[Refactor]: Refactor layer wise lr decay (#375)

* [Refactor]: Refactor layer wise lr decay

* [Fix]: Remove is_dcn_module and add log

* [Fix]: Fix lint

* [Fix]: Check relative pos bias table is not decayed

* [Fix]: Fix lint

* [Fix]: Fix UT

mmselfsup/engine/optimizers/layer_decay_optim_wrapper_constructor.py

@@ -1,6 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 import json
-from typing import Dict, List, Optional, Union
+from typing import List
 
 import torch
 from mmengine.dist import get_dist_info
@@ -13,7 +13,7 @@ from mmselfsup.registry import (OPTIM_WRAPPER_CONSTRUCTORS, OPTIM_WRAPPERS,
 
 
 def get_layer_id_for_vit(var_name: str, max_layer_id: int) -> int:
-    """Get the layer id to set the different learning rates.
+    """Get the layer id to set the different learning rates for ViT.
 
     Args:
         var_name (str): The key of the model.
@@ -36,7 +36,7 @@ def get_layer_id_for_vit(var_name: str, max_layer_id: int) -> int:
 
 def get_layer_id_for_swin(var_name: str, max_layer_id: int,
                           depths: List[int]) -> int:
-    """Get the layer id to set the different learning rates.
+    """Get the layer id to set the different learning rates for Swin.
 
     Args:
         var_name (str): The key of the model.
@@ -65,31 +65,37 @@ class LearningRateDecayOptimWrapperConstructor(DefaultOptimWrapperConstructor):
     """Different learning rates are set for different layers of backbone.
 
     Note: Currently, this optimizer constructor is built for ViT and Swin.
+
+    In addition to applying layer-wise learning rate decay schedule, this
+    module will not apply weight decay to ``normalization parameters``,
+    ``bias``, ``position embedding``, ``class token``, and
+    ``relative position bias table, automatically. What's more, the
+    ``paramwise_cfg`` in the base module will be ignored.
     """
 
-    def add_params(self,
+    def add_params(self, params: List[dict], module: nn.Module,
-                   params: List[dict],
+                   optimizer_cfg: dict, **kwargs) -> None:
-                   module: nn.Module,
-                   optimizer_cfg: Dict,
-                   prefix: str = '',
-                   is_dcn_module: Optional[Union[int, float]] = None) -> None:
         """Add all parameters of module to the params list.
 
         The parameters of the given module will be added to the list of param
         groups, with specific rules defined by paramwise_cfg.
 
         Args:
-            params (list[dict]): A list of param groups, it will be modified
+            params (List[dict]): A list of param groups, it will be modified
                 in place.
             module (nn.Module): The module to be added.
-            optimizer_cfg (Dict): The configuration of optimizer.
+            optimizer_cfg (dict): The configuration of optimizer.
-            prefix (str): The prefix of the module
+            prefix (str): The prefix of the module.
-            is_dcn_module (int|float|None): If the current module is a
-                submodule of DCN, `is_dcn_module` will be passed to
-                control conv_offset layer's learning rate. Defaults to None.
         """
         logger = MMLogger.get_current_instance()
 
+        # Check if self.param_cfg is not None
+        if len(self.paramwise_cfg) > 0:
+            logger.info('The paramwise_cfg will be ignored, and normalization \
+                parameters, bias, position embedding, class token and \
+                    relative position bias table will not be decayed by \
+                        default.')
+
         model_type = optimizer_cfg.pop('model_type', None)
         # model_type should not be None
         assert model_type is not None, 'When using layer-wise learning rate \
@@ -113,8 +119,11 @@ class LearningRateDecayOptimWrapperConstructor(DefaultOptimWrapperConstructor):
         for name, param in module.named_parameters():
             if not param.requires_grad:
                 continue  # frozen weights
+            # will not decay normalization params, bias, position embedding
+            # class token, relative position bias table
             if len(param.shape) == 1 or name.endswith('.bias') or name in (
-                    'backbone.pos_embed', 'backbone.cls_token'):
+                    'backbone.pos_embed', 'backbone.cls_token'
+            ) or name.endswith('.relative_position_bias_table'):
                 group_name = 'no_decay'
                 this_weight_decay = 0.
             else:

tests/test_engine/test_optimizers/test_layer_decay_optim_wrapper_constructor.py

@@ -118,3 +118,7 @@ def test_learning_rate_decay_optimizer_wrapper_constructor():
     assert optimizer_wrapper.optimizer.param_groups[-1]['lr_scale'] == 1.0
     assert optimizer_wrapper.optimizer.param_groups[-3]['lr_scale'] == 2.0
     assert optimizer_wrapper.optimizer.param_groups[-5]['lr_scale'] == 4.0
+    # check relative pos bias table is not decayed
+    assert optimizer_wrapper.optimizer.param_groups[-4][
+        'weight_decay'] == 0.0 and 'backbone.stages.3.blocks.1.attn.w_msa.relative_position_bias_table' in optimizer_wrapper.optimizer.param_groups[  # noqa
+            -4]['param_names']