import collections.abc as collections import numbers import os import stat import tempfile from abc import ABCMeta, abstractmethod from collections import OrderedDict from pathlib import Path from typing import Any, Callable, cast, Dict, List, Mapping, NamedTuple, Optional, Union import torch import torch.nn as nn from packaging.version import Version if Version(torch.__version__) >= Version("1.9.0"): from torch.distributed.optim import ZeroRedundancyOptimizer HAVE_ZERO = True else: HAVE_ZERO = False import ignite.distributed as idist from ignite.base import Serializable from ignite.engine import Engine, Events from ignite.utils import _tree_apply2, _tree_map __all__ = ["Checkpoint", "DiskSaver", "ModelCheckpoint", "BaseSaveHandler"] class BaseSaveHandler(metaclass=ABCMeta): """Base class for save handlers Methods to override: - :meth:`~ignite.handlers.checkpoint.BaseSaveHandler.__call__` - :meth:`~ignite.handlers.checkpoint.BaseSaveHandler.remove` Note: In derived class, please, make sure that in distributed configuration overridden methods are called by a single process. Distributed configuration on XLA devices should be treated slightly differently: for saving checkpoint with `xm.save() `_ all processes should pass into the function. Otherwise, application gets stuck. """ @abstractmethod def __call__(self, checkpoint: Mapping, filename: str, metadata: Optional[Mapping] = None) -> None: """Method to save `checkpoint` with `filename`. Additionally, metadata dictionary is provided. Metadata contains: - `basename`: file prefix (if provided) with checkpoint name, e.g. `epoch_checkpoint`. - `score_name`: score name if provided, e.g `val_acc`. - `priority`: checkpoint priority value (higher is better), e.g. `12` or `0.6554435` Args: checkpoint: checkpoint dictionary to save. filename: filename associated with checkpoint. metadata: metadata on checkpoint to save. """ @abstractmethod def remove(self, filename: str) -> None: """Method to remove saved checkpoint. Args: filename: filename associated with checkpoint. """ class Checkpoint(Serializable): """Checkpoint handler can be used to periodically save and load objects which have attribute ``state_dict/load_state_dict``. This class can use specific save handlers to store on the disk or a cloud storage, etc. The Checkpoint handler (if used with :class:`~ignite.handlers.DiskSaver`) also handles automatically moving data on TPU to CPU before writing the checkpoint. Args: to_save: Dictionary with the objects to save. Objects should have implemented ``state_dict`` and ``load_state_dict`` methods. If contains objects of type torch `DistributedDataParallel`_ or `DataParallel`_, their internal wrapped model is automatically saved (to avoid additional key ``module.`` in the state dictionary). save_handler: String, function or callable object. used to save engine and other provided objects. Function receives two objects: checkpoint as a dictionary and filename. If ``save_handler`` is callable class, it can inherit of :class:`~ignite.handlers.checkpoint.BaseSaveHandler` and optionally implement ``remove`` method to keep a fixed number of saved checkpoints. In case if user needs to save engine's checkpoint on a disk, ``save_handler`` can be defined with :class:`~ignite.handlers.DiskSaver` or a string specifying directory name can be passed to ``save_handler``. filename_prefix: Prefix for the file name to which objects will be saved. See Note for details. score_function: If not None, it should be a function taking a single argument, :class:`~ignite.engine.engine.Engine` object, and returning a score (`float`). Objects with highest scores will be retained. score_name: If ``score_function`` not None, it is possible to store its value using ``score_name``. If ``score_function`` is None, ``score_name`` can be used alone to define ``score_function`` as ``Checkpoint.get_default_score_fn(score_name)`` by default. n_saved: Number of objects that should be kept on disk. Older files will be removed. If set to `None`, all objects are kept. global_step_transform: global step transform function to output a desired global step. Input of the function is ``(engine, event_name)``. Output of function should be an integer. Default is None, global_step based on attached engine. If provided, uses function output as global_step. To setup global step from another engine, please use :meth:`~ignite.handlers.global_step_from_engine`. filename_pattern: If ``filename_pattern`` is provided, this pattern will be used to render checkpoint filenames. If the pattern is not defined, the default pattern would be used. See Note for details. include_self: Whether to include the `state_dict` of this object in the checkpoint. If `True`, then there must not be another object in ``to_save`` with key ``checkpointer``. greater_or_equal: if `True`, the latest equally scored model is stored. Otherwise, the first model. Default, `False`. save_on_rank: Which rank to save the objects on, in the distributed configuration. If ``save_handler`` is string or :class:`~pathlib.Path`, this is also used to instantiate a :class:`~ignite.handlers.DiskSaver`. .. _DistributedDataParallel: https://pytorch.org/docs/stable/generated/ torch.nn.parallel.DistributedDataParallel.html .. _DataParallel: https://pytorch.org/docs/stable/generated/torch.nn.DataParallel.html Note: This class stores a single file as a dictionary of provided objects to save. The filename is defined by ``filename_pattern`` and by default has the following structure: ``{filename_prefix}_{name}_{suffix}.{ext}`` where - ``filename_prefix`` is the argument passed to the constructor, - `name` is the key in ``to_save`` if a single object is to store, otherwise `name` is "checkpoint". - `suffix` is composed as following ``{global_step}_{score_name}={score}``. +----------------+------------+-----------------------+----------------------------------------------+ | score_function | score_name | global_step_transform | suffix | +================+============+=======================+==============================================+ | None | None | None | ``{engine.state.iteration}`` | +----------------+------------+-----------------------+----------------------------------------------+ | X | None | None | ``{score}`` | +----------------+------------+-----------------------+----------------------------------------------+ | X | None | X | ``{global_step}_{score}`` | +----------------+------------+-----------------------+----------------------------------------------+ | X | X | X | ``{global_step}_{score_name}={score}`` | +----------------+------------+-----------------------+----------------------------------------------+ | None | None | X | ``{global_step}`` | +----------------+------------+-----------------------+----------------------------------------------+ | X | X | None | ``{score_name}={score}`` | +----------------+------------+-----------------------+----------------------------------------------+ Above `global_step` defined by the output of `global_step_transform` and `score` defined by the output of `score_function`. By default, none of ``score_function``, ``score_name``, ``global_step_transform`` is defined, then suffix is setup by attached engine's current iteration. The filename will be `{filename_prefix}_{name}_{engine.state.iteration}.{ext}`. For example, ``score_name="neg_val_loss"`` and ``score_function`` that returns `-loss` (as objects with highest scores will be retained), then saved filename will be ``{filename_prefix}_{name}_neg_val_loss=-0.1234.pt``. Note: If ``filename_pattern`` is given, it will be used to render the filenames. ``filename_pattern`` is a string that can contain ``{filename_prefix}``, ``{name}``, ``{score}``, ``{score_name}`` and ``{global_step}`` as templates. For example, let ``filename_pattern="{global_step}-{name}-{score}.pt"`` then the saved filename will be ``30000-checkpoint-94.pt`` **Warning:** Please, keep in mind that if filename collide with already used one to saved a checkpoint, new checkpoint will replace the older one. This means that filename like ``checkpoint.pt`` will be saved every call and will always be overwritten by newer checkpoints. Note: To get the last stored filename, handler exposes attribute ``last_checkpoint``: .. code-block:: python handler = Checkpoint(...) ... print(handler.last_checkpoint) > checkpoint_12345.pt Note: This class is distributed configuration-friendly: it is not required to instantiate the class in rank 0 only process. This class supports automatically distributed configuration and if used with :class:`~ignite.handlers.DiskSaver`, checkpoint is stored by rank 0 process. .. warning:: When running on XLA devices or using :class:`~torch.distributed.optim.ZeroRedundancyOptimizer`, it should be run in all processes, otherwise application can get stuck while saving the checkpoint. .. code-block:: python # Wrong: # if idist.get_rank() == 0: # handler = Checkpoint(...) # trainer.add_event_handler(Events.ITERATION_COMPLETED(every=1000), handler) # Correct: handler = Checkpoint(...) trainer.add_event_handler(Events.ITERATION_COMPLETED(every=1000), handler) Examples: Attach the handler to make checkpoints during training: .. code-block:: python from ignite.engine import Engine, Events from ignite.handlers import Checkpoint trainer = ... model = ... optimizer = ... lr_scheduler = ... to_save = {'model': model, 'optimizer': optimizer, 'lr_scheduler': lr_scheduler, 'trainer': trainer} if (checkpoint_iters): # A: Output is "checkpoint_.pt" handler = Checkpoint( to_save, '/tmp/models', n_saved=2 ) trainer.add_event_handler(Events.ITERATION_COMPLETED(every=1000), handler) else: # B:Output is "checkpoint_.pt" gst = lambda *_: trainer.state.epoch handler = Checkpoint( to_save, '/tmp/models', n_saved=2, global_step_transform=gst ) trainer.add_event_handler(Events.EPOCH_COMPLETED, handler) trainer.run(data_loader, max_epochs=6) > A: ["checkpoint_7000.pt", "checkpoint_8000.pt", ] > B: ["checkpoint_5.pt", "checkpoint_6.pt", ] Attach the handler to an evaluator to save best model during the training according to computed validation metric: .. code-block:: python from ignite.engine import Engine, Events from ignite.handlers import Checkpoint, global_step_from_engine trainer = ... evaluator = ... # Setup Accuracy metric computation on evaluator. # evaluator.state.metrics contain 'accuracy', # which will be used to define ``score_function`` automatically. # Run evaluation on epoch completed event # ... to_save = {'model': model} handler = Checkpoint( to_save, '/tmp/models', n_saved=2, filename_prefix='best', score_name="accuracy", global_step_transform=global_step_from_engine(trainer) ) evaluator.add_event_handler(Events.COMPLETED, handler) trainer.run(data_loader, max_epochs=10) > ["best_model_9_accuracy=0.77.pt", "best_model_10_accuracy=0.78.pt", ] Customise the ``save_handler``: .. code-block:: python handler = Checkpoint( to_save, save_handler=DiskSaver('/tmp/models', create_dir=True, **kwargs), n_saved=2 ) .. versionchanged:: 0.4.3 - Checkpoint can save model with same filename. - Added ``greater_or_equal`` argument. .. versionchanged:: 0.4.7 - `score_name` can be used to define `score_function` automatically without providing `score_function`. - `save_handler` automatically saves to disk if path to directory is provided. - `save_on_rank` saves objects on this rank in a distributed configuration. """ Item = NamedTuple("Item", [("priority", int), ("filename", str)]) _state_dict_all_req_keys = ("_saved",) def __init__( self, to_save: Mapping, save_handler: Union[str, Path, Callable, BaseSaveHandler], filename_prefix: str = "", score_function: Optional[Callable] = None, score_name: Optional[str] = None, n_saved: Union[int, None] = 1, global_step_transform: Optional[Callable] = None, filename_pattern: Optional[str] = None, include_self: bool = False, greater_or_equal: bool = False, save_on_rank: int = 0, ): if not isinstance(to_save, collections.Mapping): raise TypeError(f"Argument `to_save` should be a dictionary, but given {type(to_save)}") self._check_objects(to_save, "state_dict") if include_self: if not isinstance(to_save, collections.MutableMapping): raise TypeError( f"If `include_self` is True, then `to_save` must be mutable, but given {type(to_save)}." ) if "checkpointer" in to_save: raise ValueError(f"Cannot have key 'checkpointer' if `include_self` is True: {to_save}") if not ( isinstance(save_handler, str) or isinstance(save_handler, Path) or callable(save_handler) or isinstance(save_handler, BaseSaveHandler) ): raise TypeError( "Argument `save_handler` should be a string or Path object or callable or inherit from BaseSaveHandler" ) if global_step_transform is not None and not callable(global_step_transform): raise TypeError(f"global_step_transform should be a function, got {type(global_step_transform)} instead.") self.to_save = to_save self.filename_prefix = filename_prefix if isinstance(save_handler, str) or isinstance(save_handler, Path): self.save_handler = DiskSaver(save_handler, create_dir=True, save_on_rank=save_on_rank) else: self.save_handler = save_handler # type: ignore self.score_function = score_function self.score_name = score_name if self.score_name is not None and self.score_function is None: self.score_function = self.get_default_score_fn(self.score_name) self.n_saved = n_saved self.ext = "pt" self.global_step_transform = global_step_transform self.filename_pattern = filename_pattern self._saved: List["Checkpoint.Item"] = [] self.include_self = include_self self.greater_or_equal = greater_or_equal self.save_on_rank = save_on_rank def _get_filename_pattern(self, global_step: Optional[int]) -> str: if self.filename_pattern is None: filename_pattern = self.setup_filename_pattern( with_prefix=len(self.filename_prefix) > 0, with_score=self.score_function is not None, with_score_name=self.score_name is not None, with_global_step=global_step is not None, ) else: filename_pattern = self.filename_pattern return filename_pattern def reset(self) -> None: """Method to reset saved checkpoint names. Use this method if the engine will independently run multiple times: .. code-block:: python from ignite.handlers import Checkpoint trainer = ... checkpointer = Checkpoint(...) trainer.add_event_handler(Events.COMPLETED, checkpointer) trainer.add_event_handler(Events.STARTED, checkpointer.reset) # fold 0 trainer.run(data0, max_epochs=max_epochs) print("Last checkpoint:", checkpointer.last_checkpoint) # fold 1 trainer.run(data1, max_epochs=max_epochs) print("Last checkpoint:", checkpointer.last_checkpoint) .. versionadded:: 0.4.3 """ self._saved = [] @property def last_checkpoint(self) -> Optional[Union[str, Path]]: if len(self._saved) < 1: return None if not isinstance(self.save_handler, DiskSaver): return self._saved[-1].filename return self.save_handler.dirname / self._saved[-1].filename def _check_lt_n_saved(self, or_equal: bool = False) -> bool: if self.n_saved is None: return True return len(self._saved) < self.n_saved + int(or_equal) def _compare_fn(self, new: Union[int, float]) -> bool: if self.greater_or_equal: return new >= self._saved[0].priority else: return new > self._saved[0].priority def __call__(self, engine: Engine) -> None: global_step = None if self.global_step_transform is not None: global_step = self.global_step_transform(engine, engine.last_event_name) if self.score_function is not None: priority = self.score_function(engine) if not isinstance(priority, numbers.Number): raise ValueError("Output of score_function should be a number") else: if global_step is None: global_step = engine.state.get_event_attrib_value(Events.ITERATION_COMPLETED) priority = global_step if self._check_lt_n_saved() or self._compare_fn(priority): priority_str = f"{priority}" if isinstance(priority, numbers.Integral) else f"{priority:.4f}" checkpoint = self._setup_checkpoint() name = "checkpoint" if len(checkpoint) == 1: for k in checkpoint: name = k checkpoint = checkpoint[name] filename_pattern = self._get_filename_pattern(global_step) filename_dict = { "filename_prefix": self.filename_prefix, "ext": self.ext, "name": name, "score_name": self.score_name, "score": priority_str if (self.score_function is not None) else None, "global_step": global_step, } filename = filename_pattern.format(**filename_dict) metadata = { "basename": f"{self.filename_prefix}{'_' * int(len(self.filename_prefix) > 0)}{name}", "score_name": self.score_name, "priority": priority, } try: index = list(map(lambda it: it.filename == filename, self._saved)).index(True) to_remove = True except ValueError: index = 0 to_remove = not self._check_lt_n_saved() if to_remove: item = self._saved.pop(index) if isinstance(self.save_handler, BaseSaveHandler): self.save_handler.remove(item.filename) self._saved.append(Checkpoint.Item(priority, filename)) self._saved.sort(key=lambda it: it[0]) if self.include_self: # Now that we've updated _saved, we can add our own state_dict. checkpoint["checkpointer"] = self.state_dict() try: self.save_handler(checkpoint, filename, metadata) except TypeError: self.save_handler(checkpoint, filename) def _setup_checkpoint(self) -> Dict[str, Any]: if self.to_save is not None: def func(obj: Any, **kwargs: Any) -> Dict: if isinstance(obj, (nn.DataParallel, nn.parallel.DistributedDataParallel)): obj = obj.module elif HAVE_ZERO and isinstance(obj, ZeroRedundancyOptimizer): obj.consolidate_state_dict(to=self.save_on_rank) if self.save_on_rank != idist.get_rank(): return {} return obj.state_dict() return cast(Dict[str, Any], _tree_map(func, self.to_save)) return {} @staticmethod def setup_filename_pattern( with_prefix: bool = True, with_score: bool = True, with_score_name: bool = True, with_global_step: bool = True ) -> str: """Helper method to get the default filename pattern for a checkpoint. Args: with_prefix: If True, the ``filename_prefix`` is added to the filename pattern: ``{filename_prefix}_{name}...``. Default, True. with_score: If True, ``score`` is added to the filename pattern: ``..._{score}.{ext}``. Default, True. At least one of ``with_score`` and ``with_global_step`` should be True. with_score_name: If True, ``score_name`` is added to the filename pattern: ``..._{score_name}={score}.{ext}``. If activated, argument ``with_score`` should be also True, otherwise an error is raised. Default, True. with_global_step: If True, ``{global_step}`` is added to the filename pattern: ``...{name}_{global_step}...``. At least one of ``with_score`` and ``with_global_step`` should be True. Examples: .. code-block:: python from ignite.handlers import Checkpoint filename_pattern = Checkpoint.setup_filename_pattern() print(filename_pattern) > "{filename_prefix}_{name}_{global_step}_{score_name}={score}.{ext}" .. versionadded:: 0.4.3 """ filename_pattern = "{name}" if not (with_global_step or with_score): raise ValueError("At least one of with_score and with_global_step should be True.") if with_global_step: filename_pattern += "_{global_step}" if with_score_name and with_score: filename_pattern += "_{score_name}={score}" elif with_score: filename_pattern += "_{score}" elif with_score_name: raise ValueError("If with_score_name is True, with_score should be also True") if with_prefix: filename_pattern = "{filename_prefix}_" + filename_pattern filename_pattern += ".{ext}" return filename_pattern @staticmethod def _check_objects(objs: Mapping, attr: str) -> None: def func(obj: Any, **kwargs: Any) -> None: if not hasattr(obj, attr): raise TypeError(f"Object {type(obj)} should have `{attr}` method") _tree_map(func, objs) @staticmethod def load_objects(to_load: Mapping, checkpoint: Union[str, Mapping, Path], **kwargs: Any) -> None: """Helper method to apply ``load_state_dict`` on the objects from ``to_load`` using states from ``checkpoint``. Args: to_load: a dictionary with objects, e.g. `{"model": model, "optimizer": optimizer, ...}` checkpoint: a path, a string filepath or a dictionary with state_dicts to load, e.g. `{"model": model_state_dict, "optimizer": opt_state_dict}`. If `to_load` contains a single key, then checkpoint can contain directly corresponding state_dict. kwargs: Keyword arguments accepted for `nn.Module.load_state_dict()`. Passing `strict=False` enables the user to load part of the pretrained model (useful for example, in Transfer Learning) Examples: .. code-block:: python import tempfile from pathlib import Path import torch from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint, Checkpoint trainer = Engine(lambda engine, batch: None) with tempfile.TemporaryDirectory() as tmpdirname: handler = ModelCheckpoint(tmpdirname, 'myprefix', n_saved=None, create_dir=True) model = torch.nn.Linear(3, 3) optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) to_save = {"weights": model, "optimizer": optimizer} trainer.add_event_handler(Events.EPOCH_COMPLETED(every=2), handler, to_save) trainer.run(torch.randn(10, 1), 5) to_load = to_save checkpoint_fp = Path(tmpdirname) / 'myprefix_checkpoint_40.pt' checkpoint = torch.load(checkpoint_fp) Checkpoint.load_objects(to_load=to_load, checkpoint=checkpoint) # or using a string for checkpoint filepath to_load = to_save checkpoint_fp = Path(tmpdirname) / 'myprefix_checkpoint_40.pt' Checkpoint.load_objects(to_load=to_load, checkpoint=checkpoint_fp) Note: If ``to_load`` contains objects of type torch `DistributedDataParallel`_ or `DataParallel`_, method ``load_state_dict`` will applied to their internal wrapped model (``obj.module``). .. _DistributedDataParallel: https://pytorch.org/docs/stable/generated/ torch.nn.parallel.DistributedDataParallel.html .. _DataParallel: https://pytorch.org/docs/stable/generated/torch.nn.DataParallel.html """ if not isinstance(checkpoint, (collections.Mapping, str, Path)): raise TypeError(f"Argument checkpoint should be a string or a dictionary, but given {type(checkpoint)}") Checkpoint._check_objects(to_load, "load_state_dict") if isinstance(checkpoint, (str, Path)): checkpoint_obj = torch.load(checkpoint) else: checkpoint_obj = checkpoint def _load_object(obj: Any, chkpt_obj: Any) -> None: if isinstance(obj, (nn.DataParallel, nn.parallel.DistributedDataParallel)): obj = obj.module if isinstance(obj, torch.nn.Module): obj.load_state_dict(chkpt_obj, **kwargs) else: obj.load_state_dict(chkpt_obj) if len(to_load) == 1: # single object and checkpoint is directly a state_dict key, obj = list(to_load.items())[0] if key not in checkpoint_obj: _load_object(obj, checkpoint_obj) return _tree_apply2(_load_object, to_load, checkpoint_obj) def reload_objects(self, to_load: Mapping, load_kwargs: Optional[Dict] = None, **filename_components: Any) -> None: """Helper method to apply ``load_state_dict`` on the objects from ``to_load``. Filename components such as name, score and global state can be configured. Args: to_load: a dictionary with objects, e.g. `{"model": model, "optimizer": optimizer, ...}` load_kwargs: Keyword arguments accepted for `nn.Module.load_state_dict()`. Passing `strict=False` enables the user to load part of the pretrained model (useful for example, in Transfer Learning) filename_components: Filename components used to define the checkpoint file path. Keyword arguments accepted are `name`, `score` and `global_state`. Examples: .. code-block:: python import tempfile import torch from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint trainer = Engine(lambda engine, batch: None) with tempfile.TemporaryDirectory() as tmpdirname: checkpoint = ModelCheckpoint(tmpdirname, 'myprefix', n_saved=None, create_dir=True) model = torch.nn.Linear(3, 3) optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) to_save = {"weights": model, "optimizer": optimizer} trainer.add_event_handler(Events.EPOCH_COMPLETED(every=2), checkpoint, to_save) trainer.run(torch.randn(10, 1), 5) to_load = to_save # load checkpoint myprefix_checkpoint_40.pt checkpoint.reload_objects(to_load=to_load, global_step=40) Note: If ``to_load`` contains objects of type torch `DistributedDataParallel`_ or `DataParallel`_, method ``load_state_dict`` will applied to their internal wrapped model (``obj.module``). Note: This method works only when the ``save_handler`` is of types string, :class:`~pathlib.Path` or :class:`~ignite.handlers.checkpoint.DiskSaver`. .. _DistributedDataParallel: https://pytorch.org/docs/stable/generated/ torch.nn.parallel.DistributedDataParallel.html .. _DataParallel: https://pytorch.org/docs/stable/generated/torch.nn.DataParallel.html """ if not isinstance(self.save_handler, DiskSaver): raise AttributeError( f"Checkpoint's `save_handler` should be of type `DiskSaver`, given {type(self.save_handler)}" ) global_step = filename_components.get("global_step", None) filename_pattern = self._get_filename_pattern(global_step) checkpoint = self._setup_checkpoint() name = "checkpoint" if len(checkpoint) == 1: for k in checkpoint: name = k name = filename_components.get("name", name) score = filename_components.get("score", None) filename_dict = { "filename_prefix": self.filename_prefix, "ext": self.ext, "name": name, "score_name": self.score_name, "score": score, "global_step": global_step, } checkpoint_fp = filename_pattern.format(**filename_dict) path = self.save_handler.dirname / checkpoint_fp load_kwargs = {} if load_kwargs is None else load_kwargs Checkpoint.load_objects(to_load=to_load, checkpoint=path, **load_kwargs) def state_dict(self) -> OrderedDict: """Method returns state dict with saved items: list of ``(priority, filename)`` pairs. Can be used to save internal state of the class. """ # TODO: this method should use _state_dict_all_req_keys return OrderedDict([("_saved", [(p, f) for p, f in self._saved])]) def load_state_dict(self, state_dict: Mapping) -> None: """Method replaces internal state of the class with provided state dict data. Args: state_dict: a dict with "saved" key and list of ``(priority, filename)`` pairs as values. """ super().load_state_dict(state_dict) self._saved = [Checkpoint.Item(p, f) for p, f in state_dict["_saved"]] @staticmethod def get_default_score_fn(metric_name: str, score_sign: float = 1.0) -> Callable: """Helper method to get default score function based on the metric name. Args: metric_name: metric name to get the value from ``engine.state.metrics``. Engine is the one to which :class:`~ignite.handlers.checkpoint.Checkpoint` handler is added. score_sign: sign of the score: 1.0 or -1.0. For error-like metrics, e.g. smaller is better, a negative score sign should be used (objects with larger score are retained). Default, 1.0. Examples: .. code-block:: python from ignite.handlers import Checkpoint best_acc_score = Checkpoint.get_default_score_fn("accuracy") best_model_handler = Checkpoint( to_save, save_handler, score_name="val_accuracy", score_function=best_acc_score ) evaluator.add_event_handler(Events.COMPLETED, best_model_handler) Usage with error-like metric: .. code-block:: python from ignite.handlers import Checkpoint neg_loss_score = Checkpoint.get_default_score_fn("loss", -1.0) best_model_handler = Checkpoint( to_save, save_handler, score_name="val_neg_loss", score_function=neg_loss_score ) evaluator.add_event_handler(Events.COMPLETED, best_model_handler) .. versionadded:: 0.4.3 """ if score_sign not in (1.0, -1.0): raise ValueError("Argument score_sign should be 1 or -1") def wrapper(engine: Engine) -> float: return score_sign * engine.state.metrics[metric_name] return wrapper class DiskSaver(BaseSaveHandler): """Handler that saves input checkpoint on a disk. Args: dirname: Directory path where the checkpoint will be saved atomic: if True, checkpoint is serialized to a temporary file, and then moved to final destination, so that files are guaranteed to not be damaged (for example if exception occurs during saving). create_dir: if True, will create directory ``dirname`` if it doesnt exist. require_empty: If True, will raise exception if there are any files in the directory ``dirname``. save_on_rank: The rank on which the checkpoint will be saved. Used in distributed configuration. kwargs: Accepted keyword arguments for `torch.save` or `xm.save`. .. versionchanged:: 0.4.2 Accept ``kwargs`` for `torch.save` or `xm.save`. .. versionchanged:: 0.4.10 Argument ``save_on_rank`` was added to specify the rank on which checkpoint should be saved. """ def __init__( self, dirname: Union[str, Path], atomic: bool = True, create_dir: bool = True, require_empty: bool = True, save_on_rank: int = 0, **kwargs: Any, ): self.dirname = Path(dirname).expanduser() self._atomic = atomic self.save_on_rank = save_on_rank if idist.get_rank() == save_on_rank: self._check_and_setup(self.dirname, create_dir, require_empty) self.kwargs = kwargs @staticmethod def _check_and_setup(dirname: Path, create_dir: bool, require_empty: bool) -> None: if create_dir: if not dirname.exists(): dirname.mkdir(parents=True) # Ensure that dirname exists if not dirname.exists(): raise ValueError(f"Directory path '{dirname}' is not found") if require_empty: matched = [fname for fname in os.listdir(dirname) if fname.endswith(".pt")] if len(matched) > 0: raise ValueError( f"Files {matched} with extension '.pt' are already present " f"in the directory {dirname}. If you want to use this " "directory anyway, pass `require_empty=False`." "" ) def __call__(self, checkpoint: Mapping, filename: str, metadata: Optional[Mapping] = None) -> None: path = self.dirname / filename if idist.has_xla_support: import torch_xla.core.xla_model as xm # all tpu procs should enter here as internally performs sync across device self._save_func(checkpoint, path, xm.save) elif self.save_on_rank == idist.get_rank(): self._save_func(checkpoint, path, torch.save) def _save_func(self, checkpoint: Mapping, path: Path, func: Callable) -> None: if not self._atomic: func(checkpoint, path, **self.kwargs) else: tmp = tempfile.NamedTemporaryFile(delete=False, dir=self.dirname) tmp_file = tmp.file tmp_name = tmp.name try: func(checkpoint, tmp_file, **self.kwargs) except BaseException: tmp.close() os.remove(tmp_name) raise else: tmp.close() os.replace(tmp.name, path) # append group/others read mode os.chmod(path, os.stat(path).st_mode | stat.S_IRGRP | stat.S_IROTH) def remove(self, filename: str) -> None: if idist.get_rank() == self.save_on_rank: path = self.dirname / filename path.unlink() class ModelCheckpoint(Checkpoint): """ModelCheckpoint handler, inherits from :class:`~ignite.handlers.checkpoint.Checkpoint`, can be used to periodically save objects to disk only. If needed to store checkpoints to another storage type, please consider :class:`~ignite.handlers.checkpoint.Checkpoint`. It also provides `last_checkpoint` attribute to show the last saved checkpoint. This handler expects two arguments: - an :class:`~ignite.engine.engine.Engine` object - a `dict` mapping names (`str`) to objects that should be saved to disk. See Examples for further details. .. warning:: Behaviour of this class has been changed since v0.3.0. There is no more internal counter that has been used to indicate the number of save actions. User could see its value `step_number` in the filename, e.g. `{filename_prefix}_{name}_{step_number}.pt`. Actually, `step_number` is replaced by current engine's epoch if `score_function` is specified and current iteration otherwise. A single `pt` file is created instead of multiple files. Args: dirname: Directory path where objects will be saved. filename_prefix: Prefix for the file names to which objects will be saved. See Notes of :class:`~ignite.handlers.checkpoint.Checkpoint` for more details. score_function: if not None, it should be a function taking a single argument, an :class:`~ignite.engine.engine.Engine` object, and return a score (`float`). Objects with highest scores will be retained. score_name: if ``score_function`` not None, it is possible to store its value using `score_name`. See Examples of :class:`~ignite.handlers.checkpoint.Checkpoint` for more details. n_saved: Number of objects that should be kept on disk. Older files will be removed. If set to `None`, all objects are kept. atomic: If True, objects are serialized to a temporary file, and then moved to final destination, so that files are guaranteed to not be damaged (for example if exception occurs during saving). require_empty: If True, will raise exception if there are any files starting with ``filename_prefix`` in the directory ``dirname``. create_dir: If True, will create directory ``dirname`` if it does not exist. global_step_transform: global step transform function to output a desired global step. Input of the function is `(engine, event_name)`. Output of function should be an integer. Default is None, global_step based on attached engine. If provided, uses function output as global_step. To setup global step from another engine, please use :meth:`~ignite.handlers.global_step_from_engine`. filename_pattern: If ``filename_pattern`` is provided, this pattern will be used to render checkpoint filenames. If the pattern is not defined, the default pattern would be used. See :class:`~ignite.handlers.checkpoint.Checkpoint` for details. include_self: Whether to include the `state_dict` of this object in the checkpoint. If `True`, then there must not be another object in ``to_save`` with key ``checkpointer``. greater_or_equal: if `True`, the latest equally scored model is stored. Otherwise, the first model. Default, `False`. save_on_rank: Which rank to save the objects on, in the distributed configuration. Used to instantiate a :class:`~ignite.handlers.DiskSaver` and is also passed to the parent class. kwargs: Accepted keyword arguments for `torch.save` or `xm.save` in `DiskSaver`. .. versionchanged:: 0.4.2 Accept ``kwargs`` for `torch.save` or `xm.save` .. versionchanged:: 0.4.9 Accept ``filename_pattern`` and ``greater_or_equal`` for parity with :class:`~ignite.handlers.checkpoint.Checkpoint` .. versionchanged:: 0.4.10 Added `save_on_rank` arg to save objects on this rank in a distributed configuration Examples: .. testcode:: python import os from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint from torch import nn trainer = Engine(lambda engine, batch: None) handler = ModelCheckpoint('/tmp/models', 'myprefix', n_saved=2, create_dir=True, require_empty=False) model = nn.Linear(3, 3) trainer.add_event_handler(Events.EPOCH_COMPLETED(every=2), handler, {'mymodel': model}) trainer.run([0, 1, 2, 3, 4], max_epochs=6) print(sorted(os.listdir('/tmp/models'))) print(handler.last_checkpoint) .. testoutput:: python ['myprefix_mymodel_20.pt', 'myprefix_mymodel_30.pt'] /tmp/models/myprefix_mymodel_30.pt """ def __init__( self, dirname: Union[str, Path], filename_prefix: str = "", score_function: Optional[Callable] = None, score_name: Optional[str] = None, n_saved: Union[int, None] = 1, atomic: bool = True, require_empty: bool = True, create_dir: bool = True, global_step_transform: Optional[Callable] = None, filename_pattern: Optional[str] = None, include_self: bool = False, greater_or_equal: bool = False, save_on_rank: int = 0, **kwargs: Any, ): disk_saver = DiskSaver( dirname, atomic=atomic, create_dir=create_dir, require_empty=require_empty, save_on_rank=save_on_rank, **kwargs, ) super(ModelCheckpoint, self).__init__( to_save={}, save_handler=disk_saver, filename_prefix=filename_prefix, score_function=score_function, score_name=score_name, n_saved=n_saved, global_step_transform=global_step_transform, filename_pattern=filename_pattern, include_self=include_self, greater_or_equal=greater_or_equal, save_on_rank=save_on_rank, ) @property def last_checkpoint(self) -> Optional[Union[str, Path]]: if len(self._saved) < 1: return None if not isinstance(self.save_handler, DiskSaver): raise RuntimeError(f"Internal error, save_handler should be DiskSaver, but has {type(self.save_handler)}.") return self.save_handler.dirname / self._saved[-1].filename def __call__(self, engine: Engine, to_save: Mapping): # type: ignore if len(to_save) == 0: raise RuntimeError("No objects to checkpoint found.") self._check_objects(to_save, "state_dict") self.to_save = to_save super(ModelCheckpoint, self).__call__(engine)