Source code for composer.algorithms.stochastic_depth.stochastic_depth

# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0

"""Modules and layers for applying the Stochastic Depth algorithm."""

from __future__ import annotations

import functools
import logging
from typing import Optional, Union

import torch
from torchvision.models.resnet import Bottleneck

from composer.algorithms.stochastic_depth.stochastic_layers import make_resnet_bottleneck_stochastic
from composer.core import Algorithm, Event, State
from composer.core.time import Time, TimeUnit
from composer.loggers import Logger
from composer.utils import module_surgery

log = logging.getLogger(__name__)

_VALID_LAYER_DISTRIBUTIONS = ('uniform', 'linear')

_VALID_STOCHASTIC_METHODS = ('block', 'sample')

_STOCHASTIC_LAYER_MAPPING = {'ResNetBottleneck': (Bottleneck, make_resnet_bottleneck_stochastic)}

__all__ = ['apply_stochastic_depth', 'StochasticDepth']


[docs]def apply_stochastic_depth( model: torch.nn.Module, target_layer_name: str, stochastic_method: str = 'block', drop_rate: float = 0.2, drop_distribution: str = 'linear', ) -> None: """Applies Stochastic Depth (`Huang et al, 2016 <https://arxiv.org/abs/1603.09382>`_) to the specified model. The algorithm replaces the specified target layer with a stochastic version of the layer. The stochastic layer will randomly drop either samples or the layer itself depending on the stochastic method specified. The block-wise version follows the original paper. The sample-wise version follows the implementation used for EfficientNet in the `Tensorflow/TPU repo <https://github.com/tensorflow/tpu>`_. .. note:: Stochastic Depth only works on instances of :class:`torchvision.models.resnet.ResNet` for now. Args: model (torch.nn.Module): model containing modules to be replaced with stochastic versions. target_layer_name (str): Block to replace with a stochastic block equivalent. The name must be registered in ``STOCHASTIC_LAYER_MAPPING`` dictionary with the target layer class and the stochastic layer class. Currently, only :class:`torchvision.models.resnet.Bottleneck` is supported. stochastic_method (str, optional): The version of stochastic depth to use. ``"block"`` randomly drops blocks during training. ``"sample"`` randomly drops samples within a block during training. Default: ``"block"``. drop_rate (float, optional): The base probability of dropping a layer or sample. Must be between 0.0 and 1.0. Default: `0.2``. drop_distribution (str, optional): How ``drop_rate`` is distributed across layers. Value must be one of ``"uniform"`` or ``"linear"``. ``"uniform"`` assigns the same ``drop_rate`` across all layers. ``"linear"`` linearly increases the drop rate across layer depth, starting with 0 drop rate and ending with ``drop_rate``. Default: ``"linear"``. Example: .. testcode:: import composer.functional as cf from torchvision import models model = models.resnet50() cf.apply_stochastic_depth( model, target_layer_name='ResNetBottleneck' ) """ _validate_stochastic_hparams( target_layer_name=target_layer_name, stochastic_method=stochastic_method, drop_rate=drop_rate, drop_distribution=drop_distribution, ) transforms = {} target_layer, stochastic_converter = _STOCHASTIC_LAYER_MAPPING[target_layer_name] module_count = module_surgery.count_module_instances(model, target_layer) stochastic_from_target_layer = functools.partial( stochastic_converter, drop_rate=drop_rate, drop_distribution=drop_distribution, module_count=module_count, stochastic_method=stochastic_method, ) transforms[target_layer] = stochastic_from_target_layer module_surgery.replace_module_classes(model, policies=transforms)
[docs]class StochasticDepth(Algorithm): """Applies Stochastic Depth (`Huang et al, 2016 <https://arxiv.org/abs/1603.09382>`_) to the specified model. The algorithm replaces the specified target layer with a stochastic version of the layer. The stochastic layer will randomly drop either samples or the layer itself depending on the stochastic method specified. The block-wise version follows the original paper. The sample-wise version follows the implementation used for EfficientNet in the `Tensorflow/TPU repo <https://github.com/tensorflow/tpu>`_. Runs on :attr:`.Event.INIT`, as well as :attr:`.Event.BATCH_START` if ``drop_warmup > 0``. .. note:: Stochastic Depth only works on instances of :class:`torchvision.models.resnet.ResNet` for now. Args: target_layer_name (str): Block to replace with a stochastic block equivalent. The name must be registered in ``STOCHASTIC_LAYER_MAPPING`` dictionary with the target layer class and the stochastic layer class. Currently, only :class:`torchvision.models.resnet.Bottleneck` is supported. stochastic_method (str, optional): The version of stochastic depth to use. ``"block"`` randomly drops blocks during training. ``"sample"`` randomly drops samples within a block during training. Default: ``"block"``. drop_rate (float, optional): The base probability of dropping a layer or sample. Must be between 0.0 and 1.0. Default: ``0.2``. drop_distribution (str, optional): How ``drop_rate`` is distributed across layers. Value must be one of ``"uniform"`` or ``"linear"``. ``"uniform"`` assigns the same ``drop_rate`` across all layers. ``"linear"`` linearly increases the drop rate across layer depth, starting with 0 drop rate and ending with ``drop_rate``. Default: ``"linear"``. drop_warmup (str | Time | float, optional): A :class:`Time` object, time-string, or float on ``[0.0, 1.0]`` representing the fraction of the training duration to linearly increase the drop probability to `linear_drop_rate`. Default: ``0.0``. """ def __init__( self, target_layer_name: str, stochastic_method: str = 'block', drop_rate: float = 0.2, drop_distribution: str = 'linear', drop_warmup: Union[float, Time, str] = 0.0, ): log.warning( 'Stochastic depth has known issues of weight mismatch when loading from a checkpoint, which will cause an error when resuming without `load_weights_only=True`.', ) if drop_rate == 0.0: log.warning('Stochastic Depth will have no effect when drop_rate set to 0') self.target_layer_name = target_layer_name self.stochastic_method = stochastic_method self.drop_rate = drop_rate self.drop_distribution = drop_distribution if isinstance(drop_warmup, str): drop_warmup = Time.from_timestring(drop_warmup) if isinstance(drop_warmup, float): drop_warmup = Time(drop_warmup, TimeUnit.DURATION) self.drop_warmup = drop_warmup self.num_stochastic_layers = 0 # Initial count of stochastic layers _validate_stochastic_hparams( stochastic_method=self.stochastic_method, target_layer_name=self.target_layer_name, drop_rate=self.drop_rate, drop_distribution=self.drop_distribution, drop_warmup=str(self.drop_warmup), ) def __repr__(self) -> str: return f"{self.__class__.__name__}(target_layer_name='{self.target_layer_name}',stochastic_method='{self.stochastic_method}',drop_rate={self.drop_rate},drop_distribution='{self.drop_distribution}',drop_warmup={repr(self.drop_warmup)})" @property def find_unused_parameters(self) -> bool: return self.stochastic_method == 'block' @staticmethod def required_on_load() -> bool: return True def match(self, event: Event, state: State) -> bool: return (event == Event.INIT) or (event == Event.BATCH_START and self.drop_warmup > 0.0) def apply(self, event: Event, state: State, logger: Logger) -> Optional[int]: assert state.model is not None target_block, _ = _STOCHASTIC_LAYER_MAPPING[self.target_layer_name] if event == Event.INIT: if module_surgery.count_module_instances(state.model, target_block) == 0: log.warning(f'No {self.target_layer_name} found in model! Algorithm will function as a no-op.') apply_stochastic_depth( state.model, target_layer_name=self.target_layer_name, stochastic_method=self.stochastic_method, drop_rate=self.drop_rate, drop_distribution=self.drop_distribution, ) self.num_stochastic_layers = module_surgery.count_module_instances(state.model, target_block) logger.log_metrics({'stochastic_depth/num_stochastic_layers': self.num_stochastic_layers}) elif event == Event.BATCH_START and self.num_stochastic_layers: elapsed_duration = state.get_elapsed_duration() assert elapsed_duration is not None, 'elapsed duration is set on BATCH_START' if elapsed_duration < self.drop_warmup: current_drop_rate = float(elapsed_duration / self.drop_warmup) * self.drop_rate _update_drop_rate( module=state.model, target_block=target_block, drop_rate=current_drop_rate, drop_distribution=self.drop_distribution, module_count=self.num_stochastic_layers, ) else: current_drop_rate = self.drop_rate logger.log_metrics({'stochastic_depth/drop_rate': current_drop_rate})
def _validate_stochastic_hparams( target_layer_name: str, stochastic_method: str, drop_rate: float, drop_distribution: str, drop_warmup: str = '0dur', ): """Helper function to validate the Stochastic Depth hyperparameter values. """ if stochastic_method and (stochastic_method not in _VALID_STOCHASTIC_METHODS): raise ValueError( f'stochastic_method {stochastic_method} is not supported.' f' Must be one of {_VALID_STOCHASTIC_METHODS}', ) if target_layer_name and (target_layer_name not in _STOCHASTIC_LAYER_MAPPING): raise ValueError( f'target_layer_name {target_layer_name} is not supported with {stochastic_method}.' f' Must be one of {list(_STOCHASTIC_LAYER_MAPPING.keys())}', ) if drop_rate and (drop_rate < 0 or drop_rate > 1): raise ValueError(f'drop_rate must be between 0 and 1: {drop_rate}') if drop_distribution and (drop_distribution not in _VALID_LAYER_DISTRIBUTIONS): raise ValueError( f'drop_distribution "{drop_distribution}" is' f' not supported. Must be one of {list(_VALID_LAYER_DISTRIBUTIONS)}', ) if stochastic_method == 'sample' and Time.from_timestring(drop_warmup).value != 0: raise ValueError(f'drop_warmup can not be used with "sample" stochastic_method') def _update_drop_rate( module: torch.nn.Module, target_block: type[torch.nn.Module], drop_rate: float, drop_distribution: str, module_count: int, module_id: int = 0, ): """Recursively updates a module's drop_rate attributes with a new value. """ for child in module.children(): if isinstance(child, target_block) and hasattr(child, 'drop_rate'): module_id += 1 if drop_distribution == 'linear': current_drop_rate = (module_id / module_count) * drop_rate # type: ignore else: current_drop_rate = drop_rate child.drop_rate = torch.tensor(current_drop_rate) module_id = _update_drop_rate(child, target_block, drop_rate, drop_distribution, module_count, module_id) return module_id