Shortcuts

Source code for pytorch_lightning.plugins.training_type.training_type_plugin

# Copyright The PyTorch Lightning 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 contextlib
from abc import ABC, abstractmethod
from typing import Any, Callable, Dict, Generator, List, Mapping, Optional, Tuple, TypeVar, Union

import torch
from torch import Tensor
from torch.nn import Module
from torch.optim import Optimizer
from torch.optim.lr_scheduler import _LRScheduler
from torch.utils.data import DataLoader

import pytorch_lightning as pl
from pytorch_lightning.overrides.base import unwrap_lightning_module
from pytorch_lightning.plugins import TorchCheckpointIO
from pytorch_lightning.plugins.io.checkpoint_plugin import CheckpointIO
from pytorch_lightning.plugins.precision import PrecisionPlugin
from pytorch_lightning.trainer.states import TrainerFn
from pytorch_lightning.utilities import rank_zero_deprecation
from pytorch_lightning.utilities.apply_func import apply_to_collection, move_data_to_device
from pytorch_lightning.utilities.distributed import ReduceOp
from pytorch_lightning.utilities.model_helpers import is_overridden
from pytorch_lightning.utilities.types import _PATH, STEP_OUTPUT

TBroadcast = TypeVar("TBroadcast")


[docs]class TrainingTypePlugin(ABC): """Base class for all training type plugins that change the behaviour of the training, validation and test- loop.""" def __init__( self, checkpoint_io: Optional[CheckpointIO] = None, precision_plugin: Optional[PrecisionPlugin] = None ) -> None: self._model: Optional[Module] = None checkpoint_io = checkpoint_io if checkpoint_io is not None else TorchCheckpointIO() self._checkpoint_io = checkpoint_io self._precision_plugin = precision_plugin if precision_plugin is not None else PrecisionPlugin() self.optimizers: List[Optimizer] = [] self.lr_schedulers: List[_LRScheduler] = [] self.optimizer_frequencies: List[int] = [] if is_overridden("post_dispatch", self, parent=TrainingTypePlugin): rank_zero_deprecation( f"`{self.__class__.__name__}.post_dispatch()` has been deprecated in v1.6 and will be removed in v1.7." f" Move your implementation to `{self.__class__.__name__}.teardown()` instead." ) @property def checkpoint_io(self) -> CheckpointIO: return self._checkpoint_io @property def precision_plugin(self) -> PrecisionPlugin: return self._precision_plugin @checkpoint_io.setter def checkpoint_io(self, plugin: CheckpointIO) -> None: self._checkpoint_io = plugin
[docs] def connect(self, model: Module) -> None: """Called by the accelerator to connect the accelerator and the model with this plugin.""" self.model = model
[docs] def setup_environment(self) -> None: """Setup any processes or distributed connections. This is called before the LightningModule/DataModule setup hook which allows the user to access the accelerator environment before setup is complete. """
[docs] def setup_optimizers(self, trainer: "pl.Trainer") -> None: """Creates optimizers and schedulers. Args: trainer: the Trainer, these optimizers should be connected to """ if trainer.state.fn not in (TrainerFn.FITTING, TrainerFn.TUNING): return optimizers, lr_schedulers, optimizer_frequencies = self.init_optimizers( trainer=trainer, model=self.lightning_module ) self.optimizers = optimizers self.lr_schedulers = lr_schedulers self.optimizer_frequencies = optimizer_frequencies
[docs] def setup(self, trainer: "pl.Trainer") -> None: """Setup plugins for the trainer fit and creates optimizers. Args: trainer: the trainer instance """ self.setup_optimizers(trainer) self.setup_precision_plugin()
[docs] def setup_precision_plugin(self) -> None: """Attaches the precision plugin to the accelerator.""" model, optimizers, schedulers = self.precision_plugin.connect(self.model, self.optimizers, self.lr_schedulers) self.model = model self.optimizers = optimizers self.lr_schedulers = schedulers
def _move_optimizer_state(self, device: Optional[torch.device] = None) -> None: """Moves the state of the optimizers to the appropriate device if needed.""" for opt in self.optimizers: for p, v in opt.state.items(): # `self.root_device` would raise error if called outside the spawn process # while training on 8 and more cores. opt.state[p] = apply_to_collection(v, torch.Tensor, move_data_to_device, device or self.root_device)
[docs] def optimizer_state(self, optimizer: Optimizer) -> Dict[str, Tensor]: """Returns state of an optimizer. Allows for syncing/collating optimizer state from processes in custom plugins. """ return optimizer.state_dict()
[docs] def backward(self, closure_loss: Tensor, *args: Any, **kwargs: Any) -> Tensor: """Forwards backward-calls to the precision plugin. Args: closure_loss: a tensor holding the loss value to backpropagate """ self.pre_backward(closure_loss) closure_loss = self.precision_plugin.pre_backward(self.lightning_module, closure_loss) self.precision_plugin.backward(self.lightning_module, closure_loss, *args, **kwargs) closure_loss = self.precision_plugin.post_backward(self.lightning_module, closure_loss) self.post_backward(closure_loss) return closure_loss
[docs] def optimizer_step( self, optimizer: Optimizer, opt_idx: int, closure: Callable[[], Any], model: Optional[Union["pl.LightningModule", Module]] = None, **kwargs: Any, ) -> None: """performs the actual optimizer step. Args: optimizer: the optimizer performing the step opt_idx: index of the current optimizer closure: closure calculating the loss value model: reference to the model, optionally defining optimizer step related hooks **kwargs: Any extra arguments to ``optimizer.step`` """ model = model or self.lightning_module self.precision_plugin.optimizer_step(model, optimizer, opt_idx, closure, **kwargs)
[docs] def optimizer_zero_grad(self, current_epoch: int, batch_idx: int, optimizer: Optimizer, opt_idx: int) -> None: """Zeros all model parameter's gradients.""" model_ref = self.lightning_module model_ref.optimizer_zero_grad(current_epoch, batch_idx, optimizer, opt_idx)
def _setup_model_and_optimizers(self, model: Module, optimizers: List[Optimizer]) -> Tuple[Module, List[Optimizer]]: """Setup a model and multiple optimizers together. The returned objects are expected to be in the same order they were passed in. The default implementation will call :meth:`_setup_model` and :meth:`_setup_optimizer` on the inputs. """ # TODO (@awaelchli): standardize this across all plugins in Lightning and Lite. Related refactor: #7324 model = self._setup_model(model) optimizers = [self._setup_optimizer(optimizer) for optimizer in optimizers] return model, optimizers def _setup_model(self, model: Module) -> Module: """Performs setup for the model, e.g., by wrapping it by another class.""" # TODO (@awaelchli): standardize this across all plugins in Lightning and Lite. Related refactor: #7324 return model def _setup_optimizer(self, optimizer: Optimizer) -> Optimizer: """Performs setup for the optimizer, e.g., by wrapping it by another class.""" # TODO (@awaelchli): standardize this across all plugins in Lightning and Lite. Related refactor: #7324 return optimizer
[docs] def batch_to_device(self, batch: Any, device: Optional[torch.device] = None, dataloader_idx: int = 0) -> Any: """Moves the batch to the correct device. The returned batch is of the same type as the input batch, just having all tensors on the correct device. Args: batch: The batch of samples to move to the correct device device: The target device dataloader_idx: The index of the dataloader to which the batch belongs. """ model = self.lightning_module device = device or self.root_device if model is not None: return model._apply_batch_transfer_handler(batch, device=device, dataloader_idx=dataloader_idx) return move_data_to_device(batch, device)
@property @abstractmethod def on_gpu(self) -> bool: """Returns whether the current process is done on GPU.""" @property @abstractmethod def on_tpu(self) -> bool: """Returns whether the current process is done on TPU.""" @property @abstractmethod def root_device(self) -> torch.device: """Returns the root device."""
[docs] @abstractmethod def model_to_device(self) -> None: """Moves the model to the correct device."""
@property @abstractmethod def is_global_zero(self) -> bool: """Whether the current process is the rank zero process not only on the local node, but for all nodes."""
[docs] @abstractmethod def reduce( self, tensor: Union[torch.Tensor, Any], group: Optional[Any] = None, reduce_op: Optional[Union[ReduceOp, str]] = "mean", ) -> Union[torch.Tensor, Any]: """Reduces the given tensor (e.g. across GPUs/processes). Args: tensor: the tensor to sync and reduce group: the process group to reduce reduce_op: the reduction operation. Defaults to 'mean'. Can also be a string 'sum' or ReduceOp. """
[docs] @abstractmethod def barrier(self, name: Optional[str] = None) -> None: """Synchronizes all processes which blocks processes until the whole group enters this function. Args: name: an optional name to pass into barrier. """
[docs] @abstractmethod def broadcast(self, obj: TBroadcast, src: int = 0) -> TBroadcast: """Broadcasts an object to all processes. Args: obj: the object to broadcast src: source rank """
[docs] @abstractmethod def all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None, sync_grads: bool = False) -> torch.Tensor: """Perform an all_gather on all processes. Args: tensor: the tensor to all_gather group: the process group to gather results from sync_grads: flag that allows users to synchronize gradients for all_gather op """
[docs] def reduce_boolean_decision(self, decision: bool) -> bool: """Reduce the early stopping decision across all processes.""" return decision
[docs] def pre_backward(self, closure_loss: torch.Tensor) -> None: """Run before precision plugin executes backward."""
[docs] def post_backward(self, closure_loss: torch.Tensor) -> None: """Run after precision plugin executes backward."""
@property def model(self) -> Optional[Module]: """Returns the potentially wrapped LightningModule.""" return self._model @model.setter def model(self, new_model: Optional[Module]) -> None: self._model = new_model @property def lightning_module(self) -> Optional["pl.LightningModule"]: """Returns the pure LightningModule without potential wrappers.""" return unwrap_lightning_module(self._model) if self._model is not None else None def load_checkpoint(self, checkpoint_path: _PATH) -> Dict[str, Any]: torch.cuda.empty_cache() return self.checkpoint_io.load_checkpoint(checkpoint_path) def load_model_state_dict(self, checkpoint: Mapping[str, Any]) -> None: self.lightning_module.load_state_dict(checkpoint["state_dict"]) def load_optimizer_state_dict(self, checkpoint: Mapping[str, Any]) -> None: optimizer_states = checkpoint["optimizer_states"] for optimizer, opt_state in zip(self.optimizers, optimizer_states): optimizer.load_state_dict(opt_state) def start_training(self, trainer: "pl.Trainer") -> Any: # double dispatch to initiate the training loop return trainer.run_stage() def start_evaluating(self, trainer: "pl.Trainer") -> Any: # double dispatch to initiate the test loop return trainer.run_stage() def start_predicting(self, trainer: "pl.Trainer") -> Any: # double dispatch to initiate the predicting loop return trainer.run_stage()
[docs] def training_step(self, *args, **kwargs) -> STEP_OUTPUT: """The actual training step. See :meth:`~pytorch_lightning.core.lightning.LightningModule.training_step` for more details """ with self.precision_plugin.train_step_context(): return self.model.training_step(*args, **kwargs)
def post_training_step(self): pass
[docs] def validation_step(self, *args, **kwargs) -> Optional[STEP_OUTPUT]: """The actual validation step. See :meth:`~pytorch_lightning.core.lightning.LightningModule.validation_step` for more details """ with self.precision_plugin.val_step_context(): return self.model.validation_step(*args, **kwargs)
[docs] def test_step(self, *args, **kwargs) -> Optional[STEP_OUTPUT]: """The actual test step. See :meth:`~pytorch_lightning.core.lightning.LightningModule.test_step` for more details """ with self.precision_plugin.test_step_context(): return self.model.test_step(*args, **kwargs)
[docs] def predict_step(self, *args, **kwargs) -> STEP_OUTPUT: """The actual predict step. See :meth:`~pytorch_lightning.core.lightning.LightningModule.predict_step` for more details """ with self.precision_plugin.predict_step_context(): return self.model.predict_step(*args, **kwargs)
def training_step_end(self, output): return output def validation_step_end(self, output): return output def test_step_end(self, output): return output
[docs] def process_dataloader(self, dataloader: DataLoader) -> DataLoader: """Wraps the dataloader if necessary. Args: dataloader: iterable. Ideally of type: :class:`torch.utils.data.DataLoader` """ return dataloader
def init_optimizers(self, trainer: "pl.Trainer", model: "pl.LightningModule"): return trainer.init_optimizers(model) @property def restore_checkpoint_after_pre_dispatch(self) -> bool: """Override to delay restoring from checkpoint till after pre-dispatch. This is useful when the plugin requires all the setup hooks to run before loading checkpoint. Returns: If true, restore checkpoint after pre_dispatch. """ return False @property def lightning_restore_optimizer_and_schedulers(self) -> bool: """Override to disable Lightning restoring optimizers/schedulers. This is useful for plugins which manage restoring optimizers/schedulers. """ return True @property def handles_gradient_accumulation(self) -> bool: """Whether the plugin handles gradient accumulation internally.""" return False
[docs] def lightning_module_state_dict(self) -> Dict[str, Union[Any, Tensor]]: """Returns model state.""" model = self.lightning_module return model.state_dict()
[docs] def save_checkpoint(self, checkpoint: Dict[str, Any], filepath: _PATH) -> None: """Save model/training states as a checkpoint file through state-dump and file-write. Args: checkpoint: dict containing model and trainer state filepath: write-target file's path """ if self.should_rank_save_checkpoint: return self.checkpoint_io.save_checkpoint(checkpoint, filepath)
[docs] def remove_checkpoint(self, filepath: _PATH) -> None: """Remove checkpoint filepath from the filesystem. Args: filepath: Path to checkpoint """ if self.should_rank_save_checkpoint: return self.checkpoint_io.remove_checkpoint(filepath)
[docs] @contextlib.contextmanager def model_sharded_context(self) -> Generator: """Provide hook to create modules in a distributed aware context. This is useful for when we'd like to shard the model instantly, which is useful for extremely large models which can save memory and initialization time. Returns: Model parallel context. """ yield
[docs] @abstractmethod def teardown(self) -> None: """This method is called to teardown the training process. It is the right place to release memory and free other resources. """
@classmethod def register_plugins(cls, plugin_registry) -> None: pass @property def should_rank_save_checkpoint(self) -> bool: """Returns whether the checkpoint should be saved (rank based)""" return self.is_global_zero
[docs] def on_train_start(self) -> None: """Called when train begins.""" pass
[docs] def on_validation_start(self) -> None: """Called when validation begins.""" pass
[docs] def on_test_start(self) -> None: """Called when test begins.""" pass
[docs] def on_predict_start(self) -> None: """Called when predict begins.""" pass
[docs] def on_train_end(self) -> None: """Called when train ends.""" pass
[docs] def on_validation_end(self) -> None: """Called when validation ends.""" pass
[docs] def on_test_end(self) -> None: """Called when test end.""" pass
[docs] def on_predict_end(self): """Called when predict ends.""" pass
[docs] def on_train_batch_start(self, batch: Any, batch_idx: int, dataloader_idx: int = 0) -> None: """Called in the training loop before anything happens for that batch.""" pass
[docs] def pre_dispatch(self, trainer: "pl.Trainer") -> None: """Hook to do something before the training/evaluation/prediction starts.""" self._move_optimizer_state()
[docs] def dispatch(self, trainer: "pl.Trainer") -> None: """Hook to do something before the training/evaluation/prediction starts.""" self.precision_plugin.dispatch(trainer)
[docs] def post_dispatch(self, trainer: "pl.Trainer") -> None: r""" .. deprecated:: v1.6 This method has been deprecated in v1.6 and will be removed in v1.7. Use :meth:`teardown` instead. Hook to do something after the training/evaluation/prediction finishes. """

© Copyright Copyright (c) 2018-2021, William Falcon et al... Revision 46f718d2.

Built with Sphinx using a theme provided by Read the Docs.
Read the Docs v: latest
Versions
latest
stable
1.5.4
1.5.3
1.5.2
1.5.1
1.5.0
1.4.9
1.4.8
1.4.7
1.4.6
1.4.5
1.4.4
1.4.3
1.4.2
1.4.1
1.4.0
1.3.8
1.3.7
1.3.6
1.3.5
1.3.4
1.3.3
1.3.2
1.3.1
1.3.0
1.2.10
1.2.8
1.2.7
1.2.6
1.2.5
1.2.4
1.2.3
1.2.2
1.2.1
1.2.0
1.1.8
1.1.7
1.1.6
1.1.5
1.1.4
1.1.3
1.1.2
1.1.1
1.1.0
1.0.8
1.0.7
1.0.6
1.0.5
1.0.4
1.0.3
1.0.2
1.0.1
1.0.0
0.10.0
0.9.0
0.8.5
0.8.4
0.8.3
0.8.2
0.8.1
0.8.0
0.7.6
0.7.5
0.7.4
0.7.3
0.7.2
0.7.1
0.7.0
0.6.0
0.5.3
0.4.9
ipynb-update
docs-search
Downloads
html
On Read the Docs
Project Home
Builds

Free document hosting provided by Read the Docs.