cherry-pick fix cascade (#2382)

dygraph/ppdet/modeling/heads/bbox_head.py

@@ -266,6 +266,11 @@ class BBoxHead(nn.Layer):
         reg_name = 'loss_bbox_reg'
         loss_bbox = {}
 
+        loss_weight = 1.
+        if fg_inds.numel() == 0:
+            fg_inds = paddle.zeros([1], dtype='int32')
+            loss_weight = 0.
+
         if cls_agnostic_bbox_reg:
             reg_delta = paddle.gather(deltas, fg_inds)
         else:
@@ -292,8 +297,8 @@ class BBoxHead(nn.Layer):
         loss_bbox_reg = paddle.abs(reg_delta - reg_target).sum(
         ) / tgt_labels.shape[0]
 
-        loss_bbox[cls_name] = loss_bbox_cls
-        loss_bbox[reg_name] = loss_bbox_reg
+        loss_bbox[cls_name] = loss_bbox_cls * loss_weight
+        loss_bbox[reg_name] = loss_bbox_reg * loss_weight
 
         return loss_bbox
 

dygraph/ppdet/modeling/heads/cascade_head.py

@@ -196,17 +196,7 @@ class CascadeHead(BBoxHead):
                 if self.training:
                     rois, rois_num, targets = self.bbox_assigner(
                         rois, rois_num, inputs, i, is_cascade=True)
-                    tgt_labels = targets[0]
-                    tgt_labels = paddle.concat(tgt_labels) if len(
-                        tgt_labels) > 1 else tgt_labels[0]
-                    tgt_labels.stop_gradient = True
-                    fg_inds = paddle.nonzero(
-                        paddle.logical_and(tgt_labels >= 0, tgt_labels <
-                                           self.num_classes)).flatten()
-                    if fg_inds.numel() == 0:
-                        targets_list.append(targets_list[-1])
-                    else:
-                        targets_list.append(targets)
+                    targets_list.append(targets)
 
             rois_feat = self.roi_extractor(body_feats, rois, rois_num)
             bbox_feat = self.head(rois_feat, i)