fix cascade series models (#1588)

ppdet/modeling/architectures/cascade_mask_rcnn.py

@@ -135,6 +135,7 @@ class CascadeMaskRCNN(object):
 
         proposals = None
         bbox_pred = None
+        max_overlap = None
         for i in range(3):
             if i > 0:
                 refined_bbox = self._decode_box(
@@ -146,10 +147,14 @@ class CascadeMaskRCNN(object):
 
             if mode == 'train':
                 outs = self.bbox_assigner(
-                    input_rois=refined_bbox, feed_vars=feed_vars, curr_stage=i)
+                    input_rois=refined_bbox,
+                    feed_vars=feed_vars,
+                    curr_stage=i,
+                    max_overlap=max_overlap)
 
                 proposals = outs[0]
-                rcnn_target_list.append(outs)
+                max_overlap = outs[-1]
+                rcnn_target_list.append(outs[:-1])
             else:
                 proposals = refined_bbox
             proposal_list.append(proposals)

ppdet/modeling/architectures/cascade_rcnn.py

@@ -128,6 +128,7 @@ class CascadeRCNN(object):
 
         proposals = None
         bbox_pred = None
+        max_overlap = None
         for i in range(3):
             if i > 0:
                 refined_bbox = self._decode_box(
@@ -139,10 +140,14 @@ class CascadeRCNN(object):
 
             if mode == 'train':
                 outs = self.bbox_assigner(
-                    input_rois=refined_bbox, feed_vars=feed_vars, curr_stage=i)
+                    input_rois=refined_bbox,
+                    feed_vars=feed_vars,
+                    curr_stage=i,
+                    max_overlap=max_overlap)
 
                 proposals = outs[0]
-                rcnn_target_list.append(outs)
+                max_overlap = outs[-1]
+                rcnn_target_list.append(outs[:-1])
             else:
                 proposals = refined_bbox
             proposal_list.append(proposals)

ppdet/modeling/architectures/cascade_rcnn_cls_aware.py

@@ -117,6 +117,7 @@ class CascadeRCNNClsAware(object):
 
         self.cascade_decoded_box = []
         self.cascade_cls_prob = []
+        max_overlap = None
 
         for stage in range(3):
             if stage > 0:
@@ -126,9 +127,13 @@ class CascadeRCNNClsAware(object):
             if mode == "train":
                 self.cascade_var_v[stage].stop_gradient = True
                 outs = self.bbox_assigner(
-                    input_rois=pool_rois, feed_vars=feed_vars, curr_stage=stage)
+                    input_rois=pool_rois,
+                    feed_vars=feed_vars,
+                    curr_stage=stage,
+                    max_overlap=max_overlap)
                 pool_rois = outs[0]
-                rcnn_target_list.append(outs)
+                max_overlap = outs[-1]
+                rcnn_target_list.append(outs[:-1])
 
             # extract roi features
             roi_feat = self.roi_extractor(body_feats, pool_rois, spatial_scale)

ppdet/modeling/architectures/htc.py

@@ -158,13 +158,18 @@ class HybridTaskCascade(object):
         bbox_pred = None
         outs = None
         refined_bbox = rpn_rois
+        max_overlap = None
         for i in range(self.num_stage):
             # BBox Branch
             if mode == 'train':
                 outs = self.bbox_assigner(
-                    input_rois=refined_bbox, feed_vars=feed_vars, curr_stage=i)
+                    input_rois=refined_bbox,
+                    feed_vars=feed_vars,
+                    curr_stage=i,
+                    max_overlap=max_overlap)
                 proposals = outs[0]
-                rcnn_target_list.append(outs)
+                max_overlap = outs[-1]
+                rcnn_target_list.append(outs[:-1])
             else:
                 proposals = refined_bbox
             proposal_list.append(proposals)

ppdet/modeling/target_assigners.py

@@ -53,7 +53,7 @@ class CascadeBBoxAssigner(object):
         self.use_random = shuffle_before_sample
         self.class_aware = class_aware
 
-    def __call__(self, input_rois, feed_vars, curr_stage):
+    def __call__(self, input_rois, feed_vars, curr_stage, max_overlap=None):
 
         curr_bbox_reg_w = [
             1. / self.bbox_reg_weights[curr_stage],
@@ -76,5 +76,7 @@ class CascadeBBoxAssigner(object):
             class_nums=self.class_nums if self.class_aware else 2,
             is_cls_agnostic=not self.class_aware,
             is_cascade_rcnn=True
-            if curr_stage > 0 and not self.class_aware else False)
+            if curr_stage > 0 and not self.class_aware else False,
+            max_overlap=max_overlap,
+            return_max_overlap=True)
         return outs