fix distribute train in cascade_rcnn (#2246)

dygraph/ppdet/modeling/heads/bbox_head.py

@@ -265,10 +265,6 @@ class BBoxHead(nn.Layer):
         reg_name = 'loss_bbox_reg'
         loss_bbox = {}
 
-        if fg_inds.numel() == 0:
-            loss_bbox[cls_name] = paddle.to_tensor(0., dtype='float32')
-            loss_bbox[reg_name] = paddle.to_tensor(0., dtype='float32')
-            return loss_bbox
         if cls_agnostic_bbox_reg:
             reg_delta = paddle.gather(deltas, fg_inds)
         else:

dygraph/ppdet/modeling/heads/cascade_head.py

@@ -196,7 +196,17 @@ class CascadeHead(BBoxHead):
                 if self.training:
                     rois, rois_num, targets = self.bbox_assigner(
                         rois, rois_num, inputs, i, is_cascade=True)
-                    targets_list.append(targets)
+                    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)
 
             rois_feat = self.roi_extractor(body_feats, rois, rois_num)
             bbox_feat = self.head(rois_feat, i)
@@ -227,6 +237,8 @@ class CascadeHead(BBoxHead):
             clip_box = clip_bbox(boxes_per_image, im_shape[i])
             if self.training:
                 keep = nonempty_bbox(clip_box)
+                if keep.shape[0] == 0:
+                    continue
                 clip_box = paddle.gather(clip_box, keep)
             rois.append(clip_box)
         rois_num = paddle.concat([paddle.shape(r)[0] for r in rois])