Albert0147/G-SFDA

How to reproduce the results on office-home?

Closed this issue · 4 comments

SkeletonOne commented

It seems that we need the split of the train/test of the source/target datasets for office-home in the utils.py, could you provide the txt files?

Albert0147 commented

Just as VisDA, we randomly split them (only for source), check this line in utils.py

SkeletonOne commented

@Albert0147 Thanks for your reply. I am still a little bit confused:

  1. I need to go through this line then randomly split them;
  2. If I have a different split setting, I may get different results from your published results.
Albert0147 commented

Yeah, that could happen, especially If the datasets are imbalanced, since we found different performance of source model can lead to different target performance after adaptation. But for VisDA and Office-Home, the difference should be small if using different splitting (i.e., with different random seeds).

SkeletonOne commented

@Albert0147 Got it, Thanks!

  • 😄1