Why use neg_avg_factor = (1 - p_neg_weight).sum() ?

[Icecream-blue-sky]

In loss function:

neg_avg_factor = (1 - p_neg_weight).sum()
...
cls_neg_loss = sum(cls_neg_loss_list) / neg_avg_factor

Why not use neg_avg_factor = (p_neg_weight).sum() ?

[strongwolf]

The value of (p_neg_weight).sum() is too small and it will cause unstable training.