SSDLoss with hard negative mining

[sandylaker]

In the implementation of the SSDLoss, the cls_loss of the negative anchors were set to 0,
but later you sum cls[pos|neg]. This is the same as cls[pos], since the loss of all the negative anchors were neglected. So maybe cls_loss[cls_targets<0] = 0 is redundant? The negative mining has already generated a mask that only takes a part of negative anchors into account. Or did I understand wrongly?