lightning/pytorch_lightning/core/datamodule.py

408 lines
14 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.
import functools
import inspect
from abc import abstractmethod
from argparse import ArgumentParser, Namespace
from typing import Any, List, Optional, Tuple, Union
from torch.utils.data import DataLoader
from pytorch_lightning.utilities import parsing, rank_zero_only, rank_zero_warn
class _DataModuleWrapper(type):
def __call__(cls, *args, **kwargs):
"""A wrapper for LightningDataModule that:
1. Runs user defined subclass's __init__
2. Assures prepare_data() runs on rank 0
3. Lets you check prepare_data and setup to see if they've been called
"""
# Track prepare_data calls and make sure it runs on rank zero
cls.prepare_data = track_data_hook_calls(rank_zero_only(cls.prepare_data))
# Track setup calls
cls.setup = track_data_hook_calls(cls.setup)
# Get instance of LightningDataModule by mocking its __init__ via __call__
obj = type.__call__(cls, *args, **kwargs)
return obj
def track_data_hook_calls(fn):
"""A decorator that checks if prepare_data/setup have been called.
- When dm.prepare_data() is called, dm.has_prepared_data gets set to True
- When dm.setup('fit') is called, dm.has_setup_fit gets set to True
- When dm.setup('test') is called, dm.has_setup_test gets set to True
- When dm.setup() is called without stage arg, both dm.has_setup_fit and dm.has_setup_test get set to True
Args:
fn (function): Function that will be tracked to see if it has been called.
Returns:
function: Decorated function that tracks its call status and saves it to private attrs in its obj instance.
"""
@functools.wraps(fn)
def wrapped_fn(*args, **kwargs):
# The object instance from which setup or prepare_data was called
obj = args[0]
# If calling setup, we check the stage and assign stage-specific bool args
if fn.__name__ == 'setup':
# Get stage either by grabbing from args or checking kwargs.
# If not provided, set call status of 'fit' and 'test' to True.
# We do this so __attach_datamodule in trainer.py doesn't mistakenly call setup('test') on trainer.test()
stage = args[1] if len(args) > 1 else kwargs.get('stage', None)
if stage == 'fit' or stage is None:
obj._has_setup_fit = True
if stage == 'test' or stage is None:
obj._has_setup_test = True
if fn.__name__ == 'prepare_data':
obj._has_prepared_data = True
return fn(*args, **kwargs)
return wrapped_fn
class LightningDataModule(object, metaclass=_DataModuleWrapper): # pragma: no cover
"""
A DataModule standardizes the training, val, test splits, data preparation and transforms.
The main advantage is consistent data splits, data preparation and transforms across models.
Example::
class MyDataModule(LightningDataModule):
def __init__(self):
super().__init__()
def prepare_data(self):
# download, split, etc...
# only called on 1 GPU/TPU in distributed
def setup(self):
# make assignments here (val/train/test split)
# called on every process in DDP
def train_dataloader(self):
train_split = Dataset(...)
return DataLoader(train_split)
def val_dataloader(self):
val_split = Dataset(...)
return DataLoader(val_split)
def test_dataloader(self):
test_split = Dataset(...)
return DataLoader(test_split)
A DataModule implements 5 key methods:
* **prepare_data** (things to do on 1 GPU/TPU not on every GPU/TPU in distributed mode).
* **setup** (things to do on every accelerator in distributed mode).
* **train_dataloader** the training dataloader.
* **val_dataloader** the val dataloader(s).
* **test_dataloader** the test dataloader(s).
This allows you to share a full dataset without explaining how to download,
split transform and process the data
"""
name: str = ...
def __init__(
self, train_transforms=None, val_transforms=None, test_transforms=None,
):
super().__init__()
self._train_transforms = train_transforms
self._val_transforms = val_transforms
self._test_transforms = test_transforms
self.dims = ()
# Private attrs to keep track of whether or not data hooks have been called yet
self._has_prepared_data = False
self._has_setup_fit = False
self._has_setup_test = False
@property
def train_transforms(self):
"""
Optional transforms (or collection of transforms) you can apply to train dataset
"""
return self._train_transforms
@train_transforms.setter
def train_transforms(self, t):
self._train_transforms = t
@property
def val_transforms(self):
"""
Optional transforms (or collection of transforms) you can apply to validation dataset
"""
return self._val_transforms
@val_transforms.setter
def val_transforms(self, t):
self._val_transforms = t
@property
def test_transforms(self):
"""
Optional transforms (or collection of transforms) you can apply to test dataset
"""
return self._test_transforms
@test_transforms.setter
def test_transforms(self, t):
self._test_transforms = t
def size(self, dim=None) -> Union[Tuple, int]:
"""
Return the dimension of each input either as a tuple or list of tuples.
"""
if dim is not None:
return self.dims[dim]
return self.dims
@property
def has_prepared_data(self):
"""Return bool letting you know if datamodule.prepare_data() has been called or not.
Returns:
bool: True if datamodule.prepare_data() has been called. False by default.
"""
return self._has_prepared_data
@property
def has_setup_fit(self):
"""Return bool letting you know if datamodule.setup('fit') has been called or not.
Returns:
bool: True if datamodule.setup('fit') has been called. False by default.
"""
return self._has_setup_fit
@property
def has_setup_test(self):
"""Return bool letting you know if datamodule.setup('test') has been called or not.
Returns:
bool: True if datamodule.setup('test') has been called. False by default.
"""
return self._has_setup_test
@abstractmethod
def prepare_data(self, *args, **kwargs):
"""
Use this to download and prepare data.
In distributed (GPU, TPU), this will only be called once.
.. warning:: Do not assign anything to the datamodule in this step since this will only be called on 1 GPU.
Pseudocode::
dm.prepare_data()
dm.setup()
Example::
def prepare_data(self):
download_imagenet()
clean_imagenet()
cache_imagenet()
"""
@abstractmethod
def setup(self, stage: Optional[str] = None):
"""
Use this to load your data from file, split it, etc. You are safe to make state assignments here.
This hook is called on every process when using DDP.
Example::
def setup(self, stage):
data = load_data(...)
self.train_ds, self.val_ds, self.test_ds = split_data(data)
"""
@abstractmethod
def train_dataloader(self, *args, **kwargs) -> DataLoader:
"""
Implement a PyTorch DataLoader for training.
Return:
Single PyTorch :class:`~torch.utils.data.DataLoader`.
Note:
Lightning adds the correct sampler for distributed and arbitrary hardware.
There is no need to set it yourself.
Example::
def train_dataloader(self):
dataset = MNIST(root=PATH, train=True, transform=transforms.ToTensor(), download=False)
loader = torch.utils.data.DataLoader(dataset=dataset)
return loader
"""
rank_zero_warn('`train_dataloader` must be implemented to be used with the Lightning Trainer')
@abstractmethod
def val_dataloader(self, *args, **kwargs) -> Union[DataLoader, List[DataLoader]]:
r"""
Implement a PyTorch DataLoader for training.
Return:
Single PyTorch :class:`~torch.utils.data.DataLoader`.
Note:
Lightning adds the correct sampler for distributed and arbitrary hardware.
There is no need to set it yourself.
Note:
You can also return a list of DataLoaders
Example::
def val_dataloader(self):
dataset = MNIST(root=PATH, train=False, transform=transforms.ToTensor(), download=False)
loader = torch.utils.data.DataLoader(dataset=dataset, shuffle=False)
return loader
"""
@abstractmethod
def test_dataloader(self, *args, **kwargs) -> Union[DataLoader, List[DataLoader]]:
r"""
Implement a PyTorch DataLoader for training.
Return:
Single PyTorch :class:`~torch.utils.data.DataLoader`.
Note:
Lightning adds the correct sampler for distributed and arbitrary hardware.
There is no need to set it yourself.
Note:
You can also return a list of DataLoaders
Example::
def test_dataloader(self):
dataset = MNIST(root=PATH, train=False, transform=transforms.ToTensor(), download=False)
loader = torch.utils.data.DataLoader(dataset=dataset, shuffle=False)
return loader
"""
@classmethod
def add_argparse_args(cls, parent_parser: ArgumentParser) -> ArgumentParser:
r"""Extends existing argparse by default `LightningDataModule` attributes.
"""
parser = ArgumentParser(parents=[parent_parser], add_help=False,)
added_args = [x.dest for x in parser._actions]
blacklist = ['kwargs']
depr_arg_names = blacklist + added_args
depr_arg_names = set(depr_arg_names)
allowed_types = (str, float, int, bool)
# TODO: get "help" from docstring :)
for arg, arg_types, arg_default in (
at for at in cls.get_init_arguments_and_types() if at[0] not in depr_arg_names
):
arg_types = [at for at in allowed_types if at in arg_types]
if not arg_types:
# skip argument with not supported type
continue
arg_kwargs = {}
if bool in arg_types:
arg_kwargs.update(nargs="?")
# if the only arg type is bool
if len(arg_types) == 1:
# redefine the type for ArgParser needed
def use_type(x):
return bool(parsing.str_to_bool(x))
else:
# filter out the bool as we need to use more general
use_type = [at for at in arg_types if at is not bool][0]
else:
use_type = arg_types[0]
if arg_default == inspect._empty:
arg_default = None
parser.add_argument(
f'--{arg}',
dest=arg,
default=arg_default,
type=use_type,
help=f'autogenerated by plb.{cls.__name__}',
**arg_kwargs,
)
return parser
@classmethod
def from_argparse_args(cls, args: Union[Namespace, ArgumentParser], **kwargs):
"""
Create an instance from CLI arguments.
Args:
args: The parser or namespace to take arguments from. Only known arguments will be
parsed and passed to the :class:`LightningDataModule`.
**kwargs: Additional keyword arguments that may override ones in the parser or namespace.
These must be valid DataModule arguments.
Example::
parser = ArgumentParser(add_help=False)
parser = LightningDataModule.add_argparse_args(parser)
module = LightningDataModule.from_argparse_args(args)
"""
if isinstance(args, ArgumentParser):
args = cls.parse_argparser(args)
params = vars(args)
# we only want to pass in valid DataModule args, the rest may be user specific
valid_kwargs = inspect.signature(cls.__init__).parameters
datamodule_kwargs = dict((name, params[name]) for name in valid_kwargs if name in params)
datamodule_kwargs.update(**kwargs)
return cls(**datamodule_kwargs)
@classmethod
def get_init_arguments_and_types(cls) -> List[Tuple[str, Tuple, Any]]:
r"""Scans the DataModule signature and returns argument names, types and default values.
Returns:
List with tuples of 3 values:
(argument name, set with argument types, argument default value).
"""
datamodule_default_params = inspect.signature(cls.__init__).parameters
name_type_default = []
for arg in datamodule_default_params:
arg_type = datamodule_default_params[arg].annotation
arg_default = datamodule_default_params[arg].default
try:
arg_types = tuple(arg_type.__args__)
except AttributeError:
arg_types = (arg_type,)
name_type_default.append((arg, arg_types, arg_default))
return name_type_default