dandingbudanding / bam Goto Github PK

CA+SA
only SA
these two experiments

I wonder if CA can improve the preforence.

I know my question is really low level question, but I can't find the first checkpoint when I did.
Am I missing something or do I need to download your checkpoint for training? Can't I make my checkpoint with my own datasets?

I read the codes.
Sometimes it's put on the tail of the body (e.g. EDSR, AWSRN).
Sometimes it's put on the tail of base blocks (e.g. RCAN, DRLN).

Why don't you put it on the tail of base blocks of EDSR (Residual block) and AWSRN (AWRU or LFB)? Maybe there would be more increase of performance?

By the way, AWSRN adds wn before all convolutional layers, but the attention module don't. Maybe this will cause some performance decrease (lr of AWSRN is 1e-3).

I think you forgot to upload the MDSR and DRLN python files in your release.

Hi, Thank you for sharing the code.

I retrained the IMDN and IMDN_BALANCED_ATTENTION without using pretrained models. However, the PSNR and SSIM are lower than the results shown in the paper.

Is there any tricks? Thanks.