def shutdown(self, force=False):
"""Shuts down workers and releases resources.
Args:
force (bool): If True, forcefully kill all workers. If False,
attempt a graceful shutdown first, and then forcefully kill if
unsuccessful.
"""
self.worker_group.shutdown(force=force)
self.worker_group = DeactivatedWorkerGroup()
def _start_workers(self, num_workers):
worker_config = self.config.copy()
batch_size_per_worker = self._configure_and_split_batch(num_workers)
if batch_size_per_worker:
worker_config[BATCH_SIZE] = batch_size_per_worker
params = dict(
training_operator_cls=self.training_operator_cls,
config=worker_config,
serialize_data_creation=self.serialize_data_creation,
use_fp16=self.use_fp16,
use_gpu=self.use_gpu,
use_tqdm=self.use_tqdm,
scheduler_step_freq=self.scheduler_step_freq,
)
dist_params = dict(
backend=self.backend,
add_dist_sampler=self.add_dist_sampler,
wrap_ddp=self.wrap_ddp,
)
worker_args = {
"max_workers": self.max_replicas,
"params": params,
"dist_params": dist_params,
"initialization_hook": self.initialization_hook,
"num_cpus_per_worker": self.num_cpus_per_worker,
"use_gpu": self.use_gpu,
"timeout_s": self.timeout_s,
}
if self.use_local:
self.worker_group = LocalWorkerGroup(**worker_args)
else:
self.worker_group = RemoteWorkerGroup(**worker_args)
# TODO(amogkam): If not enough resources are available to create
# num_workers workers, this command will hang. Instead,
# start_workers should take into account available resources when
# determining how many workers to create.
return self.worker_group.start_workers(num_workers)
class TorchTrainer:
"""Train a PyTorch model using distributed PyTorch.
Launches a set of actors which connect via distributed PyTorch and
coordinate gradient updates to train the provided model. If Ray is not
initialized, TorchTrainer will automatically initialize a local Ray
cluster for you. Be sure to run `ray.init(address="auto")` to leverage
multi-node training.
.. code-block:: python
class MyTrainingOperator(TrainingOperator):
def setup(self, config):
model = nn.Linear(1, 1)
optimizer = torch.optim.SGD(
model.parameters(), lr=config.get("lr", 1e-4))
loss = torch.nn.MSELoss()
batch_size = config["batch_size"]
train_data, val_data = LinearDataset(2, 5), LinearDataset(2, 5)
train_loader = DataLoader(train_data, batch_size=batch_size)
val_loader = DataLoader(val_data, batch_size=batch_size)
self.model, self.optimizer = self.register(
models=model,
optimizers=optimizer,
criterion=loss)
self.register_data(
train_loader=train_loader,
validation_loader=val_loader)
trainer = TorchTrainer(
training_operator_cls=MyTrainingOperator,
config={"batch_size": 32},
use_gpu=True
)
for i in range(4):
trainer.train()
Args:
training_operator_cls (type): Custom training operator class
that subclasses the TrainingOperator class. This class
will be copied onto all remote workers and used to specify
training components and custom training and validation operations.
initialization_hook (function): A function to call on all training
workers when they are first initialized. This could be useful to
set environment variables for all the worker processes.
config (dict): Custom configuration value to be passed to
all operator constructors.
num_workers (int): the number of workers used in distributed
training. If 1, the worker will not be wrapped with
DistributedDataParallel. TorchTrainer will scale down the number
of workers if enough resources are not available, and will scale
back up once they are. The total number of
workers will never exceed `num_workers` amount.
num_cpus_per_worker (int): Sets the cpu requirement for each worker.
use_gpu (bool): Sets resource allocation for workers to 1 GPU
if true, and automatically moves both the model and optimizer
to the available CUDA device.
backend (string): backend used by distributed PyTorch. Currently
support "nccl", "gloo", and "auto". If "auto", RaySGD will
automatically use "nccl" if `use_gpu` is True, and "gloo"
otherwise.
wrap_ddp (bool): Whether to automatically wrap DistributedDataParallel
over each model. If False, you are expected to call it yourself.
timeout_s (float): Seconds before the torch process group
times out. Useful when machines are unreliable.
add_dist_sampler (bool): Whether to automatically add a
DistributedSampler to all created dataloaders. Only applicable
if num_workers > 1.
use_fp16 (bool): Enables mixed precision training via apex if apex
is installed. This is automatically done after the model and
optimizers are constructed and will work for multi-model training.
Please see https://github.com/NVIDIA/apex for more details.
apex_args (dict|None): Dict containing keyword args for amp.initialize.
See https://nvidia.github.io/apex/amp.html#module-apex.amp. By
default, the models and optimizers are passed in. Consider using
"num_losses" if operating over multiple models and optimizers.
scheduler_step_freq: "batch", "epoch", "manual", or None. This will
determine when ``scheduler.step`` is called. If "batch",
``step`` will be called after every optimizer step. If "epoch",
``step`` will be called after one pass of the DataLoader. If
"manual", the scheduler will not be incremented automatically -
you are expected to call ``trainer.update_scheduler`` manually.
If a scheduler is passed in, this value is expected to not be None.
use_local (bool): If True, 1 worker will be a local worker running
on the driver process, and all other workers will be remote. If
False, all workers will be remote. Set this to True for easy
debugging of worker on driver process, but could also
lead to issues with Cuda devices. Defaults to False.
"""
# TODO: Implement autoscaling. If num_workers=-1, the trainer will use as
# many resources as available. Upon each train call, TorchTrainer will
# query the Ray global state for total available resources and resize
# its remote workers to consume all available resources.
def __init__(
self,
*,
training_operator_cls,
initialization_hook=None,
config=None,
num_workers=1,
num_cpus_per_worker=1,
use_gpu="auto",
backend="auto",
wrap_ddp=True,
timeout_s=NCCL_TIMEOUT_S,
use_fp16=False,
use_tqdm=False,
apex_args=None,
add_dist_sampler=True,
scheduler_step_freq=None,
use_local=False,
# Deprecated Args.
num_replicas=None,
batch_size=None,
model_creator=None,
data_creator=None,
optimizer_creator=None,
scheduler_creator=None,
loss_creator=None,
serialize_data_creation=None,
data_loader_args=None,
):
if (model_creator or data_creator or optimizer_creator
or scheduler_creator or loss_creator):
raise DeprecationWarning(
"Creator functions are deprecated. You should create a "
"custom TrainingOperator, override setup, and register all "
"training state there. See TrainingOperator for more info. "
"If you would still like to use creator functions, you can "
"do CustomOperator = TrainingOperator.from_creators("
"model_creator, ...) and pass in CustomOperator into "
"TorchTrainer.")
if use_local and log_once("use_local"):
logger.warning("use_local is set to True. This could lead to "
"issues with Cuda devices. If you are seeing this "
"issue, try setting use_local to False. For more "
"information, see "
"https://github.com/ray-project/ray/issues/9202.")
if num_workers > 1 and not dist.is_available():
raise ValueError(
("Distributed PyTorch is not supported on macOS. "
"To run without distributed PyTorch, set 'num_workers=1'. "
"For more information, see "
"https://github.com/pytorch/examples/issues/467."))
if num_replicas is not None:
raise DeprecationWarning(
"num_replicas is deprecated. Use num_workers instead.")
if batch_size is not None:
raise DeprecationWarning(
"batch_size is deprecated. Use config={'batch_size': N} "
"specify a batch size for each worker or "
"config={ray.util.sgd.utils.BATCH_SIZE: N} to specify a "
"batch size to be used across all workers.")
if serialize_data_creation is True:
if log_once("serialize_data_creation"):
logging.warning(
"serialize_data_creation is deprecated and will be "
"ignored. If you require serialized data loading you "
"should implement this in TrainingOperator.setup. "
"You may find FileLock useful here.")
if data_loader_args:
raise DeprecationWarning(
"data_loader_args is deprecated. You can return a "
"torch.utils.data.DataLoader in data_creator. Ray will "
"automatically set a DistributedSampler if a DataLoader is "
"returned and num_workers > 1.")
self.training_operator_cls = training_operator_cls
self.initialization_hook = initialization_hook
self.config = {} if config is None else config
if use_gpu == "auto":
use_gpu = torch.cuda.is_available()
_remind_gpu_usage(use_gpu)
if backend == "auto":
backend = "nccl" if use_gpu else "gloo"
logger.debug(f"Using {backend} as backend.")
self.backend = backend
self.num_cpus_per_worker = num_cpus_per_worker
self.use_gpu = use_gpu
self.max_replicas = num_workers
self.serialize_data_creation = serialize_data_creation
self.wrap_ddp = wrap_ddp
self.timeout_s = timeout_s
self.use_fp16 = use_fp16
self.use_tqdm = use_tqdm
self.add_dist_sampler = add_dist_sampler
self.use_local = use_local
if apex_args and not isinstance(apex_args, dict):
raise ValueError("apex_args needs to be a dict object.")
self.apex_args = apex_args
self.temp_dir = tempfile.mkdtemp(prefix="raysgd")
self._num_failures = 0
self._last_resize = float("-inf")
if scheduler_step_freq:
_validate_scheduler_step_freq(scheduler_step_freq)
self.scheduler_step_freq = scheduler_step_freq
if not ray.is_initialized() and self.max_replicas > 1:
logger.info("Automatically initializing single-node Ray. To use "
"multi-node training, be sure to run `ray.init("
"address='auto')` before instantiating the Trainer.")
ray.init()
self._start_workers(self.max_replicas)
def _configure_and_split_batch(self, num_workers):
"""If sgd.utils.BATCH_SIZE is provided, split among workers."""
if BATCH_SIZE not in self.config:
return
# Compute batch size per worker
logger.debug("BATCH_SIZE parameter detected. Splitting among workers.")
batch_size = self.config[BATCH_SIZE]
batch_size_per_worker = batch_size // num_workers
if batch_size % num_workers > 0:
new_batch_size = batch_size_per_worker * num_workers
logger.warning(
("Changing batch size from {old_batch_size} to "
"{new_batch_size} to evenly distribute batches across "
"{num_workers} workers.").format(
old_batch_size=batch_size,
new_batch_size=new_batch_size,
num_workers=num_workers))
self.config[BATCH_SIZE] = new_batch_size
return batch_size_per_worker
def _start_workers(self, num_workers):
worker_config = self.config.copy()
batch_size_per_worker = self._configure_and_split_batch(num_workers)
if batch_size_per_worker:
worker_config[BATCH_SIZE] = batch_size_per_worker
params = dict(training_operator_cls=self.training_operator_cls,
config=worker_config,
serialize_data_creation=self.serialize_data_creation,
use_fp16=self.use_fp16,
use_gpu=self.use_gpu,
use_tqdm=self.use_tqdm,
apex_args=self.apex_args,
scheduler_step_freq=self.scheduler_step_freq)
dist_params = dict(backend=self.backend,
add_dist_sampler=self.add_dist_sampler,
wrap_ddp=self.wrap_ddp)
worker_args = {
"max_workers": num_workers,
"params": params,
"dist_params": dist_params,
"initialization_hook": self.initialization_hook,
"num_cpus_per_worker": self.num_cpus_per_worker,
"use_gpu": self.use_gpu,
"timeout_s": self.timeout_s
}
if self.use_local:
self.worker_group = LocalWorkerGroup(**worker_args)
else:
self.worker_group = RemoteWorkerGroup(**worker_args)
# TODO(amogkam): If not enough resources are available to create
# num_workers workers, this command will hang. Instead,
# start_workers should take into account available resources when
# determining how many workers to create.
self.worker_group.start_workers(num_workers)
def _resize_worker_group(self, max_retries=10):
"""Resizes the number of remote workers based on available resources.
Total number of workers will never exceed `num_workers` amount.
Args:
max_retries (int): How many times to attempt to resize workers
before failing.
"""
state_dict = self.state_dict()
old_workers = self.worker_group.num_workers
self.worker_group.reset()
time.sleep(1)
for i in range(max_retries):
new_workers = self.worker_group.new_workers_size()
if new_workers:
self._last_resize = time.time()
self._start_workers(int(new_workers))
self.load_state_dict(state_dict, blocking=True)
if self.use_local and new_workers == 1 and old_workers > 1:
# Major hack. If we go from LocalDistributedRunner to a
# standard TorchRunner we have to manually reset the
# dummy actor handle global vars.
# TODO(amog): Refactor LocalDistributedTorchRunner to
# not use global variables for resource reservation.
ray.util.sgd.torch.distributed_torch_runner\
._dummy_cuda_actor = None
ray.util.sgd.torch.distributed_torch_runner\
._dummy_cpu_actor = None
return
else:
delay = 2**i
logger.warning("No new workers found. Retrying in %d sec." %
delay)
time.sleep(delay)
raise RuntimeError("Exceeded max_retries for relaunching workers.")
def train(self,
num_steps=None,
profile=False,
reduce_results=True,
max_retries=3,
info=None,
dataset=None):
"""Runs a training epoch.
Calls `operator.train_epoch()` on N parallel workers simultaneously
underneath the hood.
Set `max_retries` to enable fault handling in case of
instance preemption.
Args:
num_steps (int): Number of batches to compute update steps on
per worker. This corresponds also to the number of times
``TrainingOperator.train_batch`` is called per worker.
profile (bool): Returns time stats for the training procedure.
reduce_results (bool): Whether to average all metrics across
all workers into one dict. If a metric is a non-numerical
value (or nested dictionaries), one value will be randomly
selected among the workers. If False, returns a list of dicts.
max_retries (int): Must be non-negative. If set to N, TorchTrainer
will detect and recover from training failure. The recovery
process will kill all current workers, query the Ray
global state for total available resources, and re-launch up to
the available resources. Behavior is not well-defined
in case of shared cluster usage. Defaults to 3.
info (dict): Optional dictionary passed to the training
operator for ``train_epoch`` and ``train_batch``.
dataset (Dataset): Optional dataset to train with. If specified,
the dataloader passed in via data_creator will be ignored.
Returns:
(dict | list) A dictionary of metrics for training.
You can provide custom metrics by implementing a custom
training loop. If ``reduce_results=False``, this will return a
list of metric dictionaries whose length will be equal to
``num_workers``.
"""
assert max_retries >= 0, "`max_retries` must be non-negative."
assert isinstance(dataset, Dataset) is not None \
or self.data_creator, \
"Must specify either a data creator or a dataset"
if self.worker_group.should_scale_up():
logger.info("Resize opportunity detected. Attempting to scale up.")
self._resize_worker_group()
success, worker_stats = self.worker_group.train(num_steps=num_steps,
profile=profile,
info=info,
dataset=dataset)
# Fault handling
for i in range(max_retries):
if success:
break
else:
self._num_failures += 1
self._resize_worker_group()
logger.info("Retrying training step with %d workers." %
self.worker_group.num_workers)
success, worker_stats = self.worker_group.train(
num_steps=num_steps,
profile=profile,
info=info,
dataset=dataset)
if not success:
raise RuntimeError("Training run failed.")
if reduce_results:
return self._process_stats(worker_stats)
else:
return worker_stats
def _process_stats(self, worker_stats):
stats = {
NUM_SAMPLES:
sum(stats.pop(NUM_SAMPLES, np.nan) for stats in worker_stats)
}
for stat_key in worker_stats[0]:
if isinstance(worker_stats[0][stat_key], numbers.Number):
stats[stat_key] = np.nanmean(
[s.get(stat_key, np.nan) for s in worker_stats])
else:
stats[stat_key] = worker_stats[0][stat_key]
return stats
def apply_all_workers(self, fn):
"""Run a function on all operators on the workers.
Args:
fn (Callable): A function that takes in no arguments.
Returns:
A list of objects returned by ``fn`` on each worker.
"""
return self.worker_group.apply_all_workers(fn)
def apply_all_operators(self, fn):
"""Run a function on all operators on the workers.
Args:
fn (Callable[TrainingOperator]): A function that takes in a
TrainingOperator.
Returns:
A list of objects returned by ``fn`` on each operator.
"""
return self.worker_group.apply_all_operators(fn)
def validate(self,
num_steps=None,
profile=False,
reduce_results=True,
info=None):
"""Evaluates the model on the validation data set.
Args:
num_steps (int): Number of batches to compute update steps on
per worker. This corresponds also to the number of times
``TrainingOperator.validate_batch`` is called per worker.
profile (bool): Returns time stats for the evaluation procedure.
reduce_results (bool): Whether to average all metrics across
all workers into one dict. If a metric is a non-numerical
value (or nested dictionaries), one value will be randomly
selected among the workers. If False, returns a list of dicts.
info (dict): Optional dictionary passed to the training
operator for `validate` and `validate_batch`.
Returns:
A dictionary of metrics for validation.
You can provide custom metrics by passing in a custom
``training_operator_cls``.
"""
worker_stats = self.worker_group.validate(num_steps=num_steps,
profile=profile,
info=info)
if reduce_results:
return self._process_stats(worker_stats)
else:
return worker_stats
def update_scheduler(self, metric):
"""Calls ``scheduler.step(metric)`` on all registered schedulers.
This is useful for lr_schedulers such as ``ReduceLROnPlateau``.
"""
self.worker_group.apply_all_operators(
lambda op: [sched.step(metric) for sched in op._schedulers])
def get_model(self):
"""Returns the learned model(s)."""
unwrapped = []
models = self.worker_group.get_model()
for model in models:
unwrapped += [model.module if hasattr(model, "module") else model]
if len(unwrapped) == 1:
return unwrapped[0]
return unwrapped
def get_local_operator(self):
"""Returns the local TrainingOperator object.
Be careful not to perturb its state, or else you can cause the system
to enter an inconsistent state.
Returns:
TrainingOperator: The local TrainingOperator object.
"""
return self.worker_group.get_local_operator()
def state_dict(self):
return self.worker_group.state_dict()
def load_state_dict(self, state_dict, blocking=False):
self.worker_group.load_state_dict(state_dict, blocking=blocking)
def save(self, checkpoint):
"""Saves the Trainer state to the provided checkpoint path.
Args:
checkpoint (str): Path to target checkpoint file.
"""
torch.save(self.state_dict(), checkpoint)
return checkpoint
def load(self, checkpoint):
"""Loads the Trainer and all workers from the provided checkpoint.
Args:
checkpoint (str): Path to target checkpoint file.
"""
state_dict = torch.load(checkpoint)
self.load_state_dict(state_dict)
def restore(self, *args):
raise DeprecationWarning("Use `TorchTrainer.load()` instead.")
def shutdown(self, force=False):
"""Shuts down workers and releases resources.
Args:
force (bool): If True, forcefully kill all workers. If False,
attempt a graceful shutdown first, and then forcefully kill if
unsuccessful.
"""
self.worker_group.shutdown(force=force)
self.worker_group = DeactivatedWorkerGroup()
@classmethod
def as_trainable(cls, *args, override_tune_step=None, **kwargs):
"""Creates a BaseTorchTrainable class compatible with Tune.
Any configuration parameters will be overridden by the Tune
Trial configuration. You can also pass in a custom
``override_tune_step`` to implement your own iterative optimization
routine and override the default implementation.
.. code-block:: python
def step(trainer, info):
# Implement custom objective function here.
train_stats = trainer.train()
...
# Return the metrics to report to tune.
# Do not call tune.report here.
return train_stats
TorchTrainable = TorchTrainer.as_trainable(
training_operator_cls=MyTrainingOperator,
num_gpus=2,
override_tune_step=step
)
analysis = tune.run(
TorchTrainable,
config={"lr": tune.grid_search([0.01, 0.1])}
)
Args:
override_tune_step (Callable[[TorchTrainer, Dict], Dict]): A
function to override the default training step to be used
for Ray Tune. It accepts two arguments: the first one is an
instance of your TorchTrainer, and the second one is a info
dictionary, containing information about the Trainer
state. If None is passed in, the default step
function will be
used: run 1 epoch of training, 1 epoch of validation,
and report both results to Tune. Passing in
``override_tune_step`` is useful to define
custom step functions, for example if you need to
manually update the scheduler or want to run more than 1
training epoch for each tune iteration.
"""
if override_tune_step is not None:
callback_args = inspect.signature(override_tune_step)
if not len(callback_args.parameters) == 2:
raise ValueError("override_tune_step must take in exactly 2 "
"arguments. The passed in function "
"currently takes in {} "
"args".format(
str(len(callback_args.parameters))))
class TorchTrainable(BaseTorchTrainable):
@classmethod
def default_resource_request(cls, config):
num_workers = config.get("num_workers",
kwargs.get("num_workers", 1))
num_cpus_per_worker = config.get(
"num_cpus_per_worker", kwargs.get("num_cpus_per_worker",
1))
use_gpu = config.get("use_gpu", kwargs.get("use_gpu"))
use_local = config.get("use_local",
kwargs.get("use_local", False))
if use_local:
remote_worker_count = num_workers - 1
local_cpus = 1
local_gpus = int(use_gpu)
else:
remote_worker_count = num_workers
local_cpus = 0
local_gpus = 0
return Resources(
cpu=int(local_cpus * num_cpus_per_worker),
gpu=int(local_gpus),
extra_cpu=int(remote_worker_count * num_cpus_per_worker),
extra_gpu=int(int(use_gpu) * remote_worker_count))
def step(self):
if override_tune_step is not None:
output = override_tune_step(
self._trainer, {"iteration": self.training_iteration})
return output
else:
return super(TorchTrainable, self).step()
def _create_trainer(self, tune_config):
"""Overrides the provided config with Tune config."""
provided_config = kwargs.get("config", {}).copy()
provided_config.update(tune_config)
kwargs["config"] = provided_config
trainer = TorchTrainer(*args, **kwargs)
return trainer
return TorchTrainable
