revise group norm (5)

configs/mask_rcnn_r50_fpn_gn_2x.py

@@ -9,6 +9,7 @@ model = dict(
         out_indices=(0, 1, 2, 3),
         frozen_stages=1,
         style='pytorch',
+        # Note: eval_mode and frozen are required args for backbone
         normalize=dict(
             type='GN',
             num_groups=32,

mmdet/models/backbones/resnet.py

@@ -37,8 +37,8 @@ class BasicBlock(nn.Module):
 
         self.norm1_name, norm1 = build_norm_layer(normalize, planes, postfix=1)
         self.norm2_name, norm2 = build_norm_layer(normalize, planes, postfix=2)
-        self.add_module(self.norm1, norm1)
-        self.add_module(self.norm2, norm2)
+        self.add_module(self.norm1_name, norm1)
+        self.add_module(self.norm2_name, norm2)
 
         self.relu = nn.ReLU(inplace=True)
         self.conv2 = conv3x3(planes, planes)
@@ -325,12 +325,12 @@ class ResNet(nn.Module):
         self.relu = nn.ReLU(inplace=True)
         self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
 
+    def _freeze_stages(self):
         if self.frozen_stages >= 0:
             for m in [self.conv1, self.norm1]:
                 for param in m.parameters():
                     param.requires_grad = False
 
-    def _freeze_stages(self):
         for i in range(1, self.frozen_stages + 1):
             m = getattr(self, 'layer{}'.format(i))
             for param in m.parameters():

mmdet/models/utils/norm.py

@@ -31,6 +31,13 @@ def build_norm_layer(cfg, num_features, postfix=''):
     assert isinstance(cfg, dict) and 'type' in cfg
     cfg_ = cfg.copy()
 
+    # eval_mode is supported and popped out for processing in module
+    # having pretrained weight only (e.g. backbone)
+    # raise an exception if eval_mode is in here
+    if 'eval_mode' in cfg:
+        raise Exception('eval_mode for modules without pretrained weights '
+                        'is not supported')
+
     layer_type = cfg_.pop('type')
     if layer_type not in norm_cfg:
         raise KeyError('Unrecognized norm type {}'.format(layer_type))