import logging
from typing import Any, Callable, Dict, List, Optional, Type, Union
import ray
from ray._private.thirdparty.tabulate.tabulate import tabulate
from ray.air.config import RunConfig, ScalingConfig
from ray.train import BackendConfig, Checkpoint, TrainingIterator
from ray.train._internal import session
from ray.train._internal.backend_executor import BackendExecutor, TrialInfo
from ray.train._internal.data_config import DataConfig
from ray.train._internal.session import _TrainingResult, get_session
from ray.train._internal.utils import construct_train_func, count_required_parameters
from ray.train.trainer import BaseTrainer, GenDataset
from ray.util.annotations import DeveloperAPI, PublicAPI
from ray.widgets import Template
from ray.widgets.util import repr_with_fallback
logger = logging.getLogger(__name__)
[docs]@DeveloperAPI
class DataParallelTrainer(BaseTrainer):
"""A Trainer for data parallel training.
You should subclass this Trainer if your Trainer follows SPMD (single program,
multiple data) programming paradigm - you want multiple processes to run the same
function, but on different data.
This Trainer runs the function ``train_loop_per_worker`` on multiple Ray
Actors.
The ``train_loop_per_worker`` function is expected to take in either 0 or 1
arguments:
.. testcode::
def train_loop_per_worker():
...
.. testcode::
def train_loop_per_worker(config: Dict):
...
If ``train_loop_per_worker`` accepts an argument, then
``train_loop_config`` will be passed in as the argument. This is useful if you
want to tune the values in ``train_loop_config`` as hyperparameters.
If the ``datasets`` dict contains a training dataset (denoted by
the "train" key), then it will be split into multiple dataset
shards that can then be accessed by ``train.get_dataset_shard("train")`` inside
``train_loop_per_worker``. All the other datasets will not be split and
``train.get_dataset_shard(...)`` will return the the entire Dataset.
Inside the ``train_loop_per_worker`` function, you can use any of the
:ref:`Ray Train loop methods <train-loop-api>`.
.. testcode::
from ray import train
def train_loop_per_worker():
# Report intermediate results for callbacks or logging and
# checkpoint data.
train.report(...)
# Returns dict of last saved checkpoint.
train.get_checkpoint()
# Returns the Dataset shard for the given key.
train.get_dataset_shard("my_dataset")
# Returns the total number of workers executing training.
train.get_context().get_world_size()
# Returns the rank of this worker.
train.get_context().get_world_rank()
# Returns the rank of the worker on the current node.
train.get_context().get_local_rank()
Any returns from the ``train_loop_per_worker`` will be discarded and not
used or persisted anywhere.
**How do I use DataParallelTrainer or any of its subclasses?**
Example:
.. testcode::
import ray
from ray import train
from ray.train import ScalingConfig
from ray.train.data_parallel_trainer import DataParallelTrainer
def train_loop_for_worker():
dataset_shard_for_this_worker = train.get_dataset_shard("train")
# 3 items for 3 workers, each worker gets 1 item
batches = list(dataset_shard_for_this_worker.iter_batches(batch_size=1))
assert len(batches) == 1
train_dataset = ray.data.from_items([1, 2, 3])
assert train_dataset.count() == 3
trainer = DataParallelTrainer(
train_loop_for_worker,
scaling_config=ScalingConfig(num_workers=3),
datasets={"train": train_dataset},
)
result = trainer.fit()
.. testoutput::
:hide:
...
**How do I develop on top of DataParallelTrainer?**
In many cases, using DataParallelTrainer directly is sufficient to execute
functions on multiple actors.
However, you may want to subclass ``DataParallelTrainer`` and create a custom
Trainer for the following 2 use cases:
- **Use Case 1:** You want to do data parallel training, but want to have
a predefined ``training_loop_per_worker``.
- **Use Case 2:** You want to implement a custom
:py:class:`~ray.train.backend.Backend` that automatically handles
additional setup or teardown logic on each actor, so that the users of this
new trainer do not have to implement this logic. For example, a
``TensorflowTrainer`` can be built on top of ``DataParallelTrainer``
that automatically handles setting the proper environment variables for
distributed Tensorflow on each actor.
For 1, you can set a predefined training loop in __init__
.. testcode::
from ray.train.data_parallel_trainer import DataParallelTrainer
class MyDataParallelTrainer(DataParallelTrainer):
def __init__(self, *args, **kwargs):
predefined_train_loop_per_worker = lambda: 1
super().__init__(predefined_train_loop_per_worker, *args, **kwargs)
For 2, you can implement the ``ray.train.Backend`` and ``ray.train.BackendConfig``
interfaces.
.. testcode::
from dataclasses import dataclass
from ray.train.backend import Backend, BackendConfig
class MyBackend(Backend):
def on_start(self, worker_group, backend_config):
def set_env_var(env_var_value):
import os
os.environ["MY_ENV_VAR"] = env_var_value
worker_group.execute(set_env_var, backend_config.env_var)
@dataclass
class MyBackendConfig(BackendConfig):
env_var: str = "default_value"
def backend_cls(self):
return MyBackend
class MyTrainer(DataParallelTrainer):
def __init__(self, train_loop_per_worker, my_backend_config:
MyBackendConfig, **kwargs):
super().__init__(
train_loop_per_worker,
backend_config=my_backend_config, **kwargs)
Args:
train_loop_per_worker: The training function to execute.
This can either take in no arguments or a ``config`` dict.
train_loop_config: Configurations to pass into
``train_loop_per_worker`` if it accepts an argument.
backend_config: Configuration for setting up a Backend (e.g. Torch,
Tensorflow, Horovod) on each worker to enable distributed
communication. If no Backend should be set up, then set this to None.
scaling_config: Configuration for how to scale data parallel training.
dataset_config: Configuration for dataset ingest. This is merged with the
default dataset config for the given trainer (`cls._dataset_config`).
run_config: Configuration for the execution of the training run.
datasets: Ray Datasets to use for training and evaluation.
This is a dict where the key is the name of the dataset, which
can be accessed from within the ``train_loop_per_worker`` by calling
``train.get_dataset_shard(dataset_key)``.
By default, all datasets are sharded equally across workers.
This can be configured via ``dataset_config``.
metadata: Dict that should be made available via
`train.get_context().get_metadata()` and in `checkpoint.get_metadata()`
for checkpoints saved from this Trainer. Must be JSON-serializable.
resume_from_checkpoint: A checkpoint to resume training from.
"""
# Exposed here for testing purposes. Should never need
# to be overriden.
_backend_executor_cls: Type[BackendExecutor] = BackendExecutor
_training_iterator_cls: Type[TrainingIterator] = TrainingIterator
_scaling_config_allowed_keys = BaseTrainer._scaling_config_allowed_keys + [
"num_workers",
"resources_per_worker",
"use_gpu",
"placement_strategy",
"accelerator_type",
]
# For backwards compatibility with the legacy dataset config API.
_dataset_config = None
_fields_for_tuner_param_space = BaseTrainer._fields_for_tuner_param_space + [
"train_loop_config"
]
def __init__(
self,
train_loop_per_worker: Union[Callable[[], None], Callable[[Dict], None]],
*,
train_loop_config: Optional[Dict] = None,
backend_config: Optional[BackendConfig] = None,
scaling_config: Optional[ScalingConfig] = None,
dataset_config: Optional[DataConfig] = None,
run_config: Optional[RunConfig] = None,
datasets: Optional[Dict[str, GenDataset]] = None,
metadata: Optional[Dict[str, Any]] = None,
resume_from_checkpoint: Optional[Checkpoint] = None,
):
self._train_loop_per_worker = train_loop_per_worker
self._train_loop_config = train_loop_config
if dataset_config is None:
dataset_config = DataConfig()
if not isinstance(dataset_config, DataConfig):
raise ValueError(
"`dataset_config` must be an instance of ray.train.DataConfig, "
f"was: {dataset_config}"
)
self._data_config = dataset_config
backend_config = (
backend_config if backend_config is not None else BackendConfig()
)
self._backend_config = backend_config
super(DataParallelTrainer, self).__init__(
scaling_config=scaling_config,
run_config=run_config,
datasets=datasets,
metadata=metadata,
resume_from_checkpoint=resume_from_checkpoint,
)
train_total_resources = self.scaling_config.total_resources
self._data_config.set_train_total_resources(
train_total_resources.get("CPU", 0),
train_total_resources.get("GPU", 0),
)
[docs] @PublicAPI(stability="beta")
@classmethod
def restore(
cls: Type["DataParallelTrainer"],
path: str,
train_loop_per_worker: Optional[
Union[Callable[[], None], Callable[[Dict], None]]
] = None,
train_loop_config: Optional[Dict] = None,
**kwargs,
) -> "DataParallelTrainer":
"""Restores a DataParallelTrainer from a previously interrupted/failed run.
Args:
train_loop_per_worker: Optionally re-specified train loop function.
This should be used to re-specify a function that is not
restorable in a new Ray cluster (e.g., it holds onto outdated
object references). This should be the same training loop
that was passed to the original trainer constructor.
train_loop_config: Optionally re-specified train config.
This should similarly be used if the original `train_loop_config`
contained outdated object references, and it should not be modified
from what was originally passed in.
See :meth:`BaseTrainer.restore() <ray.train.trainer.BaseTrainer.restore>`
for descriptions of the other arguments.
Returns:
DataParallelTrainer: A restored instance of the `DataParallelTrainer`
subclass that is calling this method.
"""
return super(DataParallelTrainer, cls).restore(
path=path,
train_loop_per_worker=train_loop_per_worker,
train_loop_config=train_loop_config,
**kwargs,
)
def _validate_attributes(self):
super()._validate_attributes()
self._validate_train_loop_per_worker(
self._train_loop_per_worker, "train_loop_per_worker"
)
def _validate_train_loop_per_worker(
self, train_loop_per_worker: Callable, fn_name: str
) -> None:
num_required_params = count_required_parameters(train_loop_per_worker)
if num_required_params > 1:
raise ValueError(
f"{fn_name} should take in 0 or 1 arguments, "
f"but it accepts {num_required_params} arguments instead."
)
@classmethod
def _validate_scaling_config(cls, scaling_config: ScalingConfig) -> ScalingConfig:
scaling_config = super(DataParallelTrainer, cls)._validate_scaling_config(
scaling_config
)
# This validation happens after the scaling config is updated from
# its specification in the Tuner `param_space`
if not scaling_config.use_gpu and "GPU" in ray.available_resources():
logger.info(
"GPUs are detected in your Ray cluster, but GPU "
"training is not enabled for this trainer. To enable "
"GPU training, make sure to set `use_gpu` to True "
"in your scaling config."
)
if scaling_config.num_workers is None:
raise ValueError(
"You must specify the 'num_workers' in `scaling_config` as either an "
f"argument of `{cls.__name__}` or through the `param_space` of a "
"`Tuner` (if performing hyperparameter tuning)."
)
if scaling_config.num_workers <= 0:
raise ValueError(
"'num_workers' in `scaling_config` must be a positive "
f"integer. Received {scaling_config.num_workers}"
)
return scaling_config
def _run_training(self, training_iterator: TrainingIterator) -> None:
"""This method loops over the `TrainingIterator`:
The actual iteration (for ... in ...) waits for the training function
on each worker to report a result and supplies it as a list of results.
Afterwards (in the body of the loop), it will report the result
to the Tune session.
The iterator ends after the training function on each worker has finished.
"""
for training_results in training_iterator:
# TODO(ml-team): add ability to report results from multiple workers.
self._propagate_results(training_results)
def _propagate_results(self, training_results: List[_TrainingResult]):
first_worker_result = training_results[0]
assert all(isinstance(result, _TrainingResult) for result in training_results)
tune_session = get_session()
# Check if any workers reported a checkpoint.
# If so, report a checkpoint pointing to the persisted location
# to Tune for book-keeping.
# NOTE: This removes the restriction for any individual worker
# (ex: global rank 0 worker) from needing to report a checkpoint.
# All workers reported a checkpoint to the same fs path, so there's
# no need to report multiple checkpoints to Tune.
worker_checkpoints = [
result.checkpoint
for result in training_results
if result.checkpoint is not None
]
at_least_one_reported_checkpoint = len(worker_checkpoints) > 0
if at_least_one_reported_checkpoint:
# Update the coordinator's checkpoint index to the latest.
# This is what keeps the checkpoint index in line with the workers.
tune_session.storage._update_checkpoint_index(first_worker_result.metrics)
# Make sure that all workers uploaded to the same location.
assert all(
checkpoint.path == tune_session.storage.checkpoint_fs_path
for checkpoint in worker_checkpoints
)
checkpoint = (
Checkpoint(
filesystem=tune_session.storage.storage_filesystem,
path=tune_session.storage.checkpoint_fs_path,
)
if at_least_one_reported_checkpoint
else None
)
tracked_training_result = _TrainingResult(
checkpoint=checkpoint,
metrics=first_worker_result.metrics,
)
logger.debug(
"Report (metrics, checkpoint) to the Tune session:\n"
f" metrics={tracked_training_result.metrics}\n"
f" checkpoint={tracked_training_result.checkpoint}"
)
# Report the metrics and checkpoint to Tune.
tune_session._report_training_result(tracked_training_result)
def training_loop(self) -> None:
scaling_config = self._validate_scaling_config(self.scaling_config)
train_loop_per_worker = construct_train_func(
self._train_loop_per_worker,
self._train_loop_config,
train_func_context=self._backend_config.train_func_context,
fn_arg_name="train_loop_per_worker",
discard_returns=True,
)
trial_info = TrialInfo(
name=session.get_trial_name(),
id=session.get_trial_id(),
resources=session.get_trial_resources(),
logdir=session.get_trial_dir(),
driver_ip=ray.util.get_node_ip_address(),
experiment_name=session.get_experiment_name(),
)
backend_executor = self._backend_executor_cls(
backend_config=self._backend_config,
trial_info=trial_info,
num_workers=scaling_config.num_workers,
resources_per_worker=scaling_config._resources_per_worker_not_none,
max_retries=0,
)
# Start the remote actors.
backend_executor.start()
training_iterator = self._training_iterator_cls(
backend_executor=backend_executor,
backend_config=self._backend_config,
train_func=train_loop_per_worker,
datasets=self.datasets,
metadata=self.metadata,
data_config=self._data_config,
checkpoint=self.starting_checkpoint,
)
self._run_training(training_iterator)
# Shutdown workers.
backend_executor.shutdown()
[docs] def get_dataset_config(self) -> DataConfig:
"""Returns a copy of this Trainer's final dataset configs.
Returns:
The merged default + user-supplied dataset config.
"""
return self._data_config
@repr_with_fallback(["ipywidgets", "8"])
def _repr_mimebundle_(self, **kwargs):
"""Returns a mimebundle with an ipywidget repr and a simple text repr.
Depending on the frontend where the data is being displayed,
different mimetypes will be used from this bundle.
See https://ipython.readthedocs.io/en/stable/config/integrating.html
for information about this method, and
https://ipywidgets.readthedocs.io/en/latest/embedding.html
for more information about the jupyter widget mimetype.
Returns:
A mimebundle containing an ipywidget repr and a simple text repr.
"""
from ipywidgets import HTML, Layout, Tab, VBox
title = HTML(f"<h2>{self.__class__.__name__}</h2>")
children = []
titles = []
if self.datasets:
children.append(self._datasets_repr_())
titles.append("Datasets")
children.append(HTML(self._data_config_repr_html_()))
titles.append("Data Config")
if self._train_loop_config:
children.append(HTML(self._train_loop_config_repr_html_()))
titles.append("Train Loop Config")
if self.scaling_config:
children.append(HTML(self.scaling_config._repr_html_()))
titles.append("Scaling Config")
if self.run_config:
children.append(HTML(self.run_config._repr_html_()))
titles.append("Run Config")
if self._backend_config:
children.append(HTML(self._backend_config._repr_html_()))
titles.append("Backend Config")
tab = Tab(children, titles=titles)
widget = VBox([title, tab], layout=Layout(width="100%"))
bundle = widget._repr_mimebundle_(**kwargs)
bundle.update(
{
"text/plain": repr(self),
}
)
return bundle
def _train_loop_config_repr_html_(self) -> str:
if self._train_loop_config:
table_data = {}
for k, v in self._train_loop_config.items():
if isinstance(v, str) or str(v).isnumeric():
table_data[k] = v
elif hasattr(v, "_repr_html_"):
table_data[k] = v._repr_html_()
else:
table_data[k] = str(v)
return Template("title_data.html.j2").render(
title="Train Loop Config",
data=Template("scrollableTable.html.j2").render(
table=tabulate(
table_data.items(),
headers=["Setting", "Value"],
showindex=False,
tablefmt="unsafehtml",
),
max_height="none",
),
)
else:
return ""
def _data_config_repr_html_(self) -> str:
# TODO make this rendering nicer.
content = [str(self._data_config)]
return Template("rendered_html_common.html.j2").render(content=content)
def _datasets_repr_(self) -> str:
from ipywidgets import HTML, Layout, VBox
content = []
if self.datasets:
for name, config in self.datasets.items():
tab = config._tab_repr_()
if tab:
content.append(
HTML(
Template("title_data.html.j2").render(
title=f"Dataset - <code>{name}</code>", data=None
)
)
)
content.append(config._tab_repr_())
return VBox(content, layout=Layout(width="100%"))