Am I missing something?

Thanks in advance!

When i was trying to train with multiple GPUs due to the limited memory, the loss was not stable and can not get the ideal results. But when i using 1 GPU to train, the loss is less than 0.1 and the training process is stable. Have you met this situation ?

Hello author, I really like your work. I would like to ask if I can’t provide the method and source code of the LAM that implements the model in the paper, sorry to bother you, thank you