junshk / frn Goto Github PK

thanks for sharing your codes.could you share with your pretrained model. and how to train or test with the code.thank you very much.

Hello, @Junshk !
I've been studying the FRN structure you proposed, and I have questions for ablation studies.

1. When you studied to get the effectiveness of down-upsampling modules, you suggested RCAN structure when down-upsampling modules were not equipped. In this case, the body part is only changed, right?

2. The next question is for auto-encoder loss. I wonder the training process if the auto-encoder loss is given. When training, I think the following cases:

• down-upsampling modules w/o FRN body are trained -> the whole model w/ FRN body is trained using the pre-trained weights for down-upsampling modules.
• down-upsampling modules w/o FRN body are trained -> the whole model w/ FRN body is trained using the trained weights for down-upsampling modules, and simultaneously the weights for down-upsampling modules are fine-tuned.
• From scratch, the whole model is trained using L1 + auto-encoder loss.

I'm strongly guessing the first case works.

Thanks in advance! :)

Hello,

Thank you for your work and the repository.
I am trying to train the model on the DIV2K dataset. If I choose a scale factor of 2 and a patchsize of 96 for HR images (48 for LR images), then the downsampling module at the beginning of the network will downscale the LR image to 24x24. After propagating the residual blocks, the upscaling module then upscales the representation by the same scale factor as the downscaling module, i.e. we get a size of 48x48. However, this is not super-resolved to the HR size. Therefore the l1 loss function is not applicable in my case. Is there anything to add in the architecture such that it gets upscaled 96x96?
Thanks!

Edit: Everything cleared out, thanks.