Motivation around setting beta to 1/9

Open davideboschetto opened this issue 2 years ago • 1 comments

Hi everyone, I was wondering what was the motivation around setting the beta parameter for the smooth L1 loss for faster RCNN to 1/9.

The line in question is here: https://github.com/pytorch/vision/blame/8a0b491e51197edd9ca9f9ad7f6d9147ac8a209d/torchvision/models/detection/roi_heads.py#L44

Is this based on some work / intution I could be able to track (but wasn't) or was it a "good guess"?

Oct 27 '23 15:10 davideboschetto

Hi @davideboschetto , looking at https://github.com/pytorch/vision/pull/2113 and the other conversations linked in this PR, it seems that 1/9 was chosen mostly for its empirical performance

Oct 27 '23 15:10 NicolasHug