Source code for pytorch_lightning.loops.loop

# 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 inspect
from abc import ABC, abstractmethod
from typing import Any, Dict, Generic, Optional, Type, TypeVar, Union

from deprecate import void
from torchmetrics import Metric

import pytorch_lightning as pl
from pytorch_lightning.trainer.connectors.logger_connector.result import _ResultCollection
from pytorch_lightning.trainer.progress import BaseProgress
from pytorch_lightning.utilities.exceptions import MisconfigurationException
from pytorch_lightning.utilities.imports import _fault_tolerant_training

T = TypeVar("T")  # the output type of `run`


class Loop(ABC, Generic[T]):
    """Basic Loops interface. All classes derived from this must implement the following properties and methods:

        * :attr:`done` (property): Condition to break the loop
        * :attr:`reset` (method): Resets the internal state between multiple calls of :attr:`run`
        * :attr:`advance` (method): Implements one step of the loop

    This class implements the following loop structure:

    .. code-block:: python

        on_run_start()

        while not done:
            on_advance_start()
            advance()
            on_advance_end()

        on_run_end()
    """

    def __init__(self) -> None:
        self._restarting = False
        self._trainer: Optional["pl.Trainer"] = None

    @property
    def trainer(self) -> "pl.Trainer":
        if self._trainer is None:
            raise RuntimeError("The loop is not attached to a Trainer.")
        return self._trainer

    @trainer.setter
    def trainer(self, trainer: "pl.Trainer") -> None:
        """Connects this loop's trainer and its children."""
        self._trainer = trainer
        for v in self.__dict__.values():
            if isinstance(v, Loop):
                v.trainer = trainer

    @property
    def restarting(self) -> bool:
        """Whether the state of this loop was reloaded and it needs to restart."""
        return self._restarting

    @restarting.setter
    def restarting(self, restarting: bool) -> None:
        """Connects this loop's restarting value and its children."""
        self._restarting = restarting
        for loop in vars(self).values():
            if isinstance(loop, Loop):
                loop.restarting = restarting

    @property
    @abstractmethod
    def done(self) -> bool:
        """Property indicating when the loop is finished.

        Example::

            @property
            def done(self):
                return self.trainer.global_step >= self.trainer.max_steps
        """

    @property
    def skip(self) -> bool:
        """Determine whether to return immediately from the call to :meth:`run`.

        Example::

            @property
            def skip(self):
                return len(self.trainer.train_dataloader) == 0
        """
        return False

    def connect(self, **kwargs: "Loop") -> None:
        """Optionally connect one or multiple loops to this one.

        Linked loops should form a tree.
        """

    def replace(self, **loops: Union["Loop", Type["Loop"]]) -> None:
        """Optionally replace one or multiple of this loop's sub-loops.

        This method takes care of instantiating the class (if necessary) with all existing arguments, connecting all
        sub-loops of the old loop to the new instance, setting the ``Trainer`` reference, and connecting the new loop to
        the parent.

        Args:
            **loops: ``Loop`` subclasses or instances. The name used should match the loop attribute name you want to
                replace.

        Raises:
            MisconfigurationException: When passing a ``Loop`` class, if the ``__init__`` arguments do not match those
                of the Loop class it replaces.
        """
        new_loops = {}

        for name, type_or_object in loops.items():
            old_loop = getattr(self, name)

            if isinstance(type_or_object, type):
                # compare the signatures
                old_parameters = inspect.signature(old_loop.__class__.__init__).parameters
                current_parameters = inspect.signature(type_or_object.__init__).parameters
                if old_parameters != current_parameters:
                    raise MisconfigurationException(
                        f"`{self.__class__.__name__}.replace({type_or_object.__name__})` can only be used if the"
                        f" `__init__` signatures match but `{old_loop.__class__.__name__}` does not."
                    )
                # instantiate the loop
                kwargs = {p: getattr(old_loop, p) for p in old_parameters if p != "self"}
                loop = type_or_object(**kwargs)
            else:
                loop = type_or_object

            # connect sub-loops
            kwargs = {n: l for n, l in old_loop.__dict__.items() if isinstance(l, Loop)}
            loop.connect(**kwargs)
            # set the trainer reference
            loop.trainer = self.trainer

            new_loops[name] = loop
        # connect to self
        self.connect(**new_loops)

    def on_skip(self) -> T:
        """The function to run when :meth:`run` should be skipped, determined by the condition in :attr:`skip`.

        Returns:
            the default output value of :meth:`on_run_end`
        """

[docs]    def run(self, *args: Any, **kwargs: Any) -> T:
        """The main entry point to the loop.

        Will frequently check the :attr:`done` condition and calls :attr:`advance`
        until :attr:`done` evaluates to ``True``.

        Override this if you wish to change the default behavior. The default implementation is:

        Example::

            def run(self, *args, **kwargs):
                if self.skip:
                    return self.on_skip()

                self.reset()
                self.on_run_start(*args, **kwargs)

                while not self.done:
                    self.advance(*args, **kwargs)

                output = self.on_run_end()
                return output

        Returns:
            The output of :attr:`on_run_end` (often outputs collected from each step of the loop)
        """
        if self.skip:
            return self.on_skip()

        self.reset()

        self.on_run_start(*args, **kwargs)

        while not self.done:
            try:
                self.on_advance_start(*args, **kwargs)
                self.advance(*args, **kwargs)
                self.on_advance_end()
                self._restarting = False
            except StopIteration:
                break
        self._restarting = False

        output = self.on_run_end()
        return output

[docs]    @abstractmethod
    def reset(self) -> None:
        """Resets the internal state of the loop at the beginning of each call to :attr:`run`.

        Example::

            def reset(self):
                # reset your internal state or add custom logic
                # if you expect run() to be called multiple times
                self.current_iteration = 0
                self.outputs = []
        """

    def on_run_start(self, *args: Any, **kwargs: Any) -> None:
        """Hook to be called as the first thing after entering :attr:`run` (except the state reset).

        Accepts all arguments passed to :attr:`run`.
        """
        void(*args, **kwargs)

    def on_advance_start(self, *args: Any, **kwargs: Any) -> None:
        """Hook to be called each time before :attr:`advance` is called.

        Accepts all arguments passed to :attr`run`.
        """
        void(*args, **kwargs)

[docs]    @abstractmethod
    def advance(self, *args: Any, **kwargs: Any) -> None:
        """Performs a single step.

        Accepts all arguments passed to :attr:`run`.

        Example::

            def advance(self, iterator):
                batch = next(iterator)
                loss = self.trainer.lightning_module.training_step(batch, batch_idx)
                ...
        """

    def on_advance_end(self) -> None:
        """Hook to be called each time after :attr:`advance` is called."""

    def on_run_end(self) -> T:
        """Hook to be called at the end of the run.

        Its return argument is returned from :attr:`run`.
        """

    def teardown(self) -> None:
        """Use to release memory etc."""

    def on_save_checkpoint(self) -> Dict:
        """Called when saving a model checkpoint, use to persist loop state.

        Returns:
            The current loop state.
        """
        return {}

    def on_load_checkpoint(self, state_dict: Dict) -> None:
        """Called when loading a model checkpoint, use to reload loop state."""

    def state_dict(self, destination: Optional[Dict] = None, prefix: str = "") -> Dict:
        """The state dict is determined by the state and progress of this loop and all its children.

        Args:
            destination: An existing dictionary to update with this loop's state. By default a new dictionary
                is returned.
            prefix: A prefix for each key in the state dictionary
        """
        if destination is None:
            destination = {}

        destination[prefix + "state_dict"] = self.on_save_checkpoint()

        # do not get the mode from `self.trainer` because it might not have been attached yet
        ft_enabled = _fault_tolerant_training()
        for k, v in self.__dict__.items():
            key = prefix + k
            if isinstance(v, BaseProgress):
                destination[key] = v.state_dict()
            elif isinstance(v, Loop):
                v.state_dict(destination, key + ".")
            elif ft_enabled and isinstance(v, _ResultCollection):
                # sync / unsync metrics
                v.sync()
                destination[key] = v.state_dict()
                v.unsync()

        return destination

    def load_state_dict(
        self,
        state_dict: Dict,
        prefix: str = "",
        metrics: Optional[Dict[str, Metric]] = None,
    ) -> None:
        """Loads the state of this loop and all its children."""
        self._load_from_state_dict(state_dict.copy(), prefix, metrics)
        for k, v in self.__dict__.items():
            if isinstance(v, Loop):
                v.load_state_dict(state_dict.copy(), prefix + k + ".")
        self.restarting = True

    def _load_from_state_dict(self, state_dict: Dict, prefix: str, metrics: Optional[Dict[str, Metric]] = None) -> None:
        trainer = self._trainer
        for k, v in self.__dict__.items():
            key = prefix + k
            if key not in state_dict:
                # compatibility with old checkpoints
                continue

            if isinstance(v, BaseProgress):
                v.load_state_dict(state_dict[key])
            elif isinstance(v, _ResultCollection) and trainer is not None and trainer.lightning_module is not None:
                metric_attributes = {
                    name: module
                    for name, module in self.trainer.lightning_module.named_modules()
                    if isinstance(module, Metric)
                }
                if metrics:
                    metric_attributes.update(metrics)

                # The `_ResultCollection` objects have 2 types of metrics: `Tensor` and `torchmetrics.Metric`.
                # When creating a checkpoint, the `Metric`s are dropped from the loop `state_dict` to serialize only
                # Python primitives. However, their states are saved with the model's `state_dict`.
                # On reload, we need to re-attach the `Metric`s back to the `_ResultCollection`.
                # The references are provided through the `metric_attributes` dictionary.
                v.load_state_dict(state_dict[key], metrics=metric_attributes, sync_fn=self.trainer.strategy.reduce)

                if not self.trainer.is_global_zero:
                    v.reset(metrics=False)

        if prefix + "state_dict" in state_dict:  # compatibility with old checkpoints
            self.on_load_checkpoint(state_dict[prefix + "state_dict"])