twmht opened this issue 7 years ago · 0 comments
Hi,
to normalize loss, you divide the loss by the total count of anchors, but the paper suggest to divide only the positive anchors.
Why did you do that?
thanks