CVMI-Lab / UHDM

(ECCV2022) This is the official PyTorch implementation of ECCV2022 paper: Towards Efficient and Scale-Robust Ultra-High-Definition Image Demoireing
Apache License 2.0
207 stars 28 forks source link

WS-ESDNet : Any code for ws-esdnet #17

Closed shubhoppo closed 1 year ago

shubhoppo commented 1 year ago

Please suggest or upload the code, which shares the learnable parameters among the three branches.

XinYu-Andy commented 1 year ago

Hi, You can just initialize one block here and re-use it in the forward.

Please suggest or upload the code, which shares the learnable parameters among the three branches.

shubhoppo commented 1 year ago

Thanks for the suggestion. Could we expect using 2 SAM block(ESDNET-L) in weight shared manner give better PSNR than ESDNET?

XinYu-Andy commented 1 year ago

Thanks for the suggestion. Could we expect using 2 SAM block(ESDNET-L) in weight shared manner give better PSNR than ESDNET?

It's possible, but I can't give you a definite answer. You may need to try it yourself.

shubhoppo commented 1 year ago

Also after weight sharing model size is compressed but inference time is still same.

XinYu-Andy commented 1 year ago

Also after weight sharing model size is compressed but inference time is still same.

This is an expected behavior since parameter sharing does not reduce computational complexity.

shubhoppo commented 1 year ago

Yeah, Got it. Thanks