fix rpn transforming bug in two stage networks (#3754)

* fix rpn transforming bug in two_stage

tools/upgrade_model_version.py

@@ -57,7 +57,7 @@ def reorder_cls_channel(val, num_classes=81):
         # fc_cls
         elif out_channels == num_classes:
             new_val = torch.cat((val[1:], val[:1]), dim=0)
-        # agnostic | retina_cls | rpn_cls
+        # agnostic | retina_cls
         else:
             new_val = val
 
@@ -89,7 +89,7 @@ def truncate_cls_channel(val, num_classes=81):
 def truncate_reg_channel(val, num_classes=81):
     # bias
     if val.dim() == 1:
-        # fc_reg|rpn_reg
+        # fc_reg
         if val.size(0) % num_classes == 0:
             new_val = val.reshape(num_classes, -1)[:num_classes - 1]
             new_val = new_val.reshape(-1)
@@ -99,7 +99,7 @@ def truncate_reg_channel(val, num_classes=81):
     # weight
     else:
         out_channels, in_channels = val.shape[:2]
-        # fc_reg|rpn_reg
+        # fc_reg
         if out_channels % num_classes == 0:
             new_val = val.reshape(num_classes, -1, in_channels,
                                   *val.shape[2:])[1:]
@@ -137,14 +137,14 @@ def convert(in_file, out_file, num_classes):
 
         # classification
         m = re.search(
-            r'(conv_cls|retina_cls|rpn_cls|fc_cls|fcos_cls|'
+            r'(conv_cls|retina_cls|fc_cls|fcos_cls|'
             r'fovea_cls).(weight|bias)', new_key)
         if m is not None:
             print(f'reorder cls channels of {new_key}')
             new_val = reorder_cls_channel(val, num_classes)
 
         # regression
-        m = re.search(r'(fc_reg|rpn_reg).(weight|bias)', new_key)
+        m = re.search(r'(fc_reg).(weight|bias)', new_key)
         if m is not None and not reg_cls_agnostic:
             print(f'truncate regression channels of {new_key}')
             new_val = truncate_reg_channel(val, num_classes)