fix potential bug in binary_cross_entropy (#846)

mmdet/models/losses/cross_entropy_loss.py

@@ -29,14 +29,13 @@ def binary_cross_entropy(pred,
     if pred.dim() != label.dim():
         label, weight = _expand_binary_labels(label, weight, pred.size(-1))
 
-    # element-wise losses
+    # weighted element-wise losses
     if weight is not None:
         weight = weight.float()
     loss = F.binary_cross_entropy_with_logits(
         pred, label.float(), weight, reduction='none')
-    # apply weights and do the reduction
-    loss = weight_reduce_loss(
-        loss, weight=weight, reduction=reduction, avg_factor=avg_factor)
+    # do the reduction for the weighted loss
+    loss = weight_reduce_loss(loss, reduction=reduction, avg_factor=avg_factor)
 
     return loss