#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
r"""
Utilities for optimization.
"""
from __future__ import annotations
from contextlib import contextmanager
from typing import Any, Callable, Dict, Generator, Iterable, NamedTuple, Optional, Union
from torch import device as Device, dtype as Dtype, Tensor
from torch.nn import Module
[docs]
class TensorCheckpoint(NamedTuple):
values: Tensor
device: Optional[Device] = None
dtype: Optional[Dtype] = None
[docs]
@contextmanager
def delattr_ctx(
instance: object, *attrs: str, enforce_hasattr: bool = False
) -> Generator[None, None, None]:
r"""Contextmanager for temporarily deleting attributes."""
try:
cache = {}
for key in attrs:
if hasattr(instance, key):
cache[key] = getattr(instance, key)
delattr(instance, key)
elif enforce_hasattr:
raise ValueError(
f"Attribute {key} missing from {type(instance)} instance."
)
yield
finally:
for key, cached_val in cache.items():
setattr(instance, key, cached_val)
[docs]
@contextmanager
def parameter_rollback_ctx(
parameters: Dict[str, Tensor],
checkpoint: Optional[Dict[str, TensorCheckpoint]] = None,
**tkwargs: Any,
) -> Generator[Dict[str, TensorCheckpoint], None, None]:
r"""Contextmanager that exits by rolling back a module's state_dict.
Args:
module: Module instance.
name_filter: Optional Boolean function used to filter items by name.
checkpoint: Optional cache of values and tensor metadata specifying the rollback
state for the module (or some subset thereof).
**tkwargs: Keyword arguments passed to `torch.Tensor.to` when copying data from
each tensor in `module.state_dict()` to the internally created checkpoint.
Only adhered to when the `checkpoint` argument is None.
Yields:
A dictionary of TensorCheckpoints for the module's state_dict. Any in-places
changes to the checkpoint will be observed at rollback time. If the checkpoint
is cleared, no rollback will occur.
"""
# Create copies of the orginal values
if checkpoint is None:
checkpoint = {
name: TensorCheckpoint(
values=param.detach().to(**tkwargs).clone(),
device=param.device,
dtype=param.dtype,
)
for name, param in parameters.items()
}
try: # yield the checkpoint dictionary to the user
yield checkpoint
finally: # restore original values of tracked parameters
if checkpoint:
for name, param in parameters.items():
if name in checkpoint:
values, device, dtype = checkpoint[name]
param.data.copy_(values.to(device=device, dtype=dtype))
[docs]
@contextmanager
def module_rollback_ctx(
module: Module,
name_filter: Optional[Callable[[str], bool]] = None,
checkpoint: Optional[Dict[str, TensorCheckpoint]] = None,
**tkwargs: Any,
) -> Generator[Dict[str, TensorCheckpoint], None, None]:
r"""Contextmanager that exits by rolling back a module's state_dict.
Args:
module: Module instance.
name_filter: Optional Boolean function used to filter items by name.
checkpoint: Optional cache of values and tensor metadata specifying the rollback
state for the module (or some subset thereof).
**tkwargs: Keyword arguments passed to `torch.Tensor.to` when copying data from
each tensor in `module.state_dict()` to the internally created checkpoint.
Only adhered to when the `checkpoint` argument is None.
Yields:
A dictionary of TensorCheckpoints for the module's state_dict. Any in-places
changes to the checkpoint will be observed at rollback time. If the checkpoint
is cleared, no rollback will occur.
"""
# Create copies of the orginal values
if checkpoint is None:
checkpoint = {
name: TensorCheckpoint(
values=values.detach().to(**tkwargs).clone(),
device=values.device,
dtype=values.dtype,
)
for name, values in module.state_dict().items()
if name_filter is None or name_filter(name)
}
try: # yield the checkpoint dictionary to the user
yield checkpoint
finally: # restore original values of tracked parameters
if checkpoint:
state_dict = module.state_dict()
for key, (values, device, dtype) in checkpoint.items():
tnsr = state_dict.get(key)
if tnsr is None:
state_dict[key] = values.to(device=device, dtype=dtype)
else:
tnsr[...] = values.to(device=device, dtype=dtype)
module.load_state_dict(state_dict)
[docs]
@contextmanager
def zero_grad_ctx(
parameters: Union[Dict[str, Tensor], Iterable[Tensor]],
zero_on_enter: bool = True,
zero_on_exit: bool = False,
) -> Generator[None, None, None]:
def zero_() -> None:
for param in (
parameters.values() if isinstance(parameters, dict) else parameters
):
if param.grad is not None:
param.grad.zero_()
if zero_on_enter:
zero_()
try:
yield
finally:
if zero_on_exit:
zero_()