fix mask rcnn, test=dygraph (#2209)

dygraph/configs/faster_rcnn/_base_/faster_rcnn_r50.yml

@@ -34,7 +34,7 @@ RPNHead:
     nms_thresh: 0.7
     pre_nms_top_n: 12000
     post_nms_top_n: 2000
-    topk_after_collect: True
+    topk_after_collect: False
   test_proposal:
     min_size: 0.0
     nms_thresh: 0.7

dygraph/configs/mask_rcnn/_base_/mask_rcnn_r50.yml

@@ -35,7 +35,7 @@ RPNHead:
     nms_thresh: 0.7
     pre_nms_top_n: 12000
     post_nms_top_n: 2000
-    topk_after_collect: True
+    topk_after_collect: False
   test_proposal:
     min_size: 0.0
     nms_thresh: 0.7

dygraph/ppdet/modeling/backbones/resnet.py

@@ -555,7 +555,7 @@ class Res5Head(nn.Layer):
     def out_shape(self):
         return [ShapeSpec(
             channels=self.feat_out,
-            stride=32, )]
+            stride=16, )]
 
     def forward(self, roi_feat, stage=0):
         y = self.res5(roi_feat)

dygraph/ppdet/modeling/bbox_utils.py

@@ -50,7 +50,7 @@ def delta2bbox(deltas, boxes, weights):
     dy = deltas[:, 1::4] / wy
     dw = deltas[:, 2::4] / ww
     dh = deltas[:, 3::4] / wh
-    # Prevent sending too large values into np.exp()
+    # Prevent sending too large values into paddle.exp()
     dw = paddle.clip(dw, max=clip_scale)
     dh = paddle.clip(dh, max=clip_scale)
 
@@ -64,7 +64,7 @@ def delta2bbox(deltas, boxes, weights):
     pred_boxes.append(pred_ctr_y - 0.5 * pred_h)
     pred_boxes.append(pred_ctr_x + 0.5 * pred_w)
     pred_boxes.append(pred_ctr_y + 0.5 * pred_h)
-    pred_boxes = paddle.concat(pred_boxes, axis=-1)
+    pred_boxes = paddle.stack(pred_boxes, axis=-1)
 
     return pred_boxes
 

dygraph/ppdet/modeling/heads/bbox_head.py

@@ -129,7 +129,7 @@ class BBoxHead(nn.Layer):
     def forward(self, body_feats=None, rois=None, rois_num=None, inputs=None):
         """
         body_feats (list[Tensor]): Feature maps from backbone
-        rois (Tensor): RoIs generated from RPN module
+        rois (list[Tensor]): RoIs generated from RPN module
         rois_num (Tensor): The number of RoIs in each image
         inputs (dict{Tensor}): The ground-truth of image
         """

dygraph/ppdet/modeling/heads/cascade_head.py

@@ -187,6 +187,7 @@ class CascadeHead(BBoxHead):
         pred_proposals = paddle.concat(proposals) if len(
             proposals) > 1 else proposals[0]
         pred_bbox = delta2bbox(deltas, pred_proposals, weights)
+        pred_bbox = paddle.reshape(pred_bbox, [-1, deltas.shape[-1]])
         num_prop = [p.shape[0] for p in proposals]
         return pred_bbox.split(num_prop)
 

dygraph/ppdet/modeling/proposal_generator/rpn_head.py

@@ -116,7 +116,6 @@ class RPNHead(nn.Layer):
 
         rois, rois_num = self._gen_proposal(scores, deltas, anchors, inputs,
                                             batch_size)
-
         if self.training:
             loss = self.get_loss(scores, deltas, anchors, inputs)
             return rois, rois_num, loss

dygraph/ppdet/py_op/post_process.py

@@ -55,6 +55,8 @@ def get_seg_res(masks, bboxes, mask_nums, image_id, label_to_cat_id_map):
             score = float(bboxes[k][1])
             label = int(bboxes[k][0])
             k = k + 1
+            if label == -1:
+                continue
             cat_id = label_to_cat_id_map[label]
             rle = mask_util.encode(
                 np.array(