MIC
MIC copied to clipboard
performance variance
Dear Lucas,
I encountered a significant performance variance while training DAFormer with MIC. The lowest performance reached was only 64.57, which is far below the reported 70.6 in the paper. However, I can consistently reproduce the results of DAFormer. I believe this might be attributed to the inherent uncertainty of MIC. Have you considered employing a deterministic algorithm to mitigate such severe performance fluctuations?
Best regards,