upgrade cascade model (#6346)

* add reg_class_agnostic

* add loss_rpn_bbox

ppdet/modeling/heads/cascade_head.py

@@ -160,7 +160,8 @@ class CascadeHead(BBoxHead):
                  bbox_weight=[[10., 10., 5., 5.], [20.0, 20.0, 10.0, 10.0],
                               [30.0, 30.0, 15.0, 15.0]],
                  num_cascade_stages=3,
-                 bbox_loss=None):
+                 bbox_loss=None,
+                 reg_class_agnostic=True):
         nn.Layer.__init__(self, )
         self.head = head
         self.roi_extractor = roi_extractor
@@ -173,6 +174,9 @@ class CascadeHead(BBoxHead):
         self.num_cascade_stages = num_cascade_stages
         self.bbox_loss = bbox_loss
 
+        self.reg_class_agnostic = reg_class_agnostic
+        num_bbox_delta = 4 if reg_class_agnostic else 4 * num_classes
+
         self.bbox_score_list = []
         self.bbox_delta_list = []
         for i in range(num_cascade_stages):
@@ -190,7 +194,7 @@ class CascadeHead(BBoxHead):
                 delta_name,
                 nn.Linear(
                     in_channel,
-                    4,
+                    num_bbox_delta,
                     weight_attr=paddle.ParamAttr(initializer=Normal(
                         mean=0.0, std=0.001))))
             self.bbox_score_list.append(bbox_score)
@@ -227,6 +231,13 @@ class CascadeHead(BBoxHead):
             bbox_feat = self.head(rois_feat, i)
             scores = self.bbox_score_list[i](bbox_feat)
             deltas = self.bbox_delta_list[i](bbox_feat)
+
+            # TODO (lyuwenyu) Is it correct for only one class ?
+            if not self.reg_class_agnostic and i < self.num_cascade_stages - 1:
+                deltas = deltas.reshape([-1, self.num_classes, 4])
+                labels = scores[:, :-1].argmax(axis=-1)
+                deltas = deltas[paddle.arange(deltas.shape[0]), labels]
+
             head_out_list.append([scores, deltas, rois])
             pred_bbox = self._get_pred_bbox(deltas, rois, self.bbox_weight[i])
 

ppdet/modeling/proposal_generator/rpn_head.py

@@ -68,6 +68,7 @@ class RPNHead(nn.Layer):
             derived by from_config
     """
     __shared__ = ['export_onnx']
+    __inject__ = ['loss_rpn_bbox']
 
     def __init__(self,
                  anchor_generator=_get_class_default_kwargs(AnchorGenerator),
@@ -76,7 +77,8 @@ class RPNHead(nn.Layer):
                                                           12000, 2000),
                  test_proposal=_get_class_default_kwargs(ProposalGenerator),
                  in_channel=1024,
-                 export_onnx=False):
+                 export_onnx=False,
+                 loss_rpn_bbox=None):
         super(RPNHead, self).__init__()
         self.anchor_generator = anchor_generator
         self.rpn_target_assign = rpn_target_assign
@@ -91,6 +93,7 @@ class RPNHead(nn.Layer):
             self.train_proposal = ProposalGenerator(**train_proposal)
         if isinstance(test_proposal, dict):
             self.test_proposal = ProposalGenerator(**test_proposal)
+        self.loss_rpn_bbox = loss_rpn_bbox
 
         num_anchors = self.anchor_generator.num_anchors
         self.rpn_feat = RPNFeat(in_channel, in_channel)
@@ -298,7 +301,12 @@ class RPNHead(nn.Layer):
             loc_tgt = paddle.concat(loc_tgt)
             loc_tgt = paddle.gather(loc_tgt, pos_ind)
             loc_tgt.stop_gradient = True
+
+            if self.loss_rpn_bbox is None:
                 loss_rpn_reg = paddle.abs(loc_pred - loc_tgt).sum()
+            else:
+                loss_rpn_reg = self.loss_rpn_bbox(loc_pred, loc_tgt).sum()
+
         return {
             'loss_rpn_cls': loss_rpn_cls / norm,
             'loss_rpn_reg': loss_rpn_reg / norm