search

jspan / PHYSICS_SR

This repository is an official PyTorch implementation of the paper "Image Formation Model Guided Deep Image Super-Resolution" from AAAI 2020.
28 stars 1 forks source link

about the results #4

Open hahaha-1 opened 3 years ago

hahaha-1 commented 3 years ago

hi, i run the code that you released，but the result not achieve the results that you published in the paper. Such as under scale factor is x4, the psnr is 32.14 and the ssim is 0.8942 on Set5 datasets in my result. I just only change the path of the dataset. can you give some advice?

jspan commented 3 years ago

@hahaha-1 Hello I tested the Set5 dataset again and everything went well. I guess your results may be caused by the wrong testset. The Set5 dataset was generated by Matlab imresize function. please check your Set5 datasets. I hope this can help you.

hahaha-1 commented 3 years ago

I tested the Set5 dataset with the pre-trained model that you released, the performance indeed achieves the results that you published in the paper. But, i tested the same Set5 datasets with the model that i trained. The result does not achieve the desired results. Do you train your model with the released parameters? Do i need change some parameters, such as epochs?

jspan commented 3 years ago

@hahaha-1 I'm not sure why it happened. I don't know the settings when you trained this code. In addition, different network initial settings will lead to different results. .We use EDSR pre-train model to initial our network. So, the results of the first stage can't be lower than EDSR results. please check the settings again.

hahaha-1 commented 3 years ago

ok, thank you again for your contribution.