revise norm config

configs/mask_rcnn_r50_fpn_gn_2x.py

@@ -1,4 +1,9 @@
 # model settings
+normalize = dict(
+    type='GN',
+    num_groups=32,
+    frozen=False)
+
 model = dict(
     type='MaskRCNN',
     pretrained='tools/resnet50-GN.path',
@@ -9,20 +14,13 @@
         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,
-            eval_mode=False,
-            frozen=False)),
+        normalize=normalize),
     neck=dict(
         type='FPN',
         in_channels=[256, 512, 1024, 2048],
         out_channels=256,
         num_outs=5,
-        normalize=dict(
-            type='GN',
-            num_groups=32)),
+        normalize=normalize),
     rpn_head=dict(
         type='RPNHead',
         in_channels=256,
@@ -50,9 +48,7 @@
         target_means=[0., 0., 0., 0.],
         target_stds=[0.1, 0.1, 0.2, 0.2],
         reg_class_agnostic=False,
-        normalize=dict(
-            type='GN',
-            num_groups=32)),
+        normalize=normalize),
     mask_roi_extractor=dict(
         type='SingleRoIExtractor',
         roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
@@ -64,9 +60,7 @@
         in_channels=256,
         conv_out_channels=256,
         num_classes=81,
-        normalize=dict(
-            type='GN',
-            num_groups=32)))
+        normalize=normalize))
 
 # model training and testing settings
 train_cfg = dict(

mmdet/models/backbones/resnet.py

@@ -236,11 +236,14 @@ class ResNet(nn.Module):
             the first 1x1 conv layer.
         frozen_stages (int): Stages to be frozen (all param fixed). -1 means
             not freezing any parameters.
-        normalize (dict): dictionary to construct norm layer. Additionally,
-            eval mode and gradent freezing are controlled by
-            eval (bool) and frozen (bool) respectively.
+        normalize (dict): dictionary to construct and config norm layer.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
         with_cp (bool): Use checkpoint or not. Using checkpoint will save some
             memory while slowing down the training speed.
+        zero_init_residual (bool): whether to use zero init for last norm layer
+            in resblocks to let them behave as identity.
     """
 
     arch_settings = {
@@ -261,8 +264,8 @@ def __init__(self,
                  frozen_stages=-1,
                  normalize=dict(
                      type='BN',
-                     eval_mode=True,
                      frozen=False),
+                 norm_eval=True,
                  with_cp=False,
                  zero_init_residual=True):
         super(ResNet, self).__init__()
@@ -278,11 +281,9 @@ def __init__(self,
         assert max(out_indices) < num_stages
         self.style = style
         self.frozen_stages = frozen_stages
-        assert (isinstance(normalize, dict) and 'eval_mode' in normalize
-                and 'frozen' in normalize)
-        self.norm_eval = normalize.pop('eval_mode')
         self.normalize = normalize
         self.with_cp = with_cp
+        self.norm_eval = norm_eval
         self.zero_init_residual = zero_init_residual
         self.block, stage_blocks = self.arch_settings[depth]
         self.stage_blocks = stage_blocks[:num_stages]
@@ -350,7 +351,6 @@ def init_weights(self, pretrained=None):
                 elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                     constant_init(m, 1)
 
-            # zero init for last norm layer https://arxiv.org/abs/1706.02677
             if self.zero_init_residual:
                 for m in self.modules():
                     if isinstance(m, Bottleneck):

mmdet/models/backbones/resnext.py

@@ -122,11 +122,14 @@ class ResNeXt(ResNet):
             the first 1x1 conv layer.
         frozen_stages (int): Stages to be frozen (all param fixed). -1 means
             not freezing any parameters.
-        normalize (dict): dictionary to construct norm layer. Additionally,
-            eval mode and gradent freezing are controlled by
-            eval (bool) and frozen (bool) respectively.
+        normalize (dict): dictionary to construct and config norm layer.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only.
         with_cp (bool): Use checkpoint or not. Using checkpoint will save some
             memory while slowing down the training speed.
+        zero_init_residual (bool): whether to use zero init for last norm layer
+            in resblocks to let them behave as identity.
     """
 
     arch_settings = {

mmdet/models/utils/norm.py

@@ -31,13 +31,6 @@ 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))