search

CR-Gjx / SCC

Pytorch implementation of "Alleviating Semantics Distortion in Unsupervised Low-Level Image-to-Image Translation via Structure Consistency Constraint." (CVPR 2022).
19 stars 1 forks source link

SCC becomes negative #3

Closed Myblackcat0216 closed 1 year ago

Myblackcat0216 commented 1 year ago

Hello, I also have a similar problem, my SCC is negative during training.I set the batch_size to be 1 and lambda_scc to be 5 following the setting suggested in the paper.

CR-Gjx commented 1 year ago

Hello, can you provide more experimental settings? e.g. datasets/ sample generated images/ experimental logs. Importantly, please set the batch_size to be 1, and try to vary lambda_scc to be smaller?

Myblackcat0216 commented 1 year ago

Sorry, I just saw this email, my datasets trainA and trainB have 506 images each.I am using the SCC loss function on a derivative project of cyclegan. The main task is the mutual conversion between the two domains during the day and night. The training log has not been saved yet, and it will be sent to you after training.

CR-Gjx commented 1 year ago

I see, I tried my loss on Day and Night datasets, and the performance is good with lambda_SCC=5. Can you try using my cyclegan script directly and set batch_size =1 ?

Myblackcat0216 commented 1 year ago

Ok, I'll try again. Thank you very much for your help

junzhin commented 1 year ago

Hi, I have a followed-up question, if we set the training batch size to be one during the training, do we need to adjust the learning rate accordingly?

CR-Gjx commented 1 year ago

I never tried to adjust the learning rate. lol

CR-Gjx commented 1 year ago

Problem solved, I will close this issue.