the normalization is weird

[twmht]

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