search

pym1024 / SWAP_Universal

Academic Free License v3.0
0 stars 0 forks source link

The running results of SWAP_CNN are much lower than the reported results. #1

Closed Andrewymd closed 3 days ago

Andrewymd commented 1 week ago

Thanks a lot for your work, could you help me figure out where the problem is? For instance, Spearman's correlation coefficient between the SWAP-Score and CIFAR-10 validation accuracy for DARTS CNNs is 0.93 I tested it three times and the results were almost the same every time. (Spearman's Correlation Coefficient: 0.5966954170713796, Spearman's Correlation Coefficient: 0.5962400067831423, Spearman's Correlation Coefficient: 0.5964299289647295)

pym1024 commented 1 week ago

Hi there,

It's due to a mistake in the line 29 of swap.py. We did some explorations with nn.functional.threshold, but forgot to switch it back to torch.sign when committing to the Github. We've fixed it, you can pull the latest version or fix it by yourself. Thanks for letting us know, and feel free to leave comments if you've any further questions.

Cheers