# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0
"""Base class for algorithms that improve a model's quality or efficiency."""
from __future__ import annotations
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING, Any, Optional
from composer.core.serializable import Serializable
if TYPE_CHECKING:
from composer.core import Event, State
from composer.loggers import Logger
__all__ = ['Algorithm']
[docs]class Algorithm(Serializable, ABC):
"""Base class for algorithms.
Algorithms are pieces of code which run at specific events (see :class:`.Event`) in the training loop.
Algorithms modify the trainer's :class:`.State`, generally with the effect of improving the model's quality
or increasing the efficiency and throughput of the training loop.
Algorithms must implement the following two methods:
+----------------+-------------------------------------------------------------------------------+
| Method | Description |
+================+===============================================================================+
| :func:`match` | returns whether the algorithm should be run given the current |
| | :class:`.Event` and :class:`.State`. |
+----------------+-------------------------------------------------------------------------------+
| :func:`apply` | Executes the algorithm's code and makes an in-place change |
| | to the :class:`.State`. |
+----------------+-------------------------------------------------------------------------------+
"""
def __init__(self, *args, **kwargs): # Stub signature for PyRight
del args, kwargs # unused
pass
@property
def find_unused_parameters(self) -> bool:
"""Indicates whether this algorithm may cause some model parameters to be unused. Defaults to False.
For example, it is used to tell :class:`torch.nn.parallel.DistributedDataParallel` (DDP) that some parameters
will be frozen during training, and hence it should not expect gradients from them. All algorithms which do any
kind of parameter freezing should override this function to return ``True``.
.. note::
DeepSpeed integration with this function returning True is not tested. It may not work as expected.
"""
return False
@property
def backwards_create_graph(self) -> bool:
"""Whether this algorithm requires the backwards pass to be differentiable. Defaults to ``False``.
If it returns ``True``, ``create_graph=True`` will be passed to :meth:`torch.Tensor.backward` which will result in
the graph of the gradient also being constructed. This allows the computation of second order derivatives.
"""
return False
[docs] @staticmethod
def required_on_load() -> bool:
"""Return `True` to indicate this algorithm is required when loading from a checkpoint which used it."""
return False
def state_dict(self) -> dict[str, Any]:
return {'repr': self.__repr__()}
[docs] @abstractmethod
def match(self, event: Event, state: State) -> bool:
"""Determines whether this algorithm should run given the current :class:`.Event` and :class:`.State`.
Examples:
To only run on a specific event (e.g., on :attr:`.Event.BEFORE_LOSS`), override match as shown below:
>>> class MyAlgorithm:
... def match(self, event, state):
... return event == Event.BEFORE_LOSS
>>> MyAlgorithm().match(Event.BEFORE_LOSS, state)
True
To run based on some value of a :class:`.State` attribute, override match as shown below:
>>> class MyAlgorithm:
... def match(self, event, state):
... return state.timestamp.epoch > 30
>>> MyAlgorithm().match(Event.BEFORE_LOSS, state)
False
See :class:`.State` for accessible attributes.
Args:
event (Event): The current event.
state (State): The current state.
Returns:
bool: True if this algorithm should run now.
"""
raise NotImplementedError(f'implement match() required for {self.__class__.__name__}')
[docs] @abstractmethod
def apply(self, event: Event, state: State, logger: Logger) -> Optional[int]:
"""Applies the algorithm to make an in-place change to the :class:`.State`.
Can optionally return an exit code to be stored in a :class:`.Trace`.
This exit code is made accessible for debugging.
Args:
event (Event): The current event.
state (State): The current state.
logger (Logger): A logger to use for logging algorithm-specific metrics.
Returns:
int or None: exit code that will be stored in :class:`.Trace` and made accessible for debugging.
"""
raise NotImplementedError(f'implement apply() required for {self.__class__.__name__}')