lightning/pytorch_lightning/trainer/trainer.py

1489 lines
57 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 inspect
import os
import warnings
from argparse import ArgumentParser, Namespace
from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
import torch
import torch.distributed as torch_distrib
import torch.multiprocessing as mp
from torch.utils.data import DataLoader
from pytorch_lightning.callbacks import Callback, EarlyStopping, ModelCheckpoint
from pytorch_lightning.core.datamodule import LightningDataModule
from pytorch_lightning.core.lightning import LightningModule
from pytorch_lightning.core.memory import ModelSummary
from pytorch_lightning.core.step_result import EvalResult
from pytorch_lightning.loggers import LightningLoggerBase
from pytorch_lightning.profiler import BaseProfiler, PassThroughProfiler, SimpleProfiler
from pytorch_lightning.trainer.auto_mix_precision import NATIVE_AMP_AVALAIBLE, TrainerAMPMixin
from pytorch_lightning.trainer.callback_config import TrainerCallbackConfigMixin
from pytorch_lightning.trainer.callback_hook import TrainerCallbackHookMixin
from pytorch_lightning.trainer.data_loading import TrainerDataLoadingMixin
from pytorch_lightning.trainer.deprecated_api import TrainerDeprecatedAPITillVer0_9, TrainerDeprecatedAPITillVer0_10
from pytorch_lightning.trainer.distrib_data_parallel import TrainerDDPMixin
from pytorch_lightning.trainer.distrib_parts import (TrainerDPMixin, _parse_gpu_ids, _parse_tpu_cores,
determine_root_gpu_device, pick_multiple_gpus)
from pytorch_lightning.trainer.evaluation_loop import TrainerEvaluationLoopMixin
from pytorch_lightning.trainer.logging import TrainerLoggingMixin
from pytorch_lightning.trainer.lr_finder import TrainerLRFinderMixin
from pytorch_lightning.trainer.model_hooks import TrainerModelHooksMixin
from pytorch_lightning.trainer.optimizers import TrainerOptimizersMixin
from pytorch_lightning.trainer.supporters import TensorRunningAccum
from pytorch_lightning.trainer.training_io import TrainerIOMixin
from pytorch_lightning.trainer.training_loop import TrainerTrainLoopMixin
from pytorch_lightning.trainer.training_tricks import TrainerTrainingTricksMixin
from pytorch_lightning.utilities import parsing, rank_zero_info, rank_zero_only, rank_zero_warn
from pytorch_lightning.utilities.debugging import InternalDebugger
from pytorch_lightning.utilities.exceptions import MisconfigurationException
from pytorch_lightning.trainer.configuration_validator import ConfigValidator
from pytorch_lightning.accelerator_backends import (
GPUBackend, TPUBackend, CPUBackend, DDPSpawnBackend, DataParallelBackend)
# warnings to ignore in trainer
warnings.filterwarnings(
'ignore', message='torch.distributed.reduce_op is deprecated, ' 'please use torch.distributed.ReduceOp instead'
)
try:
from apex import amp
except ImportError:
APEX_AVAILABLE = False
else:
APEX_AVAILABLE = True
try:
import torch_xla
import torch_xla.core.xla_model as xm
import torch_xla.distributed.xla_multiprocessing as xmp
except ImportError:
XLA_AVAILABLE = False
else:
XLA_AVAILABLE = True
try:
import horovod.torch as hvd
except (ModuleNotFoundError, ImportError):
HOROVOD_AVAILABLE = False
else:
HOROVOD_AVAILABLE = True
class Trainer(
TrainerIOMixin,
TrainerCallbackHookMixin,
TrainerModelHooksMixin,
TrainerOptimizersMixin,
TrainerAMPMixin,
TrainerDPMixin,
TrainerDDPMixin,
TrainerLoggingMixin,
TrainerTrainingTricksMixin,
TrainerDataLoadingMixin,
TrainerEvaluationLoopMixin,
TrainerTrainLoopMixin,
TrainerCallbackConfigMixin,
TrainerLRFinderMixin,
TrainerDeprecatedAPITillVer0_9,
TrainerDeprecatedAPITillVer0_10,
):
"""
Example:
>>> import torch
>>> from torch.nn import functional as F
>>> from torch.utils.data import Dataset, DataLoader
>>> # Define model
>>> class SimpleModel(LightningModule):
... def __init__(self):
... super().__init__()
... self.l1 = torch.nn.Linear(in_features=64, out_features=4)
...
... def forward(self, x):
... return torch.relu(self.l1(x.view(x.size(0), -1)))
...
... def training_step(self, batch, batch_nb):
... x, y = batch
... loss = F.cross_entropy(self(x), y)
... return {'loss': loss, 'log': {'train_loss': loss}}
...
... def test_step(self, batch, batch_nb):
... x, y = batch
... loss = F.cross_entropy(self(x), y)
... return {'loss': loss, 'log': {'test_loss': loss}}
...
... def configure_optimizers(self):
... return torch.optim.Adam(self.parameters(), lr=0.02)
...
>>> # Define dataset
>>> class SimpleDataset(Dataset):
... def __init__(self, num_samples=200):
... self.input_seq = torch.randn(num_samples, 64)
... self.output_seq = torch.randint(0, 4, (num_samples,))
...
... def __len__(self):
... return len(self.input_seq)
...
... def __getitem__(self, item):
... return self.input_seq[item], self.output_seq[item]
...
>>> train_loader = DataLoader(SimpleDataset(), batch_size=8)
>>> model = SimpleModel()
>>> # Define Trainer and fit model
>>> trainer = Trainer(max_epochs=1, progress_bar_refresh_rate=0)
>>> trainer.fit(model, train_loader)
1
>>> test_outputs = trainer.test(model, train_loader, verbose=False)
>>> len(test_outputs)
25
"""
DEPRECATED_IN_0_9 = ('use_amp', 'show_progress_bar', 'training_tqdm_dict', 'num_tpu_cores')
def __init__(
self,
logger: Union[LightningLoggerBase, Iterable[LightningLoggerBase], bool] = True,
checkpoint_callback: Union[ModelCheckpoint, bool] = True,
early_stop_callback: Optional[Union[EarlyStopping, bool]] = False,
callbacks: Optional[List[Callback]] = None,
default_root_dir: Optional[str] = None,
gradient_clip_val: float = 0,
process_position: int = 0,
num_nodes: int = 1,
num_processes: int = 1,
gpus: Optional[Union[List[int], str, int]] = None,
auto_select_gpus: bool = False,
tpu_cores: Optional[Union[List[int], str, int]] = None,
log_gpu_memory: Optional[str] = None,
progress_bar_refresh_rate: int = 1,
overfit_batches: Union[int, float] = 0.0,
track_grad_norm: Union[int, float, str] = -1,
check_val_every_n_epoch: int = 1,
fast_dev_run: bool = False,
accumulate_grad_batches: Union[int, Dict[int, int], List[list]] = 1,
max_epochs: int = 1000,
min_epochs: int = 1,
max_steps: Optional[int] = None,
min_steps: Optional[int] = None,
limit_train_batches: Union[int, float] = 1.0,
limit_val_batches: Union[int, float] = 1.0,
limit_test_batches: Union[int, float] = 1.0,
val_check_interval: Union[int, float] = 1.0,
log_save_interval: int = 100,
row_log_interval: int = 50,
distributed_backend: Optional[str] = None,
precision: int = 32,
print_nan_grads: bool = False, # backward compatible, todo: remove in v0.9.0
weights_summary: Optional[str] = ModelSummary.MODE_DEFAULT,
weights_save_path: Optional[str] = None,
num_sanity_val_steps: int = 2,
truncated_bptt_steps: Optional[int] = None,
resume_from_checkpoint: Optional[str] = None,
profiler: Optional[Union[BaseProfiler, bool]] = None,
benchmark: bool = False,
deterministic: bool = False,
reload_dataloaders_every_epoch: bool = False,
auto_lr_find: Union[bool, str] = False,
replace_sampler_ddp: bool = True,
terminate_on_nan: bool = False,
auto_scale_batch_size: Union[str, bool] = False,
prepare_data_per_node: bool = True,
amp_level: str = 'O2', # backward compatible, todo: remove in v1.0.0
num_tpu_cores: Optional[int] = None, # backward compatible, todo: remove in v0.9.0
use_amp=None, # backward compatible, todo: remove in v0.9.0
show_progress_bar=None, # backward compatible, todo: remove in v0.9.0
val_percent_check: float = None, # backward compatible, todo: remove in v0.10.0
test_percent_check: float = None, # backward compatible, todo: remove in v0.10.0
train_percent_check: float = None, # backward compatible, todo: remove in v0.10.0
overfit_pct: float = None, # backward compatible, todo: remove in v1.0.0
):
r"""
Customize every aspect of training via flags
Args:
logger: Logger (or iterable collection of loggers) for experiment tracking.
checkpoint_callback: Callback for checkpointing.
early_stop_callback (:class:`pytorch_lightning.callbacks.EarlyStopping`):
callbacks: Add a list of callbacks.
default_root_dir: Default path for logs and weights when no logger/ckpt_callback passed.
Default: ``os.getcwd()``.
gradient_clip_val: 0 means don't clip.
gradient_clip:
.. warning:: .. deprecated:: 0.7.0
Use `gradient_clip_val` instead. Will remove 0.9.0.
process_position: orders the progress bar when running multiple models on same machine.
num_nodes: number of GPU nodes for distributed training.
nb_gpu_nodes:
.. warning:: .. deprecated:: 0.7.0
Use `num_nodes` instead. Will remove 0.9.0.
gpus: Which GPUs to train on.
auto_select_gpus:
If enabled and `gpus` is an integer, pick available
gpus automatically. This is especially useful when
GPUs are configured to be in "exclusive mode", such
that only one process at a time can access them.
tpu_cores: How many TPU cores to train on (1 or 8) / Single TPU to train on [1]
num_tpu_cores: How many TPU cores to train on (1 or 8)
.. warning:: .. deprecated:: 0.7.6. Will remove 0.9.0.
log_gpu_memory: None, 'min_max', 'all'. Might slow performance
show_progress_bar:
.. warning:: .. deprecated:: 0.7.2
Set `progress_bar_refresh_rate` to positive integer to enable. Will remove 0.9.0.
progress_bar_refresh_rate: How often to refresh progress bar (in steps). Value ``0`` disables progress bar.
Ignored when a custom callback is passed to :paramref:`~Trainer.callbacks`.
overfit_batches: Overfit a percent of training data (float) or a set number of batches (int). Default: 0.0
overfit_pct:
.. warning:: .. deprecated:: 0.8.0
Use `overfit_batches` instead. Will be removed in 0.10.0.
track_grad_norm: -1 no tracking. Otherwise tracks that p-norm. May be set to 'inf' infinity-norm.
check_val_every_n_epoch: Check val every n train epochs.
fast_dev_run: runs 1 batch of train, test and val to find any bugs (ie: a sort of unit test).
accumulate_grad_batches: Accumulates grads every k batches or as set up in the dict.
max_epochs: Stop training once this number of epochs is reached.
max_nb_epochs:
.. warning:: .. deprecated:: 0.7.0
Use `max_epochs` instead. Will remove 0.9.0.
min_epochs: Force training for at least these many epochs
min_nb_epochs:
.. warning:: .. deprecated:: 0.7.0
Use `min_epochs` instead. Will remove 0.9.0.
max_steps: Stop training after this number of steps. Disabled by default (None).
min_steps: Force training for at least these number of steps. Disabled by default (None).
limit_train_batches: How much of training dataset to check (floats = percent, int = num_batches)
limit_val_batches: How much of validation dataset to check (floats = percent, int = num_batches)
limit_test_batches: How much of test dataset to check (floats = percent, int = num_batches)
train_percent_check:
.. warning:: .. deprecated:: 0.8.0
Use `limit_train_batches` instead. Will remove v0.10.0.
val_percent_check:
.. warning:: .. deprecated:: 0.8.0
Use `limit_val_batches` instead. Will remove v0.10.0.
test_percent_check:
.. warning:: .. deprecated:: 0.8.0
Use `limit_test_batches` instead. Will remove v0.10.0.
val_check_interval: How often within one training epoch to check the validation set
log_save_interval: Writes logs to disk this often
row_log_interval: How often to add logging rows (does not write to disk)
add_row_log_interval:
.. warning:: .. deprecated:: 0.7.0
Use `row_log_interval` instead. Will remove 0.9.0.
distributed_backend: The distributed backend to use (dp, ddp, ddp2, ddp_spawn, ddp_cpu)
use_amp:
.. warning:: .. deprecated:: 0.7.0
Use `precision` instead. Will remove 0.9.0.
precision: Full precision (32), half precision (16).
print_nan_grads:
.. warning:: .. deprecated:: 0.7.2
Has no effect. When detected, NaN grads will be printed automatically.
Will remove 0.9.0.
weights_summary: Prints a summary of the weights when training begins.
weights_save_path: Where to save weights if specified. Will override default_root_dir
for checkpoints only. Use this if for whatever reason you need the checkpoints
stored in a different place than the logs written in `default_root_dir`.
Defaults to `default_root_dir`.
amp_level: The optimization level to use (O1, O2, etc...).
num_sanity_val_steps: Sanity check runs n validation batches before starting the training routine.
Set it to `-1` to run all batches in all validation dataloaders. Default: 2
truncated_bptt_steps: Truncated back prop breaks performs backprop every k steps of
resume_from_checkpoint: To resume training from a specific checkpoint pass in the path here.
This can be a URL.
profiler: To profile individual steps during training and assist in
reload_dataloaders_every_epoch: Set to True to reload dataloaders every epoch
auto_lr_find: If set to True, will `initially` run a learning rate finder,
trying to optimize initial learning for faster convergence. Sets learning
rate in self.lr or self.learning_rate in the LightningModule.
To use a different key, set a string instead of True with the key name.
replace_sampler_ddp: Explicitly enables or disables sampler replacement.
If not specified this will toggled automatically ddp is used
benchmark: If true enables cudnn.benchmark.
deterministic: If true enables cudnn.deterministic
terminate_on_nan: If set to True, will terminate training (by raising a `ValueError`) at the
end of each training batch, if any of the parameters or the loss are NaN or +/-inf.
auto_scale_batch_size: If set to True, will `initially` run a batch size
finder trying to find the largest batch size that fits into memory.
The result will be stored in self.batch_size in the LightningModule.
Additionally, can be set to either `power` that estimates the batch size through
a power search or `binsearch` that estimates the batch size through a binary search.
prepare_data_per_node: If True, each LOCAL_RANK=0 will call prepare data.
Otherwise only NODE_RANK=0, LOCAL_RANK=0 will prepare data
"""
super().__init__()
self.deterministic = deterministic
torch.backends.cudnn.deterministic = self.deterministic
if self.deterministic:
# fixing non-deterministic part of horovod
# https://github.com/PyTorchLightning/pytorch-lightning/pull/1572/files#r420279383
os.environ["HOROVOD_FUSION_THRESHOLD"] = str(0)
# init the default rank if exists
# we need to call this here or NVIDIA flags and other messaging in init will show on all ranks
# this way we only show it on rank 0
if 'LOCAL_RANK' in os.environ:
rank_zero_only.rank = int(os.environ['LOCAL_RANK'])
# training bookeeping
self.total_batch_idx = 0
self.running_loss = TensorRunningAccum(window_length=20)
self.batch_idx = 0
self.progress_bar_metrics = {}
self.callback_metrics = {}
self.num_training_batches = 0
self.num_val_batches = []
self.num_test_batches = []
self.train_dataloader = None
self.test_dataloaders = None
self.val_dataloaders = None
# when true, prints test results
self.verbose_test = True
# when .test() is called, it sets this
self.tested_ckpt_path = None
# training state
self.model = None
self.testing = False
self.prepare_data_per_node = prepare_data_per_node
self.lr_schedulers = []
self.optimizers = None
self.optimizer_frequencies = []
self.global_step = 0
self.current_epoch = 0
self.interrupted = False
self.should_stop = False
self.running_sanity_check = False
self._default_root_dir = default_root_dir or os.getcwd()
self._weights_save_path = weights_save_path or self._default_root_dir
# init callbacks
self.callbacks = callbacks or []
# configure early stop callback
# creates a default one if none passed in
early_stop_callback = self.configure_early_stopping(early_stop_callback)
if early_stop_callback:
self.callbacks.append(early_stop_callback)
# configure checkpoint callback
# it is important that this is the last callback to run
# pass through the required args to figure out defaults
checkpoint_callback = self.configure_checkpoint_callback(checkpoint_callback)
if checkpoint_callback:
self.callbacks.append(checkpoint_callback)
# TODO refactor codebase (tests) to not directly reach into these callbacks
self.checkpoint_callback = checkpoint_callback
self.early_stop_callback = early_stop_callback
self.on_init_start()
# benchmarking
self.benchmark = benchmark
torch.backends.cudnn.benchmark = self.benchmark
# Transfer params
self.num_nodes = num_nodes
self.log_gpu_memory = log_gpu_memory
self.gradient_clip_val = gradient_clip_val
self.check_val_every_n_epoch = check_val_every_n_epoch
if not isinstance(track_grad_norm, (int, float)) and track_grad_norm != 'inf':
raise MisconfigurationException("track_grad_norm can be an int, a float or 'inf' (infinity norm).")
self.track_grad_norm = float(track_grad_norm)
# tpu config
if num_tpu_cores is not None:
rank_zero_warn(
"Argument `num_tpu_cores` is now set by `tpu_cores` since v0.7.6"
" and this argument will be removed in v0.9.0",
DeprecationWarning,
)
if tpu_cores is None:
tpu_cores = num_tpu_cores
self.tpu_cores = _parse_tpu_cores(tpu_cores)
self.on_tpu = self.tpu_cores is not None
self.tpu_id = self.tpu_cores[0] if isinstance(self.tpu_cores, list) else None
if num_processes != 1 and distributed_backend != "ddp_cpu":
rank_zero_warn("num_processes is only used for distributed_backend=\"ddp_cpu\". Ignoring it.")
self.num_processes = num_processes
self.weights_summary = weights_summary
self.max_epochs = max_epochs
self.min_epochs = min_epochs
self.max_steps = max_steps
self.min_steps = min_steps
if num_sanity_val_steps == -1:
self.num_sanity_val_steps = float("inf")
else:
self.num_sanity_val_steps = min(num_sanity_val_steps, limit_val_batches)
# Backward compatibility, TODO: remove in v0.9.0
if print_nan_grads:
rank_zero_warn(
"Argument `print_nan_grads` has no effect and will be removed in v0.9.0."
" NaN grads will be printed automatically when detected.",
DeprecationWarning,
)
self.reload_dataloaders_every_epoch = reload_dataloaders_every_epoch
self.auto_lr_find = auto_lr_find
self.auto_scale_batch_size = auto_scale_batch_size
self._is_data_prepared = False
self.replace_sampler_ddp = replace_sampler_ddp
self.truncated_bptt_steps = truncated_bptt_steps
self.resume_from_checkpoint = resume_from_checkpoint
self.terminate_on_nan = terminate_on_nan
self.shown_warnings = set()
self.fast_dev_run = fast_dev_run
if self.fast_dev_run:
limit_train_batches = 1
limit_val_batches = 1
limit_test_batches = 1
self.num_sanity_val_steps = 0
self.max_epochs = 1
rank_zero_info(
'Running in fast_dev_run mode: will run a full train,' ' val and test loop using a single batch'
)
# configure profiler
if profiler is True:
profiler = SimpleProfiler()
self.profiler = profiler or PassThroughProfiler()
# accumulated grads
self.accumulate_grad_batches = accumulate_grad_batches
self.configure_accumulated_gradients(accumulate_grad_batches)
# for gpus allow int, string and gpu list
if auto_select_gpus and isinstance(gpus, int):
self.gpus = pick_multiple_gpus(gpus)
else:
self.gpus = gpus
self.data_parallel_device_ids = _parse_gpu_ids(self.gpus)
self.root_gpu = determine_root_gpu_device(self.data_parallel_device_ids)
self.root_device = torch.device("cpu")
self.on_gpu = True if (self.data_parallel_device_ids and torch.cuda.is_available()) else False
# tpu state flags
self.use_tpu = False
self.tpu_local_core_rank = None
self.tpu_global_core_rank = None
# distributed backend choice
self.distributed_backend = distributed_backend
self.set_distributed_mode(distributed_backend)
# override dist backend when using tpus
if self.on_tpu:
self.init_tpu()
# init flags for SLURM+DDP to work
self.world_size = 1
self.interactive_ddp_procs = []
self.configure_slurm_ddp(self.num_nodes)
self.node_rank = self.determine_ddp_node_rank()
self.local_rank = self.determine_local_rank()
self.global_rank = 0
# NVIDIA setup
self.set_nvidia_flags(self.is_slurm_managing_tasks, self.data_parallel_device_ids)
# backward compatibility
if show_progress_bar is not None:
self.show_progress_bar = show_progress_bar
self._progress_bar_callback = self.configure_progress_bar(progress_bar_refresh_rate, process_position)
# logging
self.configure_logger(logger)
self.log_save_interval = log_save_interval
self.val_check_interval = val_check_interval
self.row_log_interval = row_log_interval
# how much of the data to use
# TODO: remove in 0.10.0
if overfit_pct is not None:
rank_zero_warn(
"Argument `overfit_pct` is now set by `overfit_batches` since v0.8.0"
" and this argument will be removed in v0.10.0",
DeprecationWarning,
)
overfit_batches = overfit_pct
# convert floats to ints
self.overfit_batches = _determine_limit_batches(overfit_batches)
# TODO: remove in 0.10.0
if val_percent_check is not None:
rank_zero_warn(
"Argument `val_percent_check` is now set by `limit_val_batches` since v0.8.0"
" and this argument will be removed in v0.10.0",
DeprecationWarning,
)
limit_val_batches = val_percent_check
# TODO: remove in 0.10.0
if test_percent_check is not None:
rank_zero_warn(
"Argument `test_percent_check` is now set by `limit_test_batches` since v0.8.0"
" and this argument will be removed in v0.10.0",
DeprecationWarning,
)
limit_test_batches = test_percent_check
# TODO: remove in 0.10.0
if train_percent_check is not None:
rank_zero_warn(
"Argument `train_percent_check` is now set by `limit_train_batches` since v0.8.0"
" and this argument will be removed in v0.10.0",
DeprecationWarning,
)
limit_train_batches = train_percent_check
self.limit_test_batches = _determine_limit_batches(limit_test_batches)
self.limit_val_batches = _determine_limit_batches(limit_val_batches)
self.limit_train_batches = _determine_limit_batches(limit_train_batches)
self.determine_data_use_amount(self.overfit_batches)
# AMP init
# These are the only lines needed after v0.8.0
# we wrap the user's forward with autocast and give it back at the end of fit
self.autocast_original_forward = None
self.precision = precision
self.scaler = None
# Backward compatibility, TODO: remove in v0.9.0
if use_amp is not None:
rank_zero_warn(
"Argument `use_amp` is now set by `precision` since v0.7.0"
" and this method will be removed in v0.9.0",
DeprecationWarning,
)
self.precision = 16 if use_amp else 32
self.amp_level = amp_level
self.init_amp()
self.on_colab_kaggle = os.getenv('COLAB_GPU') or os.getenv('KAGGLE_URL_BASE')
# tracks internal state for debugging
self.dev_debugger = InternalDebugger(self)
self.config_validator = ConfigValidator(self)
self.accelerator_backend = None
# Callback system
self.on_init_end()
@property
def is_global_zero(self) -> bool:
return self.global_rank == 0
@property
def slurm_job_id(self) -> Optional[int]:
try:
job_id = os.environ['SLURM_JOB_ID']
job_id = int(job_id)
# in interactive mode, don't make logs use the same job id
in_slurm_interactive_mode = os.environ['SLURM_JOB_NAME'] == 'bash'
if in_slurm_interactive_mode:
job_id = None
except Exception:
job_id = None
return job_id
@classmethod
def default_attributes(cls):
init_signature = inspect.signature(Trainer)
args = {}
for param_name in init_signature.parameters:
value = init_signature.parameters[param_name].default
args[param_name] = value
return args
@classmethod
def get_init_arguments_and_types(cls) -> List[Tuple[str, Tuple, Any]]:
r"""Scans the Trainer 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).
Examples:
>>> args = Trainer.get_init_arguments_and_types()
>>> import pprint
>>> pprint.pprint(sorted(args)) # doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
[('accumulate_grad_batches',
(<class 'int'>, typing.Dict[int, int], typing.List[list]),
1),
...
('callbacks',
(typing.List[pytorch_lightning.callbacks.base.Callback],
<class 'NoneType'>),
None),
('check_val_every_n_epoch', (<class 'int'>,), 1),
...
('max_epochs', (<class 'int'>,), 1000),
...
('precision', (<class 'int'>,), 32),
('prepare_data_per_node', (<class 'bool'>,), True),
('print_nan_grads', (<class 'bool'>,), False),
('process_position', (<class 'int'>,), 0),
('profiler',
(<class 'pytorch_lightning.profiler.profilers.BaseProfiler'>,
<class 'bool'>,
<class 'NoneType'>),
None),
...
"""
trainer_default_params = inspect.signature(cls).parameters
name_type_default = []
for arg in trainer_default_params:
arg_type = trainer_default_params[arg].annotation
arg_default = trainer_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
@classmethod
def get_deprecated_arg_names(cls) -> List:
"""Returns a list with deprecated Trainer arguments."""
depr_arg_names = []
for name, val in cls.__dict__.items():
if name.startswith('DEPRECATED') and isinstance(val, (tuple, list)):
depr_arg_names.extend(val)
return depr_arg_names
@classmethod
def add_argparse_args(cls, parent_parser: ArgumentParser) -> ArgumentParser:
r"""Extends existing argparse by default `Trainer` attributes.
Args:
parent_parser:
The custom cli arguments parser, which will be extended by
the Trainer default arguments.
Only arguments of the allowed types (str, float, int, bool) will
extend the `parent_parser`.
Examples:
>>> import argparse
>>> import pprint
>>> parser = argparse.ArgumentParser()
>>> parser = Trainer.add_argparse_args(parser)
>>> args = parser.parse_args([])
>>> pprint.pprint(vars(args)) # doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
{...
'check_val_every_n_epoch': 1,
'checkpoint_callback': True,
'default_root_dir': None,
'deterministic': False,
'distributed_backend': None,
'early_stop_callback': False,
...
'logger': True,
'max_epochs': 1000,
'max_steps': None,
'min_epochs': 1,
'min_steps': None,
...
'profiler': None,
'progress_bar_refresh_rate': 1,
...}
"""
parser = ArgumentParser(parents=[parent_parser], add_help=False,)
blacklist = ['kwargs']
depr_arg_names = cls.get_deprecated_arg_names() + blacklist
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 == 'gpus' or arg == 'tpu_cores':
use_type = Trainer._allowed_type
arg_default = Trainer._arg_default
parser.add_argument(
f'--{arg}',
dest=arg,
default=arg_default,
type=use_type,
help='autogenerated by pl.Trainer',
**arg_kwargs,
)
return parser
def _allowed_type(x) -> Union[int, str]:
if ',' in x:
return str(x)
else:
return int(x)
def _arg_default(x) -> Union[int, str]:
if ',' in x:
return str(x)
else:
return int(x)
@classmethod
def parse_argparser(cls, arg_parser: Union[ArgumentParser, Namespace]) -> Namespace:
"""Parse CLI arguments, required for custom bool types."""
args = arg_parser.parse_args() if isinstance(arg_parser, ArgumentParser) else arg_parser
types_default = {
arg: (arg_types, arg_default) for arg, arg_types, arg_default in cls.get_init_arguments_and_types()
}
modified_args = {}
for k, v in vars(args).items():
if k in types_default and v is None:
# We need to figure out if the None is due to using nargs="?" or if it comes from the default value
arg_types, arg_default = types_default[k]
if bool in arg_types and isinstance(arg_default, bool):
# Value has been passed as a flag => It is currently None, so we need to set it to True
# We always set to True, regardless of the default value.
# Users must pass False directly, but when passing nothing True is assumed.
# i.e. the only way to disable somthing that defaults to True is to use the long form:
# "--a_default_true_arg False" becomes False, while "--a_default_false_arg" becomes None,
# which then becomes True here.
v = True
modified_args[k] = v
return Namespace(**modified_args)
@classmethod
def from_argparse_args(cls, args: Union[Namespace, ArgumentParser], **kwargs) -> 'Trainer':
"""
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:`Trainer`.
**kwargs: Additional keyword arguments that may override ones in the parser or namespace.
These must be valid Trainer arguments.
Example:
>>> parser = ArgumentParser(add_help=False)
>>> parser = Trainer.add_argparse_args(parser)
>>> parser.add_argument('--my_custom_arg', default='something') # doctest: +SKIP
>>> args = Trainer.parse_argparser(parser.parse_args(""))
>>> trainer = Trainer.from_argparse_args(args, logger=False)
"""
if isinstance(args, ArgumentParser):
args = cls.parse_argparser(args)
params = vars(args)
# we only want to pass in valid Trainer args, the rest may be user specific
valid_kwargs = inspect.signature(cls.__init__).parameters
trainer_kwargs = dict((name, params[name]) for name in valid_kwargs if name in params)
trainer_kwargs.update(**kwargs)
return cls(**trainer_kwargs)
@property
def num_gpus(self) -> int:
gpus = self.data_parallel_device_ids
if gpus is None:
return 0
return len(gpus)
@property
def data_parallel(self) -> bool:
return self.use_dp or self.use_ddp or self.use_ddp2
@property
def progress_bar_callback(self):
return self._progress_bar_callback
@property
def progress_bar_dict(self) -> dict:
""" Read-only for progress bar metrics. """
ref_model = self.model if not self.data_parallel else self.model.module
return dict(**ref_model.get_progress_bar_dict(), **self.progress_bar_metrics)
@property
def disable_validation(self) -> bool:
""" Check if validation is disabled during training. """
return not self.enable_validation
@property
def enable_validation(self) -> bool:
""" Check if we should run validation during training. """
val_loop_enabled = self.is_overridden('validation_step') and self.limit_val_batches > 0
return val_loop_enabled or self.fast_dev_run
@property
def default_root_dir(self) -> str:
"""
The default location to save artifacts of loggers, checkpoints etc.
It is used as a fallback if logger or checkpoint callback do not define specific save paths.
"""
return os.path.normpath(self._default_root_dir)
@property
def weights_save_path(self) -> str:
"""
The default root location to save weights (checkpoints), e.g., when the
:class:`~pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint` does not define a file path.
"""
return os.path.normpath(self._weights_save_path)
# -----------------------------
# MODEL TRAINING
# -----------------------------
def fit(
self,
model: LightningModule,
train_dataloader: Optional[DataLoader] = None,
val_dataloaders: Optional[Union[DataLoader, List[DataLoader]]] = None,
datamodule: Optional[LightningDataModule] = None,
):
r"""
Runs the full optimization routine.
Args:
model: Model to fit.
train_dataloader: A Pytorch
DataLoader with training samples. If the model has
a predefined train_dataloader method this will be skipped.
val_dataloaders: Either a single
Pytorch Dataloader or a list of them, specifying validation samples.
If the model has a predefined val_dataloaders method this will be skipped
Example::
# Option 1,
# Define the train_dataloader() and val_dataloader() fxs
# in the lightningModule
# RECOMMENDED FOR MOST RESEARCH AND APPLICATIONS TO MAINTAIN READABILITY
trainer = Trainer()
model = LightningModule()
trainer.fit(model)
# Option 2
# in production cases we might want to pass different datasets to the same model
# Recommended for PRODUCTION SYSTEMS
train, val = DataLoader(...), DataLoader(...)
trainer = Trainer()
model = LightningModule()
trainer.fit(model, train_dataloader=train, val_dataloaders=val)
# Option 1 & 2 can be mixed, for example the training set can be
# defined as part of the model, and validation can then be feed to .fit()
"""
results = None
# bind logger and other properties
self.copy_trainer_model_properties(model)
# clean hparams
if hasattr(model, 'hparams'):
parsing.clean_namespace(model.hparams)
# if a datamodule comes in as the second arg, then fix it for the user
if isinstance(train_dataloader, LightningDataModule):
datamodule = train_dataloader
train_dataloader = None
self.config_validator.enforce_datamodule_dataloader_override(train_dataloader, val_dataloaders, datamodule)
# set up the passed in dataloaders (if needed)
self.__attach_dataloaders(model, train_dataloader, val_dataloaders)
self.__attach_datamodule(model, datamodule)
# check that model is configured correctly
self.config_validator.verify_loop_configurations(model)
# callbacks
self.on_fit_start()
if self.is_function_implemented('on_fit_start', model):
model.on_fit_start()
# on multi-gpu jobs we only want to manipulate (download, etc) on node_rank=0, local_rank=0
# or in the case where each node needs to do its own manipulation in which case just local_rank=0
if self.can_prepare_data():
model.prepare_data()
self._is_data_prepared = True
# Run auto batch size scaling
if self.auto_scale_batch_size:
if isinstance(self.auto_scale_batch_size, bool):
self.auto_scale_batch_size = 'power'
self.scale_batch_size(model, mode=self.auto_scale_batch_size)
model.logger = self.logger # reset logger binding
# Run learning rate finder:
if self.auto_lr_find:
self._run_lr_finder_internally(model)
model.logger = self.logger # reset logger binding
# route to appropriate start method
# when using multi-node or DDP within a node start each module in a separate process
if self.use_ddp2:
if self.is_slurm_managing_tasks:
task = int(os.environ['SLURM_LOCALID'])
# torchelastic or general non_slurm ddp2
elif 'WORLD_SIZE' in os.environ and ('GROUP_RANK' in os.environ or 'NODE_RANK' in os.environ):
task = int(os.environ['LOCAL_RANK'])
self.ddp_train(process_idx=task, mp_queue=None, model=model)
elif self.use_ddp:
# set testing if set in environ
self.testing = os.environ.get('PL_TESTING_MODE', self.testing)
if self.is_slurm_managing_tasks:
task = int(os.environ['SLURM_LOCALID'])
self.ddp_train(process_idx=task, mp_queue=None, model=model)
# torchelastic or general non_slurm ddp
elif 'WORLD_SIZE' in os.environ and ('GROUP_RANK' in os.environ or 'NODE_RANK' in os.environ):
task = int(os.environ['LOCAL_RANK'])
self.ddp_train(process_idx=task, mp_queue=None, model=model)
elif self.distributed_backend == 'ddp_cpu':
self.accelerator_backend = DDPSpawnBackend(self)
self.accelerator_backend.setup()
self.accelerator_backend.train(model, nprocs=self.num_processes)
results = self.accelerator_backend.teardown(model)
elif self.distributed_backend == 'ddp_spawn':
self.accelerator_backend = DDPSpawnBackend(self)
self.accelerator_backend.setup()
self.accelerator_backend.train(model, nprocs=self.num_processes)
results = self.accelerator_backend.teardown(model)
elif self.distributed_backend == 'ddp':
self.set_random_port()
results = self.spawn_ddp_children(model)
elif self.use_dp:
self.accelerator_backend = DataParallelBackend(self)
self.accelerator_backend.setup(model)
results = self.accelerator_backend.train()
self.accelerator_backend.teardown()
elif self.use_horovod:
results = self.horovod_train(model)
elif self.use_single_gpu:
self.accelerator_backend = GPUBackend(self)
model = self.accelerator_backend.setup(model)
results = self.accelerator_backend.train(model)
elif self.use_tpu:
self.accelerator_backend = TPUBackend(self)
self.accelerator_backend.setup()
self.accelerator_backend.train(model)
self.accelerator_backend.teardown(model)
else:
self.accelerator_backend = CPUBackend(self)
self.accelerator_backend.setup(model)
results = self.accelerator_backend.train(model)
# callbacks
self.on_fit_end()
# model hooks
if self.is_function_implemented('on_fit_end'):
model.on_fit_end()
self.teardown('fit')
if self.is_function_implemented('teardown'):
model.teardown('fit')
# return 1 when finished
# used for testing or when we need to know that training succeeded
return results or 1
def can_prepare_data(self):
if self.prepare_data_per_node:
return self.local_rank == 0
else:
return self.node_rank == 0 and self.local_rank == 0
def __attach_dataloaders(self, model, train_dataloader=None, val_dataloaders=None, test_dataloaders=None):
# when dataloader is passed via fit, patch the train_dataloader
# functions to overwrite with these implementations
if train_dataloader is not None:
model.train_dataloader = _PatchDataLoader(train_dataloader)
if val_dataloaders is not None:
model.val_dataloader = _PatchDataLoader(val_dataloaders)
if test_dataloaders is not None:
model.test_dataloader = _PatchDataLoader(test_dataloaders)
def __attach_datamodule(self, model, datamodule=None):
# We use datamodule if it's been provided on .fit or .test, otherwise we check model for it
datamodule = datamodule or getattr(model, 'datamodule', None)
# If we have a datamodule, attach necessary hooks + dataloaders
if datamodule:
if self.is_overridden('train_dataloader', datamodule):
model.train_dataloader = datamodule.train_dataloader
if self.is_overridden('val_dataloader', datamodule):
model.val_dataloader = datamodule.val_dataloader
if self.is_overridden('test_dataloader', datamodule):
model.test_dataloader = datamodule.test_dataloader
def run_pretrain_routine(self, model: LightningModule):
"""Sanity check a few things before starting actual training.
Args:
model: The model to run sanity test on.
"""
ref_model = model
if self.data_parallel:
ref_model = model.module
# give model convenience properties
ref_model.trainer = self
# set local properties on the model
self.copy_trainer_model_properties(ref_model)
# init amp. Must be done here instead of __init__ to allow ddp to work
if NATIVE_AMP_AVALAIBLE and self.precision == 16 and not self.use_tpu:
self.scaler = torch.cuda.amp.GradScaler()
# log hyper-parameters
if self.logger is not None:
# save exp to get started
self.logger.log_hyperparams(ref_model.hparams)
self.logger.save()
if self.use_ddp or self.use_ddp2:
torch_distrib.barrier()
# wait for all models to restore weights
if self.on_tpu and XLA_AVAILABLE:
# wait for all processes to catch up
torch_xla.core.xla_model.rendezvous("pl.Trainer.run_pretrain_routine")
elif self.use_horovod:
# wait for all processes to catch up
hvd.join()
# register auto-resubmit when on SLURM
self.register_slurm_signal_handlers()
# print model summary
if self.is_global_zero and self.weights_summary is not None and not self.testing:
if self.weights_summary in ModelSummary.MODES:
ref_model.summarize(mode=self.weights_summary)
else:
raise MisconfigurationException("weights_summary can be None, " + ", ".join(ModelSummary.MODES))
# track model now.
# if cluster resets state, the model will update with the saved weights
self.model = model
# restore training and model before hpc is called
self.restore_weights(model)
# when testing requested only run test and return
if self.testing:
# only load test dataloader for testing
# self.reset_test_dataloader(ref_model)
eval_loop_results, _ = self.run_evaluation(test_mode=True)
if len(eval_loop_results) == 0:
return 1
# remove the tensors from the eval results
for i, result in enumerate(eval_loop_results):
if isinstance(result, dict):
for k, v in result.items():
if isinstance(v, torch.Tensor):
result[k] = v.cpu().item()
return eval_loop_results
# run a few val batches before training starts
self._run_sanity_check(ref_model, model)
# clear cache before training
if self.on_gpu and self.root_gpu is not None:
# use context because of:
# https://discuss.pytorch.org/t/out-of-memory-when-i-use-torch-cuda-empty-cache/57898
with torch.cuda.device(f'cuda:{self.root_gpu}'):
torch.cuda.empty_cache()
# CORE TRAINING LOOP
self.train()
def _run_sanity_check(self, ref_model, model):
using_val_step = ref_model.val_dataloader is not None and self.is_overridden('validation_step')
should_sanity_check = using_val_step and self.num_sanity_val_steps > 0 and self.limit_val_batches > 0
# run tiny validation (if validation defined)
# to make sure program won't crash during val
if should_sanity_check:
self.reset_val_dataloader(ref_model)
# hook and callback
self.running_sanity_check = True
ref_model.on_sanity_check_start()
self.on_sanity_check_start()
num_loaders = len(self.val_dataloaders)
max_batches = [self.num_sanity_val_steps] * num_loaders
eval_results = self._evaluate(model, self.val_dataloaders, max_batches, False)
# allow no returns from eval
if eval_results is not None and len(eval_results) > 0:
# when we get a list back, used only the last item
if isinstance(eval_results, list):
eval_results = eval_results[-1]
if isinstance(eval_results, EvalResult):
callback_metrics = eval_results.callback_metrics
else:
_, _, _, callback_metrics, _ = self.process_output(eval_results)
self.callback_metrics = callback_metrics
self.on_sanity_check_end()
self.running_sanity_check = False
def test(
self,
model: Optional[LightningModule] = None,
test_dataloaders: Optional[Union[DataLoader, List[DataLoader]]] = None,
ckpt_path: Optional[str] = 'best',
verbose: bool = True,
datamodule: Optional[LightningDataModule] = None,
):
r"""
Separates from fit to make sure you never run on your test set until you want to.
Args:
model: The model to test.
test_dataloaders: Either a single
Pytorch Dataloader or a list of them, specifying validation samples.
ckpt_path: Either ``best`` or path to the checkpoint you wish to test.
If ``None``, use the weights from the last epoch to test. Default to ``best``.
verbose: If True, prints the test results
Returns:
The final test result dictionary. If no test_epoch_end is defined returns a list of dictionaries
Example::
# Option 1
# run test with the best checkpoint from ``ModelCheckpoint`` after fitting.
test = DataLoader(...)
trainer = Trainer()
model = LightningModule()
trainer.fit(model)
trainer.test(test_dataloaders=test)
# Option 2
# run test with the specified checkpoint after fitting
test = DataLoader(...)
trainer = Trainer()
model = LightningModule()
trainer.fit(model)
trainer.test(test_dataloaders=test, ckpt_path='path/to/checkpoint.ckpt')
# Option 3
# run test with the weights from the end of training after fitting
test = DataLoader(...)
trainer = Trainer()
model = LightningModule()
trainer.fit(model)
trainer.test(test_dataloaders=test, ckpt_path=None)
# Option 4
# run test from a loaded model. ``ckpt_path`` is ignored in this case.
test = DataLoader(...)
model = LightningModule.load_from_checkpoint('path/to/checkpoint.ckpt')
trainer = Trainer()
trainer.test(model, test_dataloaders=test)
"""
# --------------------
# SETUP HOOK
# --------------------
self.verbose_test = verbose
if self.global_rank != 0:
return
# If you supply a datamodule you can't supply train_dataloader or val_dataloaders
if test_dataloaders and datamodule:
raise MisconfigurationException(
'You cannot pass test_dataloaders to trainer.test if you supply a datamodule'
)
# Attach datamodule to get setup/prepare_data added to model before the call to it below
self.__attach_datamodule(model or self.get_model(), datamodule)
self.setup('test')
if model is not None:
results = self.__test_given_model(model, test_dataloaders)
else:
results = self.__test_using_best_weights(ckpt_path, test_dataloaders)
self.teardown('test')
return results
def __test_using_best_weights(self, ckpt_path, test_dataloaders):
model = self.get_model()
model.setup('test')
# if user requests the best checkpoint but we don't have it, error
if ckpt_path == 'best' and self.checkpoint_callback.save_top_k <= 0:
raise MisconfigurationException(
'ckpt_path is "best", but ModelCheckpoint is not configured to save the best model.'
)
# load best weights
if ckpt_path is not None:
# ckpt_path is 'best' so load the best model
if ckpt_path == 'best':
ckpt_path = self.checkpoint_callback.best_model_path
if len(ckpt_path) == 0:
rank_zero_warn(
f'.test() found no path for the best weights, {ckpt_path}. Please '
f'specify a path for a checkpoint .test(ckpt_path=PATH)'
)
return {}
ckpt = torch.load(ckpt_path, map_location=lambda storage, loc: storage)
model.load_state_dict(ckpt['state_dict'])
# attach dataloaders
if test_dataloaders is not None:
self.__attach_dataloaders(model, test_dataloaders=test_dataloaders)
# run tests
self.tested_ckpt_path = ckpt_path
self.set_random_port(force=True)
self.testing = True
os.environ['PL_TESTING_MODE'] = '1'
self.model = model
results = self.fit(model)
self.testing = False
del os.environ['PL_TESTING_MODE']
# teardown
if self.is_function_implemented('teardown'):
model_ref = self.get_model()
model_ref.teardown('test')
return results
def __test_given_model(self, model, test_dataloaders):
# setup hook
model.setup('test')
# attach data
if test_dataloaders is not None:
self.__attach_dataloaders(model, test_dataloaders=test_dataloaders)
# run test
# sets up testing so we short circuit to eval
self.set_random_port(force=True)
self.testing = True
self.model = model
results = self.fit(model)
self.testing = False
# teardown
if self.is_function_implemented('teardown'):
model.teardown('test')
return results
def barrier(self, name):
if self.use_ddp or self.use_ddp2:
pass
# torch_distrib.barrier()
if self.on_tpu and XLA_AVAILABLE:
# wait for all processes to catch up
torch_xla.core.xla_model.rendezvous(f'pl.Trainer.{name}')
class _PatchDataLoader(object):
r"""
Callable object for patching dataloaders passed into trainer.fit().
Use this class to override model.*_dataloader() and be pickle-compatible.
Args:
dataloader: Dataloader object to return when called.
"""
def __init__(self, dataloader: Union[List[DataLoader], DataLoader]):
self.dataloader = dataloader
# cannot pickle __code__ so cannot verify if PatchDataloader
# exists which shows dataloader methods have been overwritten.
# so, we hack it by using the string representation
self.patch_loader_code = str(self.__call__.__code__)
def __call__(self) -> Union[List[DataLoader], DataLoader]:
return self.dataloader
def _determine_limit_batches(batches: Union[int, float]) -> Union[int, float]:
if 0 <= batches <= 1:
return batches
elif batches > 1 and batches % 1.0 == 0:
return int(batches)
else:
raise MisconfigurationException(
f'You have passed invalid value {batches}, it has to be in (0, 1) or nature number.'
)