Some findings when running the demo.

[singer-yang]

1, nac_resnet_set12.py can not be executed, many lines need to be revised to get an executable demo. 2, after running nac_resnet_set12.py with the original parameters, 40GB files are generated. I suggest reducing the epoch for every image from 1001 to 100 because the results seem not increasing after 100 iterations.

[csjunxu]

Hi，yang, thank you for your comments, could you tell me which lins should be revised? The directory path and the noise levels should be revised, and any others?

Yes, 100 epochs are enough, and you can revise the code to not save the intermediate files when running the code, since the best results are usually obtained in the last epoch.

[singer-yang]

Hi Jun, just clone and run the demo, you will find an "input size error". Because I have not totally understood the code yet, I need some time to figure it out. By the way, there is too much model saving module in the code, leaving the nac part hard to find.

I have a question, why do you say "the best results are usually obtained in the last epoch"? What I understand is that since we do not know the ground truth, whether this self-supervised network converges remains unclear to me. Maybe you have the answer.

[csjunxu]

The reason "the best results are usually obtained in the last epoch" is that, NAC averages the outputs of current epoch and the previous epochs by a weight alpha, that is, Output_CurrentEpoch= (1-alpha)Output_CurrentEpoch + alphaOutput_PreviousEpochs, where alpha is usually 0.99. Therefore, the output is usually becoming slightly better or remaining the same with the increasing of the training epochs. This is observed in the results.

Yes, you are right that NAC produces too much files not useful.