XPixelGroup
Can't Reproduce the Results of SRx3
I'm trying to reproduce the experiment result in the paper with HAT and HAT-L (for SRx3 tasks). But I have not been able to achieve a satisfactory result yet.
Specifically, following weights are used on datasets BSDS100, manga109, Set5, Set14 and urban100:
- HAT_SRx3_ImageNet-pretrain.pth
- HAT-L_SRx3_ImageNet-pretrain.pth
The list above shows my reproduction results (in red font) compared to the results in the paper (in black font).
It is worth noting that the results obtained on BSD100 are very close to the results presented in your paper, but the results of other datasets are quite different from the expected.
I'm wondering whether I made some mistakes during experiments, or the pretrained weights you provide haven't been fine tuned for specific tasks?
My config file for HAT-L_SRx3_ImageNet-pretrain is provided as an attachment.
HAT-L_SRx3_ImageNet-pretrain.zip
