class LocalWorkerGroup(WorkerGroupInterface):
"""A group of TorchRunner workers.
1 worker runs locally, and all the other workers are remote actors.
Args:
Same as RemoteWorkerGroup.
"""
def __init__(
self,
max_workers,
params,
dist_params,
initialization_hook,
timeout_s,
num_cpus_per_worker,
use_gpu,
):
# Invariant: These variables should never change state!
self._max_workers = max_workers
self._params = params
self._dist_params = dist_params
self._initialization_hook = initialization_hook
self._timeout_s = timeout_s
self._num_cpus_per_worker = num_cpus_per_worker
self._use_gpu = use_gpu
self.local_worker = None
self.remote_worker_group = RemoteWorkerGroup(
max_workers=max_workers - 1,
params=params,
dist_params=dist_params,
initialization_hook=initialization_hook,
timeout_s=timeout_s,
num_cpus_per_worker=num_cpus_per_worker,
use_gpu=use_gpu,
)
def start_workers(self, num_workers):
logger.debug(f"start_workers: Setting {num_workers} workers.")
if num_workers == 1:
self.local_worker = TorchRunner(**self._params)
if self._initialization_hook:
self.apply_all_workers(self._initialization_hook)
self.local_worker.setup_operator()
return True
else:
try:
# Start local worker
self.local_worker = LocalDistributedRunner(
num_cpus=self._num_cpus_per_worker,
num_gpus=int(self._use_gpu),
**{
**self._params,
**self._dist_params
},
)
self.remote_worker_group._init_dist_workers(num_workers - 1)
if self._initialization_hook:
self.apply_all_workers(self._initialization_hook)
# Compute URL for initializing distributed PyTorch.
address = setup_address()
remote_pgs = self.remote_worker_group._setup_process_group(
address=address, world_size=num_workers, starting_rank=1)
# Use the local worker as rank 0. Helps with debugging.
self.local_worker.setup_process_group(
url=address,
world_rank=0,
world_size=num_workers,
timeout=timedelta(seconds=self._timeout_s),
)
ray.get(remote_pgs)
local_node_ip = ray.util.get_node_ip_address()
rank_dict = defaultdict(int)
self.local_worker.set_local_rank(local_rank=0)
rank_dict[local_node_ip] += 1
self.remote_worker_group._setup_local_rank(rank_dict)
remote_operators = self.remote_worker_group._setup_operator()
self.local_worker.setup_operator()
ray.get(remote_operators)
return True
except RayActorError:
return False
def apply_all_operators(self, fn):
remote_calls = self.remote_worker_group._apply_all_operators(fn)
local_call = self.local_worker.apply_operator(fn)
return [local_call] + ray.get(remote_calls)
def apply_all_workers(self, fn):
remote_calls = self.remote_worker_group._apply_all_workers(fn)
local_call = self.local_worker.apply(fn)
return [local_call] + ray.get(remote_calls)
def get_local_operator(self):
return self.local_worker.training_operator
def get_model(self, to_cpu=False):
models = self.local_worker.models
if to_cpu:
models = [m.cpu() for m in models]
return models
def load_state_dict(self, state_dict, blocking=False):
# This is not the most efficient because you have to wait for
# the local worker to save then dump to buffer.
self.local_worker.load_state_dict(state_dict)
state_id = ray.put(self.local_worker.state_stream())
remote_calls = self.remote_worker_group._load_state_id(state_id)
if blocking:
ray.get(remote_calls)
def state_dict(self):
return self.local_worker.state_dict()
def should_scale_up(self):
return self.remote_worker_group.should_scale_up()
def reset(self):
"""Terminates models without giving up local resource reservation."""
if not isinstance(self.local_worker, LocalDistributedRunner):
self.local_worker.shutdown()
else:
self.local_worker.shutdown(cleanup=False)
self.remote_worker_group.reset()
self.local_worker = None
self.remote_worker_group = RemoteWorkerGroup(
max_workers=self._max_workers - 1,
params=self._params,
dist_params=self._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,
)
def new_workers_size(self):
return self.remote_worker_group.new_workers_size() + 1
def train(self, num_steps=None, profile=False, info=None, dataset=None):
params = dict(num_steps=num_steps, profile=profile, info=info)
if dataset:
dataset.set_num_shards(self.num_workers)
remote_worker_stats = self.remote_worker_group._train(
num_steps, profile, info, dataset)
try:
if dataset:
params["iterator"] = dataset.get_shard(self.num_workers - 1)
local_worker_stats = self.local_worker.train_epoch(**params)
except RuntimeError as err:
if "gloo" in err.args[0] and "Timed out" in err.args[0]:
logger.warning(err)
return False, None
if "NCCL" in err.args[0]: # there is no specific error message
logger.warning(err)
return False, None
if "Connection closed by peer" in err.args[0]:
logger.warning(err)
return False, None
raise err
success = check_for_failure(remote_worker_stats)
if success:
return success, [local_worker_stats] + ray.get(remote_worker_stats)
return success, None
def validate(self, num_steps=None, profile=False, info=None):
params = dict(num_steps=num_steps, profile=profile, info=info)
remote_worker_stats = self.remote_worker_group._validate(params)
local_worker_stats = self.local_worker.validate(**params)
worker_stats = [local_worker_stats] + ray.get(remote_worker_stats)
return worker_stats
def shutdown(self, force=False):
self.local_worker.shutdown()
self.remote_worker_group.shutdown(force=force)
self.local_worker = None
self.remote_worker_group = DeactivatedWorkerGroup()
@property
def num_workers(self):
return self.remote_worker_group.num_workers + 1
@property
def remote_workers(self):
return self.remote_worker_group.remote_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.
.. code-block:: python
ray.init()
def model_creator(config):
return nn.Linear(1, 1)
def optimizer_creator(model, config):
return torch.optim.SGD(
model.parameters(), lr=config.get("lr", 1e-4))
def data_creator(config):
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)
return train_loader, val_loader
trainer = TorchTrainer(
model_creator=model_creator,
data_creator=data_creator,
optimizer_creator=optimizer_creator,
loss_creator=nn.MSELoss,
config={"batch_size": 32},
use_gpu=True
)
for i in range(4):
trainer.train()
Args:
model_creator (dict -> Model(s)): Constructor function that takes in
config and returns the model(s) to be optimized. These must be
``torch.nn.Module`` objects. If multiple models are returned,
a ``training_operator_cls`` must be specified. You do not need to
handle GPU/devices in this function; RaySGD will do that under
the hood.
data_creator (dict -> Iterable(s)): Constructor function
that takes in the passed config and returns one or
two Iterable objects. Note that even though two Iterable objects
can be returned, only one will be used for training, and the
other will be used for validation. If not provided, you must
provide a custom TrainingOperator.
optimizer_creator ((models, dict) -> optimizers): Constructor
function that takes in the return values from
``model_creator`` and the passed config and returns One or
more Torch optimizer objects. You do not need to handle
GPU/devices in this function; ``RaySGD`` will do that for you.
loss_creator (torch.nn.*Loss class | dict -> loss): A constructor
function for the training loss. This can be either a function that
takes in the provided config for customization or a subclass
of ``torch.nn.modules.loss._Loss``, which is most Pytorch
loss classes. For example, ``loss_creator=torch.nn.BCELoss``.
If not provided, you must provide a custom TrainingOperator.
scheduler_creator ((optimizers, dict) -> scheduler):
A constructor function for the torch scheduler. This is
a function that takes in the generated optimizers (from
``optimizer_creator``) provided config for customization.
Be sure to set ``scheduler_step_freq`` to increment the
scheduler correctly.
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
custom training and validation operations. Defaults to
TrainingOperator.
config (dict): Custom configuration value to be passed to
all creator and operator constructors.
num_workers (int): the number of workers used in distributed
training. If 1, the worker will not be wrapped with
DistributedDataParallel.
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.
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", 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.
"""
def __init__(
self,
*,
model_creator,
data_creator,
optimizer_creator,
loss_creator=None,
scheduler_creator=None,
training_operator_cls=None,
initialization_hook=None,
config=None,
num_workers=1,
use_gpu=False,
backend="auto",
use_fp16=False,
use_tqdm=False,
apex_args=None,
scheduler_step_freq="batch",
num_replicas=None,
batch_size=None,
data_loader_args=None,
):
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 not (callable(model_creator) and callable(optimizer_creator)
and callable(data_creator)):
raise ValueError(
"Must provide a callable model_creator, optimizer_creator, "
"and data_creator.")
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 data_loader_args:
raise ValueError(
"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.model_creator = model_creator
self.optimizer_creator = optimizer_creator
self.loss_creator = loss_creator
self.data_creator = data_creator
self.scheduler_creator = scheduler_creator
self.training_operator_cls = training_operator_cls
if not training_operator_cls and not loss_creator:
raise ValueError("If a loss_creator is not provided, you must "
"provide a custom training operator.")
self.initialization_hook = initialization_hook
self.config = {} if config is None else config
if backend == "auto":
backend = "nccl" if use_gpu else "gloo"
logger.debug("Using {} as backend.".format(backend))
self.backend = backend
# TODO: Have an auto "use_gpu" option to detect and use GPUs.
self.use_gpu = use_gpu
self.max_replicas = num_workers
self.use_fp16 = use_fp16
self.use_tqdm = use_tqdm
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")
_validate_scheduler_step_freq(scheduler_step_freq)
self.scheduler_step_freq = scheduler_step_freq
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):
logger.debug(f"start_workers: Setting %d workers." % 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
self.local_worker = None
self.remote_workers = []
if num_workers == 1:
# Start local worker
self.local_worker = TorchRunner(
model_creator=self.model_creator,
data_creator=self.data_creator,
optimizer_creator=self.optimizer_creator,
loss_creator=self.loss_creator,
scheduler_creator=self.scheduler_creator,
training_operator_cls=self.training_operator_cls,
config=worker_config,
use_fp16=self.use_fp16,
use_tqdm=self.use_tqdm,
apex_args=self.apex_args,
scheduler_step_freq=self.scheduler_step_freq)
if self.initialization_hook:
self.apply_all_workers(self.initialization_hook)
self.local_worker.setup()
else:
params = dict(
model_creator=self.model_creator,
data_creator=self.data_creator,
optimizer_creator=self.optimizer_creator,
loss_creator=self.loss_creator,
scheduler_creator=self.scheduler_creator,
backend=self.backend,
training_operator_cls=self.training_operator_cls,
config=worker_config,
use_fp16=self.use_fp16,
use_tqdm=self.use_tqdm,
apex_args=self.apex_args,
scheduler_step_freq=self.scheduler_step_freq)
# Start local worker
self.local_worker = LocalDistributedRunner(
num_cpus=1, num_gpus=int(self.use_gpu), **params)
# Generate actor class
RemoteRunner = ray.remote(
num_cpus=1, num_gpus=int(self.use_gpu))(DistributedTorchRunner)
# Start workers
self.remote_workers = [
RemoteRunner.remote(**params) for i in range(num_workers - 1)
]
if self.initialization_hook:
self.apply_all_workers(self.initialization_hook)
# Compute URL for initializing distributed PyTorch
ip = ray.services.get_node_ip_address()
port = self.local_worker.find_free_port()
address = "tcp://{ip}:{port}".format(ip=ip, port=port)
remote_setups = [
worker.setup.remote(address, i + 1, num_workers)
for i, worker in enumerate(self.remote_workers)
]
self.local_worker.setup(address, 0, num_workers)
# Get setup tasks in order to throw errors on failure
ray.get(remote_setups)
def train(self,
num_steps=None,
profile=False,
reduce_results=True,
max_retries=0,
checkpoint="auto",
info=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.
This corresponds also to the number of times
``TrainingOperator.train_batch`` is called.
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, 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.
checkpoint (str): Path to checkpoint to restore from if retrying.
If max_retries is set and ``checkpoint == "auto"``,
TorchTrainer will save a checkpoint before starting to train.
info (dict): Optional dictionary passed to the training
operator for ``train_epoch`` and ``train_batch``.
Returns:
(dict | list) A dictionary of metrics for training.
You can provide custom metrics by passing in a custom
``training_operator_cls``. 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."
if max_retries:
if checkpoint == "auto":
logger.debug("Retrying detected. Automatically checkpointing.")
checkpoint = self.save(
os.path.join(self.temp_dir, "tmp_checkpoint"))
elif not checkpoint:
raise ValueError("Cannot retry from empty checkpoint.")
if checkpoint and self._should_resize():
logger.info("Resize opportunity detected. Attempting to scale up.")
self._resize_workers(checkpoint=checkpoint)
success, worker_stats = self._train_epoch(
num_steps=num_steps, profile=profile, info=info)
# Fault handling
for i in range(max_retries):
if success:
break
else:
self._num_failures += 1
self._resize_workers(checkpoint=checkpoint)
logger.info("Retrying training step with %d workers." %
(len(self.remote_workers) + 1))
success, worker_stats = self._train_epoch(
num_steps=num_steps, profile=profile, info=info)
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], 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 _train_epoch(self, num_steps=None, profile=False, info=None):
params = dict(num_steps=num_steps, profile=profile, info=info)
remote_worker_stats = [
w.train_epoch.remote(**params) for w in self.remote_workers
]
try:
local_worker_stats = self.local_worker.train_epoch(**params)
except RuntimeError as err:
if "gloo" in err.args[0] and "Timed out" in err.args[0]:
logger.warning(err)
return False, None
if "NCCL" in err.args[0]: # there is no specific error message
logger.warning(err)
return False, None
raise err
success = check_for_failure(remote_worker_stats)
if success:
return success, [local_worker_stats] + ray.get(remote_worker_stats)
return success, None
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.
"""
remote_calls = [w.apply.remote(fn) for w in self.remote_workers]
local_call = self.local_worker.apply(fn)
return [local_call] + ray.get(remote_calls)
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.
"""
remote_calls = [
w.apply_operator.remote(fn) for w in self.remote_workers
]
local_call = self.local_worker.apply_operator(fn)
return [local_call] + ray.get(remote_calls)
def validate(self, num_steps=None, profile=False, info=None):
"""Evaluates the model on the validation data set.
Args:
num_steps (int): Number of batches to compute update steps on.
This corresponds also to the number of times
``TrainingOperator.validate_batch`` is called.
profile (bool): Returns time stats for the evaluation procedure.
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``.
"""
params = dict(num_steps=num_steps, profile=profile, info=info)
remote_worker_stats = [
w.validate.remote(**params) for w in self.remote_workers
]
local_worker_stats = self.local_worker.validate(**params)
return self._process_stats([local_worker_stats] +
ray.get(remote_worker_stats))
def update_scheduler(self, metric):
"""Calls ``scheduler.step(metric)`` on all schedulers.
This is useful for lr_schedulers such as ``ReduceLROnPlateau``.
"""
self.apply_all_operators(
lambda op: [sched.step(metric) for sched in op.schedulers])
def get_model(self):
"""Returns the learned model(s)."""
models = self.model_creator(self.config)
state = self.local_worker.get_state()
if len(state["models"]) == 1:
models.load_state_dict(state["models"][0])
else:
for model, state_dict in zip(models, state["models"]):
model.load_state_dict(state_dict)
return models
def state_dict(self):
return self.local_worker.get_state()
def load_state_dict(self, state):
state_id = ray.put(state)
remote_calls = [
worker.set_state.remote(state_id) for worker in self.remote_workers
]
self.local_worker.set_state(state)
ray.get(remote_calls)
def save(self, checkpoint):
"""Saves the model(s) to the provided checkpoint.
Args:
checkpoint (str): Path to target checkpoint file.
Returns:
checkpoint (str): Path to target checkpoint file.
"""
torch.save(self.state_dict(), checkpoint)
return checkpoint
def restore(self, checkpoint):
"""Restores the Trainer and all workers from the provided checkpoint.
Args:
checkpoint (str): Path to target checkpoint file.
"""
state = torch.load(checkpoint)
self.load_state_dict(state)
def shutdown(self, force=False):
"""Shuts down workers and releases resources."""
if not force:
cleanup = [
worker.shutdown.remote() for worker in self.remote_workers
]
self.local_worker.shutdown()
try:
ray.get(cleanup)
[
worker.__ray_terminate__.remote()
for worker in self.remote_workers
]
except RayActorError:
logger.warning(
"Failed to shutdown gracefully, forcing a shutdown.")
for worker in self.remote_workers:
logger.warning("Killing worker {}.".format(worker))
ray.kill(worker)
else:
self.local_worker.shutdown()
for worker in self.remote_workers:
logger.warning("Killing worker {}.".format(worker))
ray.kill(worker)
self.local_worker = None
self.remote_workers = []
def _reset(self):
"""Terminates models without giving up local resource reservation."""
self.local_worker.shutdown(cleanup=False)
for worker in self.remote_workers:
logger.warning("Killing worker {}.".format(worker))
ray.kill(worker)
self.local_worker = None
self.remote_workers = []
def _check_potential_remote_workers_size(self):
# ASSUME 1 GPU + 1 CPU is already reserved for the local worker
remote_resources = ray.available_resources()
max_remote_workers = self.max_replicas - 1
new_remote_workers = min(
remote_resources.get("CPU", 0), max_remote_workers)
if self.use_gpu:
new_remote_workers = min(
remote_resources.get("GPU", 0), new_remote_workers)
return new_remote_workers
def _resize_workers(self, checkpoint, max_retries=10):
self._reset()
assert checkpoint, "Cannot restore without checkpoint."
time.sleep(1)
for i in range(max_retries):
new_remote_workers = self._check_potential_remote_workers_size()
if new_remote_workers:
self._last_resize = time.time()
self._start_workers(int(new_remote_workers) + 1)
self.restore(checkpoint)
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 _should_resize(self):
"""Returns True if past cooldown and exists resources to scale up."""
worker_gap = self.max_replicas - 1 - len(self.remote_workers)
past_cooldown = (time.time() - self._last_resize) > RESIZE_COOLDOWN_S
if past_cooldown and worker_gap:
# Assume 1 resource is already reserved for local worker.
potential_remote_size = self._check_potential_remote_workers_size()
return potential_remote_size > 0
return False