Frankluox/Channel_Importance_FSL

Oracle Transformation

Closed this issue · 3 comments

WindFall-wyx commented

Hi, this is a solid and interesting work.
We try to reproduce Oracle Transformation following Eq(6) and F.1, but found that it may even drop 0.5% than the baseline. So could you release your Oracle Transformation code or show more details?
Thanks a lot.

Frankluox commented

Hi, thanks for your interest in our work! We will release the code today.

Frankluox commented

Hi, we have updated the code. The updated oracle transformation is an extended version of that in the ICML paper, additionally supporting features taking negative values. So you can test the transformation using any pre-trained backbones (like ResNet 12 with learkyReLU, CLIP, DINO, SwinTransformer, etc) to investigate the channel importance bias. Any further questions are welcome.

WindFall-wyx commented

Thanks for your prompt reply!