Add frozen_stages kwarg to InternImage backbone (#283)

Author: czczup

.gitignore

@@ -6,4 +6,4 @@ segmentation/convertor/
 checkpoint_dir/
 demo/
 pretrained/
-upload.py
+upload.py

detection/mmdet_custom/models/backbones/intern_image.py

@@ -575,6 +575,7 @@ def __init__(self,
                  center_feature_scale=False,  # for InternImage-H/G
                  use_dcn_v4_op=False,
                  out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
                  init_cfg=None,
                  **kwargs):
         super().__init__()
@@ -588,6 +589,8 @@ def __init__(self,
         self.init_cfg = init_cfg
         self.out_indices = out_indices
         self.level2_post_norm_block_ids = level2_post_norm_block_ids
+        self.frozen_stages = frozen_stages
+
         logger = get_root_logger()
         logger.info(f'using core type: {core_op}')
         logger.info(f'using activation layer: {act_layer}')
@@ -642,6 +645,19 @@ def __init__(self,
         self.num_layers = len(depths)
         self.apply(self._init_weights)
         self.apply(self._init_deform_weights)
+        self._freeze_stages()
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep normalization layer frozen."""
+        super(InternImage, self).train(mode)
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            for level in self.levels[:self.frozen_stages]:
+                level.eval()
+                for param in level.parameters():
+                    param.requires_grad = False
 
     def init_weights(self):
         logger = get_root_logger()

segmentation/mmseg_custom/models/backbones/intern_image.py

@@ -575,6 +575,7 @@ def __init__(self,
                  center_feature_scale=False,  # for InternImage-H/G
                  use_dcn_v4_op=False,
                  out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
                  init_cfg=None,
                  **kwargs):
         super().__init__()
@@ -588,6 +589,8 @@ def __init__(self,
         self.init_cfg = init_cfg
         self.out_indices = out_indices
         self.level2_post_norm_block_ids = level2_post_norm_block_ids
+        self.frozen_stages = frozen_stages
+
         logger = get_root_logger()
         logger.info(f'using core type: {core_op}')
         logger.info(f'using activation layer: {act_layer}')
@@ -642,6 +645,19 @@ def __init__(self,
         self.num_layers = len(depths)
         self.apply(self._init_weights)
         self.apply(self._init_deform_weights)
+        self._freeze_stages()
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep normalization layer frozen."""
+        super(InternImage, self).train(mode)
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            for level in self.levels[:self.frozen_stages]:
+                level.eval()
+                for param in level.parameters():
+                    param.requires_grad = False
 
     def init_weights(self):
         logger = get_root_logger()