gy77/add freeze hook

mmengine/hooks/__init__.py

@@ -3,6 +3,7 @@
 from .early_stopping_hook import EarlyStoppingHook
 from .ema_hook import EMAHook
 from .empty_cache_hook import EmptyCacheHook
+from .freeze_hook import FreezeHook
 from .hook import Hook
 from .iter_timer_hook import IterTimerHook
 from .logger_hook import LoggerHook
@@ -18,5 +19,5 @@
     'Hook', 'IterTimerHook', 'DistSamplerSeedHook', 'ParamSchedulerHook',
     'SyncBuffersHook', 'EmptyCacheHook', 'CheckpointHook', 'LoggerHook',
     'NaiveVisualizationHook', 'EMAHook', 'RuntimeInfoHook', 'ProfilerHook',
-    'PrepareTTAHook', 'NPUProfilerHook', 'EarlyStoppingHook'
+    'PrepareTTAHook', 'NPUProfilerHook', 'EarlyStoppingHook', 'FreezeHook'
 ]

mmengine/hooks/freeze_hook.py

@@ -0,0 +1,224 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import re
+from typing import Optional, Sequence
+
+from mmengine.logging import print_log
+from mmengine.model import BaseModel, is_model_wrapper
+from mmengine.registry import HOOKS
+from .hook import DATA_BATCH, Hook
+
+
+@HOOKS.register_module()
+class FreezeHook(Hook):
+    """FreezeHook is used to freeze/unfreeze network layers when training to a
+    specified epoch/iter.
+
+    Args:
+        freeze_layers (str): Layers of the network that will be freeze. \
+            Match by regular expression.
+        freeze_iter (int): The iter number to start freezing layers.
+        freeze_epoch (int): The epoch number to start freezing layers.
+        unfreeze_layers (str): Layers of the network that will be unfreeze. \
+            Match by regular expression.
+        unfreeze_iter (int): The iter number to start unfreezing layers.
+        unfreeze_epoch (int): The epoch number to start unfreezing layers.
+        verbose  (bool): Whether to log the requires_grad of each layer.
+
+    Notes:
+        The GPU memory usage shown in the "nvidia-smi" command does not change
+        when you freeze model layers.
+        https://discuss.pytorch.org/t/how-can-we-release-gpu-memory-cache/14530/4
+
+    Examples:
+        >>> # The simplest FreezeHook config.
+        >>> freeze_hook_cfg = dict(
+                type="FreezeHook",
+                freeze_layers="backbone.*|neck.*",
+                freeze_epoch=0
+            )
+    """
+
+    def __init__(
+        self,
+        freeze_layers: str,
+        freeze_iter: Optional[int] = None,
+        freeze_epoch: Optional[int] = None,
+        unfreeze_layers: Optional[str] = None,
+        unfreeze_iter: Optional[int] = None,
+        unfreeze_epoch: Optional[int] = None,
+        verbose: bool = False,
+    ) -> None:
+        # check arguments type
+        if not isinstance(freeze_layers, str):
+            raise TypeError('`freeze_layers` must be a str')
+        if not isinstance(freeze_iter, int) and freeze_iter is not None:
+            raise TypeError('`freeze_iter` must be a int or None')
+        if not isinstance(freeze_epoch, int) and freeze_epoch is not None:
+            raise TypeError('`freeze_epoch` must be a int or None')
+        if not isinstance(unfreeze_layers,
+                          str) and unfreeze_layers is not None:
+            raise TypeError('`unfreeze_layers` must be a str or None')
+        if not isinstance(unfreeze_iter, int) and unfreeze_iter is not None:
+            raise TypeError('`unfreeze_iter` must be a int or None')
+        if not isinstance(unfreeze_epoch, int) and unfreeze_epoch is not None:
+            raise TypeError('`unfreeze_epoch` must be a int or None')
+        if not isinstance(verbose, bool):
+            raise TypeError('`verbose`  must be a boolean')
+        # check arguments value
+        if freeze_iter is None and freeze_epoch is None:
+            raise ValueError(
+                '`freeze_iter` and `freeze_epoch` should not be both None.')
+        if freeze_iter is not None and freeze_epoch is not None:
+            raise ValueError(
+                '`freeze_iter` and `freeze_epoch` should not be both set.')
+        if freeze_iter is not None and freeze_iter < 0:
+            raise ValueError(
+                '`freeze_iter` must be greater than or equal to 0')
+        if freeze_epoch is not None and freeze_epoch < 0:
+            raise ValueError(
+                '`freeze_epoch` must be greater than or equal to 0')
+        if unfreeze_layers is not None:
+            if unfreeze_iter is None and unfreeze_epoch is None:
+                raise ValueError('`unfreeze_iter` and `unfreeze_epoch` \
+                                 should not be both None.')
+            if unfreeze_iter is not None and unfreeze_epoch is not None:
+                raise ValueError('`unfreeze_iter` and `unfreeze_epoch` \
+                        should not be both set.')
+        if unfreeze_iter is not None:
+            if unfreeze_iter < 0:
+                raise ValueError(
+                    '`unfreeze_iter` must be greater than or equal to 0')
+            if freeze_iter is None:
+                raise ValueError('`unfreeze_iter` and `freeze_iter` \
+                        should be set at the same time.')
+            if unfreeze_iter < freeze_iter:
+                raise ValueError('`unfreeze_iter` must be greater than \
+                        or equal to `freeze_iter`')
+        if unfreeze_epoch is not None:
+            if unfreeze_epoch < 0:
+                raise ValueError(
+                    '`unfreeze_epoch` must be greater than or equal to 0')
+            if freeze_epoch is None:
+                raise ValueError('`unfreeze_epoch` and `freeze_epoch` \
+                        should be set at the same time.')
+            if unfreeze_epoch < freeze_epoch:
+                raise ValueError('`unfreeze_epoch` must be greater than \
+                        or equal to `freeze_epoch`')
+
+        self.freeze_layers = freeze_layers
+        self.freeze_iter = freeze_iter
+        self.freeze_epoch = freeze_epoch
+        self.unfreeze_layers = unfreeze_layers
+        self.unfreeze_iter = unfreeze_iter
+        self.unfreeze_epoch = unfreeze_epoch
+        self.verbose = verbose
+        self.freeze_layer_names: list = []
+        self.unfreeze_layer_names: list = []
+
+    def _modify_layers_grad(self, model: BaseModel, layers: Sequence[str],
+                            requires_grad: bool):
+        """Modify the ``requires_grad`` of the specified layers.
+
+        Args:
+            model (BaseModel): A BaseModel of mmengine.
+            layers (Sequence[str]): Network layers to be modified.
+            requires_grad (bool): Whether to enable gradient.
+        """
+        if is_model_wrapper(model):
+            model = model.module
+        for k, v in model.named_parameters():
+            if k in layers:
+                v.requires_grad = requires_grad
+
+    def _log_model_grad(self, model: BaseModel):
+        """Print `requires_grad` for all network layers.
+
+        Args:
+            model (BaseModel): A BaseModel of mmengine.
+        """
+        for k, v in model.named_parameters():
+            print_log(
+                f'{k} requires_grad: {v.requires_grad}', logger='current')
+
+    def _freeze(self,
+                runner,
+                idx: int,
+                freeze_idx: int,
+                unfreeze_idx: Optional[int] = None) -> None:
+        """The main function of FreezeHook.
+
+        Args:
+            runner (Runner): The runner of the training process.
+            idx (int): The index in the train loop.
+            freeze_idx (int): The index to start freezing layers.
+            unfreeze_idx (int): The index to start unfreezing layers.
+        """
+        if idx in (freeze_idx, unfreeze_idx):
+            model = runner.model
+            if is_model_wrapper(model):
+                model = model.module
+
+            if idx == freeze_idx:
+                print_log('Start freezing layers.', logger='current')
+                self._modify_layers_grad(model, self.freeze_layer_names, False)
+                # if you want to release GPU memory cache:
+                # import torch; torch.cuda.empty_cache()
+
+            if idx == unfreeze_idx:
+                print_log('Start unfreezing layers.', logger='current')
+                self._modify_layers_grad(model, self.unfreeze_layer_names,
+                                         True)
+
+            if self.verbose:
+                self._log_model_grad(model)
+
+    def before_train(self, runner) -> None:
+        """Collect network layers that will be freeze or unfreeze.
+
+        Args:
+            runner (Runner): The runner of the training process.
+        """
+        model = runner.model
+        if is_model_wrapper(model):
+            model = model.module
+
+        layer_names = ''
+        for k, _ in model.named_parameters():
+            layer_names += f'{k}\n'
+
+        freeze_names = re.findall(self.freeze_layers, layer_names)
+        self.freeze_layer_names = freeze_names
+
+        if self.unfreeze_layers is not None:
+            unfreeze_names = re.findall(self.unfreeze_layers, layer_names)
+            self.unfreeze_layer_names = unfreeze_names
+
+    def before_train_iter(
+        self,
+        runner,
+        batch_idx: int,
+        data_batch: DATA_BATCH = None,
+        outputs: Optional[dict] = None,
+    ) -> None:
+        """Update `requires_grad` before the start of `freeze_iter`
+
+        Args:
+            runner (Runner): The runner of the training process.
+            batch_idx (int): The index of the current batch in the train loop.
+            data_batch (Sequence[dict], optional): Data from dataloader.
+                Defaults to None.
+            outputs (dict, optional): Outputs from model. Defaults to None.
+        """
+        if self.freeze_iter is not None:
+            self._freeze(runner, runner.iter, self.freeze_iter,
+                         self.unfreeze_iter)
+
+    def before_train_epoch(self, runner) -> None:
+        """Update `requires_grad` before the start of `freeze_epoch`
+
+        Args:
+            runner (Runner): The runner of the training process.
+        """
+        if self.freeze_epoch is not None:
+            self._freeze(runner, runner.epoch, self.freeze_epoch,
+                         self.unfreeze_epoch)