[Fix]: fix bug of autoassign (#5092)

* fix bug of norm pre

* fix init of fpn

* use reg feature

configs/autoassign/autoassign_r50_fpn_8x2_1x_coco.py

@@ -23,12 +23,11 @@
         start_level=1,
         add_extra_convs=True,
         extra_convs_on_inputs=True,
-        caffe2_xavier_init=True,
         num_outs=5,
-        relu_before_extra_convs=True),
+        relu_before_extra_convs=True,
+        init_cfg=dict(type='Caffe2Xavier', layer='Conv2d')),
     bbox_head=dict(
         type='AutoAssignHead',
-        norm_on_bbox=True,
         num_classes=80,
         in_channels=256,
         stacked_convs=4,

mmdet/models/dense_heads/autoassign_head.py

@@ -148,8 +148,7 @@ def __init__(self,
                  neg_loss_weight=0.75,
                  center_loss_weight=0.75,
                  **kwargs):
-        super().__init__(
-            *args, conv_bias=True, centerness_on_reg=True, **kwargs)
+        super().__init__(*args, conv_bias=True, **kwargs)
         self.center_prior = CenterPrior(
             force_topk=force_topk,
             topk=topk,
@@ -187,6 +186,31 @@ def _get_points_single(self,
                              dim=-1)
         return points
 
+    def forward_single(self, x, scale, stride):
+        """Forward features of a single scale level.
+
+        Args:
+            x (Tensor): FPN feature maps of the specified stride.
+            scale (:obj: `mmcv.cnn.Scale`): Learnable scale module to resize
+                the bbox prediction.
+            stride (int): The corresponding stride for feature maps, only
+                used to normalize the bbox prediction when self.norm_on_bbox
+                is True.
+
+        Returns:
+            tuple: scores for each class, bbox predictions and centerness \
+                predictions of input feature maps.
+        """
+        cls_score, bbox_pred, cls_feat, reg_feat = super(
+            FCOSHead, self).forward_single(x)
+        centerness = self.conv_centerness(reg_feat)
+        # scale the bbox_pred of different level
+        # float to avoid overflow when enabling FP16
+        bbox_pred = scale(bbox_pred).float()
+        bbox_pred = F.relu(bbox_pred)
+        bbox_pred *= stride
+        return cls_score, bbox_pred, centerness
+
     def get_pos_loss_single(self, cls_score, objectness, reg_loss, gt_labels,
                             center_prior_weights):
         """Calculate the positive loss of all points in gt_bboxes.
@@ -440,7 +464,6 @@ def get_targets(self, points, gt_bboxes_list):
                   (num_points, num_gt, 4).
         """
 
-        num_levels = len(points)
         concat_points = torch.cat(points, dim=0)
         # the number of points per img, per lvl
         num_points = [center.size(0) for center in points]
@@ -450,11 +473,6 @@ def get_targets(self, points, gt_bboxes_list):
             list(bbox_targets.split(num_points, 0))
             for bbox_targets in bbox_targets_list
         ]
-        if self.norm_on_bbox:
-            for i in range(num_levels):
-                for j in range(len(bbox_targets_list)):
-                    bbox_targets_list[j][
-                        i] = bbox_targets_list[j][i] / self.strides[i]
         concat_lvl_bbox_targets = [
             torch.cat(item, dim=0) for item in bbox_targets_list
         ]

mmdet/models/necks/fpn.py

@@ -40,8 +40,6 @@ class FPN(BaseModule):
             conv. Default: False.
         no_norm_on_lateral (bool): Whether to apply norm on lateral.
             Default: False.
-        caffe2_xavier_init (bool): Whether to apply caffe2_xavier_init on all
-            conv in FPN. Default: False.
         conv_cfg (dict): Config dict for convolution layer. Default: None.
         norm_cfg (dict): Config dict for normalization layer. Default: None.
         act_cfg (str): Config dict for activation layer in ConvModule.
@@ -76,7 +74,6 @@ def __init__(self,
                  extra_convs_on_inputs=True,
                  relu_before_extra_convs=False,
                  no_norm_on_lateral=False,
-                 caffe2_xavier_init=False,
                  conv_cfg=None,
                  norm_cfg=None,
                  act_cfg=None,
@@ -93,7 +90,6 @@ def __init__(self,
         self.no_norm_on_lateral = no_norm_on_lateral
         self.fp16_enabled = False
         self.upsample_cfg = upsample_cfg.copy()
-        self.caffe2_xavier_init = caffe2_xavier_init
 
         if end_level == -1:
             self.backbone_end_level = self.num_ins