File size: 8,330 Bytes
9c6594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 |
# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
from collections.abc import Generator
from contextlib import AbstractContextManager, contextmanager
from typing import Any, Callable, Optional
import torch
import torch.distributed as dist
from torch import Tensor
import pytorch_lightning as pl
from lightning_fabric.utilities.distributed import _distributed_is_initialized
from lightning_fabric.utilities.warnings import PossibleUserWarning
from pytorch_lightning.accelerators.xla import XLAAccelerator
from pytorch_lightning.callbacks.timer import Timer
from pytorch_lightning.loops import _Loop
from pytorch_lightning.loops.fetchers import _DataFetcher, _DataLoaderIterDataFetcher, _PrefetchDataFetcher
from pytorch_lightning.loops.progress import _BaseProgress
from pytorch_lightning.strategies import FSDPStrategy
from pytorch_lightning.strategies.parallel import ParallelStrategy
from pytorch_lightning.strategies.strategy import Strategy
from pytorch_lightning.trainer.states import RunningStage
from pytorch_lightning.utilities.rank_zero import rank_zero_warn
from pytorch_lightning.utilities.signature_utils import is_param_in_hook_signature
def check_finite_loss(loss: Optional[Tensor]) -> None:
"""Checks for finite loss value.
Args:
loss: the loss value to check to be finite
"""
if loss is not None and not torch.isfinite(loss).all():
raise ValueError(f"The loss returned in `training_step` is {loss}.")
def _parse_loop_limits(
min_steps: Optional[int],
max_steps: int,
min_epochs: Optional[int],
max_epochs: Optional[int],
trainer: "pl.Trainer",
) -> tuple[int, int]:
"""This utility computes the default values for the minimum and maximum number of steps and epochs given the values
the user has selected.
Args:
min_steps: Minimum number of steps.
max_steps: Maximum number of steps.
min_epochs: Minimum number of epochs.
max_epochs: Maximum number of epochs.
trainer: Trainer instance.
Returns:
The parsed limits, with default values being set for the ones that the user did not specify.
"""
if max_epochs is None:
if max_steps == -1 and not any(isinstance(cb, Timer) for cb in trainer.callbacks):
rank_zero_warn(
"`max_epochs` was not set. Setting it to 1000 epochs. To train without an epoch limit,"
" set `max_epochs=-1`.",
category=PossibleUserWarning,
)
max_epochs = 1000
else:
max_epochs = -1
if min_epochs is None and min_steps is not None:
# setting this allows FitLoop.done to re-evaluate should_stop when it gets triggered `on_fit_start`
min_epochs = 1
if min_epochs is None:
# the default value is 0 so no training will be done when should_stop is triggered `on_fit_start`
min_epochs = 0
return min_epochs, max_epochs
@contextmanager
def _block_parallel_sync_behavior(strategy: Strategy, block: bool = True) -> Generator[None, None, None]:
"""Blocks synchronization in :class:`~pytorch_lightning.strategies.parallel.ParallelStrategy`. This is useful for
example when accumulating gradients to reduce communication when it is not needed.
Args:
strategy: the strategy instance to use.
block: whether the context manager is enabled or not
Returns:
context manager with sync behaviour off
"""
if isinstance(strategy, ParallelStrategy) and block:
with strategy.block_backward_sync():
yield None
else:
yield None
def _is_max_limit_reached(current: int, maximum: int = -1) -> bool:
"""Check if the limit has been reached (if enabled).
Args:
current: the current value
maximum: the maximum value (or -1 to disable limit)
Returns:
bool: whether the limit has been reached
"""
return maximum != -1 and current >= maximum
def _reset_progress(loop: _Loop) -> None:
for v in vars(loop).values():
if isinstance(v, _BaseProgress):
v.reset()
elif isinstance(v, _Loop):
_reset_progress(v)
def _select_data_fetcher(trainer: "pl.Trainer", stage: RunningStage) -> _DataFetcher:
lightning_module = trainer.lightning_module
if stage == RunningStage.TESTING:
step_fx_name = "test_step"
elif stage == RunningStage.TRAINING:
step_fx_name = "training_step"
elif stage in (RunningStage.VALIDATING, RunningStage.SANITY_CHECKING):
step_fx_name = "validation_step"
elif stage == RunningStage.PREDICTING:
step_fx_name = "predict_step"
else:
raise RuntimeError(f"DataFetcher is unsupported for {trainer.state.stage}")
step_fx = getattr(lightning_module, step_fx_name)
if is_param_in_hook_signature(step_fx, "dataloader_iter", explicit=True):
rank_zero_warn(
f"Found `dataloader_iter` argument in the `{step_fx_name}`. Note that the support for "
"this signature is experimental and the behavior is subject to change."
)
return _DataLoaderIterDataFetcher()
return _PrefetchDataFetcher()
def _no_grad_context(loop_run: Callable) -> Callable:
def _decorator(self: _Loop, *args: Any, **kwargs: Any) -> Any:
if not isinstance(self, _Loop):
raise TypeError(f"`{type(self).__name__}` needs to be a Loop.")
if not hasattr(self, "inference_mode"):
raise TypeError(f"`{type(self).__name__}.inference_mode` needs to be defined")
context_manager: type[AbstractContextManager]
if _distributed_is_initialized() and dist.get_backend() == "gloo":
# gloo backend does not work properly.
# https://github.com/Lightning-AI/pytorch-lightning/pull/12715/files#r854569110
# TODO: explore why and possibly open an issue in PyTorch repository
context_manager = torch.no_grad
elif isinstance(self.trainer.accelerator, XLAAccelerator):
context_manager = torch.no_grad
elif isinstance(self.trainer.strategy, FSDPStrategy):
# https://github.com/pytorch/pytorch/issues/95957
context_manager = torch.no_grad
elif self.inference_mode:
context_manager = torch.inference_mode
else:
context_manager = torch.no_grad
with context_manager():
return loop_run(self, *args, **kwargs)
return _decorator
def _verify_dataloader_idx_requirement(
hooks: tuple[str, ...], is_expected: bool, stage: RunningStage, pl_module: "pl.LightningModule"
) -> None:
for hook in hooks:
fx = getattr(pl_module, hook)
# this validation only works if "dataloader_idx" is used, no other names such as "dl_idx"
param_present = is_param_in_hook_signature(fx, "dataloader_idx")
if not is_expected:
if param_present:
params = inspect.signature(fx).parameters
if "dataloader_idx" in params and params["dataloader_idx"].default is inspect.Parameter.empty:
raise RuntimeError(
f"You provided only a single `{stage.dataloader_prefix}_dataloader`, but have included "
f"`dataloader_idx` in `{type(pl_module).__name__}.{hook}()`. Either remove the"
" argument or give it a default value i.e. `dataloader_idx=0`."
)
elif not param_present:
raise RuntimeError(
f"You provided multiple `{stage.dataloader_prefix}_dataloader`, but no `dataloader_idx`"
f" argument in `{type(pl_module).__name__}.{hook}()`. Try adding `dataloader_idx=0` to its"
" signature."
)
|