search

greatlog / DAN

This is an official implementation of Unfolding the Alternating Optimization for Blind Super Resolution
231 stars 41 forks source link

About pretrained weights #20

Closed MeLonJ10 closed 3 years ago

MeLonJ10 commented 3 years ago

Hi, it' a wonderful job. Why two pretrained weights ("danv1_x4_setting1.pth" and "danv1_x4_setting2.pth") are the same?

greatlog commented 3 years ago

Thanks for your attention.

Actually, they are the same model. The model is trained under Setting 1. The result we reported for Setting 2 (x4) is also produced by this model. It is a mistake. If we had trained the model under Setting 2, it perhaps would achieve better results. But we have made a mistake in reporting the results. So we provide the model that we used for producing the reported results. The results will be corrected in a revised version.

MeLonJ10 commented 3 years ago

Thank you for the quick reply. One can use codes/scripts/generate_mod_blur_LR_bic.py to generate LR image for setting1. Can you tell me how to generate LR image for setting2. And it will be greatly appreciated. Thank you so much!

greatlog commented 3 years ago

The test images for Setting 2 are fixed, which is the dataset DIV2KRK. The training images for Setting 2 are generated on the fly. The degradation setting is in train_setting2*.yml. The details can be referred to the preprocessing function.

MeLonJ10 commented 3 years ago

I see! Thanks for your help!