ma-xu/Context-Cluster

the code question

woaixueyingying opened this issue · 5 comments

woaixueyingying commented

image
image

C in the formula for the aggregated feature G is denoted by 'mask' in the code, why is it not denoted by 'sim' at the beginning.

YangTracy commented

Yes, I also have the same question. The 'mask' in the code should be 'sim'. Maybe the code makes a mistake here.

ma-xu commented

@YangTracy @woaixueyingying Thanks very much for the detailed question. I will double-check it and reply by this week.

YangTracy commented
1chizhang commented

I also found this problem, so what exactly is happening and what effect would replacing c with the sim in eq cause?

ma-xu commented

@YangTracy @Volodymyr233 @woaixueyingying

Thanks for this issue!

Double-check with co-authors, this is a small bug, no influence to our conclusions and contributions.
We re-trained coc-tiny on ImageNet1k with same settings and achieved 72.2, an improvement of 0.4 (reported 71.8).

The slight improvement can be contributed to the soft/weighted summation, instead of mask (cn be considered as a hard version of C in Eq. 1).

We will updated all checkpoints once we finished re-training all models. Again, thanks for your kind reminder.

  • 👍2