lightning/pytorch_lightning/trainer/progress.py

272 lines
8.8 KiB
Python

# 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.
from dataclasses import asdict, dataclass, field
from typing import Type
@dataclass
class BaseProgress:
"""Mixin that implements state-loading utilities for dataclasses."""
def state_dict(self) -> dict:
return asdict(self)
def load_state_dict(self, state_dict: dict) -> None:
self.__dict__.update(state_dict)
@classmethod
def from_state_dict(cls, state_dict: dict) -> "BaseProgress":
obj = cls()
obj.load_state_dict(state_dict)
return obj
@dataclass
class ReadyCompletedTracker(BaseProgress):
"""Track an event's progress.
Args:
ready: Intended to track the number of events ready to start.
completed: Intended to be incremented after the event completes (e.g. after ``on_*_end`` runs).
These attributes should be increased in order, that is, :attr:`ready` first and :attr:`completed` last.
"""
ready: int = 0
completed: int = 0
def reset(self) -> None:
"""Reset the state."""
self.ready = 0
self.completed = 0
def reset_on_restart(self) -> None:
"""Reset the progress on restart.
If there is a failure before all attributes are increased, restore the attributes to the last fully completed
value.
"""
self.ready = self.completed
@dataclass
class StartedTracker(ReadyCompletedTracker):
"""Track an event's progress.
Args:
ready: Intended to track the number of events ready to start.
started: Intended to be incremented after the event is started (e.g. after ``on_*_start`` runs).
completed: Intended to be incremented after the event completes (e.g. after ``on_*_end`` runs).
These attributes should be increased in order, that is, :attr:`ready` first and :attr:`completed` last.
"""
started: int = 0
def reset(self) -> None:
super().reset()
self.started = 0
def reset_on_restart(self) -> None:
super().reset_on_restart()
self.started = self.completed
@dataclass
class ProcessedTracker(StartedTracker):
"""Track an event's progress.
Args:
ready: Intended to track the number of events ready to start.
started: Intended to be incremented after the event is started (e.g. after ``on_*_start`` runs).
processed: Intended to be incremented after the event is processed.
completed: Intended to be incremented after the event completes (e.g. after ``on_*_end`` runs).
These attributes should be increased in order, that is, :attr:`ready` first and :attr:`completed` last.
"""
processed: int = 0
def reset(self) -> None:
super().reset()
self.processed = 0
def reset_on_restart(self) -> None:
super().reset_on_restart()
self.processed = self.completed
@dataclass
class Progress(BaseProgress):
"""Track aggregated and current progress.
Args:
total: Intended to track the total progress of an event.
current: Intended to track the current progress of an event.
"""
total: ReadyCompletedTracker = field(default_factory=ProcessedTracker)
current: ReadyCompletedTracker = field(default_factory=ProcessedTracker)
def __post_init__(self) -> None:
if type(self.total) is not type(self.current): # noqa: E721
raise ValueError("The `total` and `current` instances should be of the same class")
def increment_ready(self) -> None:
self.total.ready += 1
self.current.ready += 1
def increment_started(self) -> None:
if not isinstance(self.total, StartedTracker):
raise TypeError(f"`{self.total.__class__.__name__}` doesn't have a `started` attribute")
self.total.started += 1
self.current.started += 1
def increment_processed(self) -> None:
if not isinstance(self.total, ProcessedTracker):
raise TypeError(f"`{self.total.__class__.__name__}` doesn't have a `processed` attribute")
self.total.processed += 1
self.current.processed += 1
def increment_completed(self) -> None:
self.total.completed += 1
self.current.completed += 1
@classmethod
def from_defaults(cls, tracker_cls: Type[ReadyCompletedTracker], **kwargs: int) -> "Progress":
"""Utility function to easily create an instance from keyword arguments to both ``Tracker``s."""
return cls(total=tracker_cls(**kwargs), current=tracker_cls(**kwargs))
def reset_on_epoch(self) -> None:
self.current.reset()
def reset_on_run(self) -> None:
self.current.reset()
def reset_on_restart(self) -> None:
self.current.reset_on_restart()
def load_state_dict(self, state_dict: dict) -> None:
self.total.load_state_dict(state_dict["total"])
self.current.load_state_dict(state_dict["current"])
@dataclass
class DataLoaderProgress(Progress):
"""Tracks dataloader progress.
These counters are local to a trainer rank. By default, they are not globally synced across all ranks.
Args:
total: Tracks the total dataloader progress.
current: Tracks the current dataloader progress.
"""
total: ReadyCompletedTracker = field(default_factory=ReadyCompletedTracker)
current: ReadyCompletedTracker = field(default_factory=ReadyCompletedTracker)
@dataclass
class BatchProgress(Progress):
"""Tracks batch progress.
These counters are local to a trainer rank. By default, they are not globally synced across all ranks.
Args:
total: Tracks the total dataloader progress.
current: Tracks the current dataloader progress.
is_last_batch: Whether the batch is the last one. This is useful for iterable datasets.
"""
is_last_batch: bool = False
def reset_on_run(self) -> None:
super().reset_on_run()
self.is_last_batch = False
def load_state_dict(self, state_dict: dict) -> None:
super().load_state_dict(state_dict)
self.is_last_batch = state_dict["is_last_batch"]
@dataclass
class SchedulerProgress(Progress):
"""Tracks scheduler progress.
These counters are local to a trainer rank. By default, they are not globally synced across all ranks.
Args:
total: Tracks the total scheduler progress.
current: Tracks the current scheduler progress.
"""
total: ReadyCompletedTracker = field(default_factory=ReadyCompletedTracker)
current: ReadyCompletedTracker = field(default_factory=ReadyCompletedTracker)
@dataclass
class OptimizerProgress(BaseProgress):
"""Track optimizer progress.
Args:
step: Tracks ``optimizer.step`` calls.
zero_grad: Tracks ``optimizer.zero_grad`` calls.
"""
step: Progress = field(default_factory=lambda: Progress.from_defaults(ReadyCompletedTracker))
zero_grad: Progress = field(default_factory=lambda: Progress.from_defaults(StartedTracker))
def reset_on_run(self) -> None:
self.step.reset_on_run()
self.zero_grad.reset_on_run()
def reset_on_restart(self) -> None:
self.step.reset_on_restart()
self.zero_grad.reset_on_restart()
def load_state_dict(self, state_dict: dict) -> None:
self.step.load_state_dict(state_dict["step"])
self.zero_grad.load_state_dict(state_dict["zero_grad"])
@dataclass
class OptimizationProgress(BaseProgress):
"""Track optimization progress.
Args:
optimizer: Tracks optimizer progress.
optimizer_position: The index of the current optimizer amongst the currently active optimizers.
Used to know which optimizer we were using when restarting.
Since not all optimizers may be active at a given time, this index is different from the ``optimizer_idx``
seen in the optimization loops.
"""
# TODO: support for multiple optimizers
optimizer: OptimizerProgress = field(default_factory=OptimizerProgress)
optimizer_position: int = 0
@property
def optimizer_steps(self) -> int:
return self.optimizer.step.total.completed
def reset_on_run(self) -> None:
self.optimizer.reset_on_run()
def reset_on_restart(self) -> None:
self.optimizer.reset_on_restart()
def load_state_dict(self, state_dict: dict) -> None:
self.optimizer.load_state_dict(state_dict["optimizer"])
self.optimizer_position = state_dict["optimizer_position"]