Yes, here is the code. It is almost identical to that in the CIFAR example, except that I’m giving the dataloaders and model to the function myself, I’m using a different loss function and I added a progress bar for training.
def training(local_rank, config):
rank = idist.get_rank()
manual_seed(config["seed"] + rank)
device = idist.device()
logger = setup_logger(name="NN-Training")
log_basic_info(logger, config)
output_path = config["output_path"]
if rank == 0:
if config["stop_iteration"] is None:
now = datetime.now().strftime("%Y%m%d-%H%M%S")
else:
now = f"stop-on-{config['stop_iteration']}"
folder_name = f"{config['model']}_backend-{idist.backend()}-{idist.get_world_size()}_{now}"
output_path = Path(output_path) / folder_name
if not output_path.exists():
output_path.mkdir(parents=True)
config["output_path"] = output_path.as_posix()
logger.info(f"Output path: {config['output_path']}")
if "cuda" in device.type:
config["cuda device name"] = torch.cuda.get_device_name(local_rank)
# Setup dataflow, model, optimizer, criterion
train_loader, test_loader = config['dataloaders']
config["num_iters_per_epoch"] = len(train_loader)
model, optimizer, criterion, lr_scheduler = initialize(config)
# Create trainer for current task
trainer = create_trainer(model, optimizer, criterion, lr_scheduler, train_loader.sampler, config, logger)
if idist.get_rank() == 0 and (not config["with_clearml"]):
common.ProgressBar(desc=f"Trainer (train)", persist=False).attach(trainer, output_transform=lambda x: {'batch loss': x['batch loss']})
# Let's now setup evaluator engine to perform model's validation and compute metrics
metrics = {
"MeanAbsoluteError": MeanAbsoluteError(),
"Loss": Loss(criterion),
}
# We define two evaluators as they wont have exactly similar roles:
# - `evaluator` will save the best model based on validation score
evaluator = create_evaluator(model, metrics=metrics, config=config)
train_evaluator = create_evaluator(model, metrics=metrics, config=config)
def run_validation(engine):
epoch = trainer.state.epoch
state = train_evaluator.run(train_loader)
log_metrics(logger, epoch, state.times["COMPLETED"], "Train", state.metrics)
state = evaluator.run(test_loader)
log_metrics(logger, epoch, state.times["COMPLETED"], "Test", state.metrics)
trainer.add_event_handler(Events.EPOCH_COMPLETED(every=config["validate_every"]) | Events.COMPLETED, run_validation)
if rank == 0:
# Setup TensorBoard logging on trainer and evaluators. Logged values are:
# - Training metrics, e.g. running average loss values
# - Learning rate
# - Evaluation train/test metrics
evaluators = {"training": train_evaluator, "test": evaluator}
tb_logger = common.setup_tb_logging(output_path, trainer, optimizer, evaluators=evaluators)
# Store 2 best models by validation mae starting from num_epochs / 2:
best_model_handler = Checkpoint(
{"model": model},
get_save_handler(config),
filename_prefix="best",
n_saved=2,
global_step_transform=global_step_from_engine(trainer),
score_name="test_mae",
score_function=Checkpoint.get_default_score_fn("MeanAbsoluteError"),
)
evaluator.add_event_handler(
Events.COMPLETED(lambda *_: trainer.state.epoch > config["num_epochs"] // 2), best_model_handler
)
# In order to check training resuming we can stop training on a given iteration
if config["stop_iteration"] is not None:
@trainer.on(Events.ITERATION_STARTED(once=config["stop_iteration"]))
def _():
logger.info(f"Stop training on {trainer.state.iteration} iteration")
trainer.terminate()
try:
trainer.run(train_loader, max_epochs=config["num_epochs"])
except Exception as e:
logger.exception("")
raise e
if rank == 0:
tb_logger.close()
def run(
model,
dataloaders,
seed=543,
data_path="/tmp/nn",
output_path="/tmp/output-nn/",
batch_size=512,
momentum=0.9,
weight_decay=1e-4,
num_workers=12,
num_epochs=24,
learning_rate=0.4,
num_warmup_epochs=4,
validate_every=1,
checkpoint_every=1000,
backend=None,
resume_from=None,
log_every_iters=15,
nproc_per_node=None,
stop_iteration=None,
with_clearml=False,
with_amp=False,
**spawn_kwargs,
):
"""Main entry to train an model on NN dataset.
Args:
seed (int): random state seed to set. Default, 543.
data_path (str): input dataset path. Default, "/tmp/nn".
output_path (str): output path. Default, "/tmp/output-nn".
model (str): model name (from torchvision) to setup model to train. Default, "resnet18".
batch_size (int): total batch size. Default, 512.
momentum (float): optimizer's momentum. Default, 0.9.
weight_decay (float): weight decay. Default, 1e-4.
num_workers (int): number of workers in the data loader. Default, 12.
num_epochs (int): number of epochs to train the model. Default, 24.
learning_rate (float): peak of piecewise linear learning rate scheduler. Default, 0.4.
num_warmup_epochs (int): number of warm-up epochs before learning rate decay. Default, 4.
validate_every (int): run model's validation every ``validate_every`` epochs. Default, 3.
checkpoint_every (int): store training checkpoint every ``checkpoint_every`` iterations. Default, 1000.
backend (str, optional): backend to use for distributed configuration. Possible values: None, "nccl", "xla-tpu",
"gloo" etc. Default, None.
nproc_per_node (int, optional): optional argument to setup number of processes per node. It is useful,
when main python process is spawning training as child processes.
resume_from (str, optional): path to checkpoint to use to resume the training from. Default, None.
log_every_iters (int): argument to log batch loss every ``log_every_iters`` iterations.
It can be 0 to disable it. Default, 15.
stop_iteration (int, optional): iteration to stop the training. Can be used to check resume from checkpoint.
with_clearml (bool): if True, experiment ClearML logger is setup. Default, False.
with_amp (bool): if True, enables native automatic mixed precision. Default, False.
**spawn_kwargs: Other kwargs to spawn run in child processes: master_addr, master_port, node_rank, nnodes
"""
# catch all local parameters
config = locals()
config['dataloaders'] = dataloaders
config.update(config["spawn_kwargs"])
del config["spawn_kwargs"]
spawn_kwargs["nproc_per_node"] = nproc_per_node
if backend == "xla-tpu" and with_amp:
raise RuntimeError("The value of with_amp should be False if backend is xla")
with idist.Parallel(backend=backend, **spawn_kwargs) as parallel:
parallel.run(training, config)
def get_dataflow(config):
# - Get train/test datasets
if idist.get_local_rank() > 0:
# Ensure that only local rank 0 download the dataset
# Thus each node will download a copy of the dataset
idist.barrier()
train_dataset, test_dataset = utils.get_train_test_datasets(config["data_path"])
if idist.get_local_rank() == 0:
# Ensure that only local rank 0 download the dataset
idist.barrier()
# Setup data loader also adapted to distributed config: nccl, gloo, xla-tpu
train_loader = idist.auto_dataloader(
train_dataset, batch_size=config["batch_size"], num_workers=config["num_workers"], shuffle=True, drop_last=True,
)
test_loader = idist.auto_dataloader(
test_dataset, batch_size=2 * config["batch_size"], num_workers=config["num_workers"], shuffle=False,
)
return train_loader, test_loader
def initialize(config):
model = config["model"]
# Adapt model for distributed settings if configured
model = idist.auto_model(model)
optimizer = optim.SGD(
model.parameters(),
lr=config["learning_rate"],
momentum=config["momentum"],
weight_decay=config["weight_decay"],
nesterov=True,
)
optimizer = idist.auto_optim(optimizer)
criterion = nn.L1Loss().to(idist.device())
le = config["num_iters_per_epoch"]
milestones_values = [
(0, 0.0),
(le * config["num_warmup_epochs"], config["learning_rate"]),
(le * config["num_epochs"], 0.0),
]
lr_scheduler = PiecewiseLinear(optimizer, param_name="lr", milestones_values=milestones_values)
return model, optimizer, criterion, lr_scheduler
def log_metrics(logger, epoch, elapsed, tag, metrics):
metrics_output = "\n".join([f"\t{k}: {v}" for k, v in metrics.items()])
logger.info(f"\nEpoch {epoch} - Evaluation time (seconds): {elapsed:.2f} - {tag} metrics:\n {metrics_output}")
def log_basic_info(logger, config):
logger.info(f"Train {config['model']}")
logger.info(f"- PyTorch version: {torch.__version__}")
logger.info(f"- Ignite version: {ignite.__version__}")
if torch.cuda.is_available():
# explicitly import cudnn as
# torch.backends.cudnn can not be pickled with hvd spawning procs
from torch.backends import cudnn
logger.info(f"- GPU Device: {torch.cuda.get_device_name(idist.get_local_rank())}")
logger.info(f"- CUDA version: {torch.version.cuda}")
logger.info(f"- CUDNN version: {cudnn.version()}")
logger.info("\n")
logger.info("Configuration:")
for key, value in config.items():
logger.info(f"\t{key}: {value}")
logger.info("\n")
if idist.get_world_size() > 1:
logger.info("\nDistributed setting:")
logger.info(f"\tbackend: {idist.backend()}")
logger.info(f"\tworld size: {idist.get_world_size()}")
logger.info("\n")
def create_trainer(model, optimizer, criterion, lr_scheduler, train_sampler, config, logger):
device = idist.device()
# Setup Ignite trainer:
# - let's define training step
# - add other common handlers:
# - TerminateOnNan,
# - handler to setup learning rate scheduling,
# - ModelCheckpoint
# - RunningAverage` on `train_step` output
# - Two progress bars on epochs and optionally on iterations
with_amp = config["with_amp"]
scaler = GradScaler(enabled=with_amp)
def train_step(engine, batch):
x, y = batch[0], batch[1]
if x.device != device:
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
model.train()
with autocast(enabled=with_amp):
y_pred = model(x)
loss = criterion(y_pred, y)
optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
return {
"batch loss": loss.item(),
}
trainer = Engine(train_step)
trainer.logger = logger
to_save = {"trainer": trainer, "model": model, "optimizer": optimizer, "lr_scheduler": lr_scheduler}
metric_names = [
"batch loss",
]
common.setup_common_training_handlers(
trainer=trainer,
train_sampler=train_sampler,
to_save=to_save,
save_every_iters=config["checkpoint_every"],
save_handler=get_save_handler(config),
lr_scheduler=lr_scheduler,
output_names=metric_names if config["log_every_iters"] > 0 else None,
with_pbars=False,
clear_cuda_cache=False,
)
resume_from = config["resume_from"]
if resume_from is not None:
checkpoint_fp = Path(resume_from)
assert checkpoint_fp.exists(), f"Checkpoint '{checkpoint_fp.as_posix()}' is not found"
logger.info(f"Resume from a checkpoint: {checkpoint_fp.as_posix()}")
checkpoint = torch.load(checkpoint_fp.as_posix(), map_location="cpu")
Checkpoint.load_objects(to_load=to_save, checkpoint=checkpoint)
return trainer
def create_evaluator(model, metrics, config, tag="val"):
with_amp = config["with_amp"]
device = idist.device()
@torch.no_grad()
def evaluate_step(engine: Engine, batch):
model.eval()
x, y = batch[0], batch[1]
if x.device != device:
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
with autocast(enabled=with_amp):
output = model(x)
return output, y
evaluator = Engine(evaluate_step)
for name, metric in metrics.items():
metric.attach(evaluator, name)
if idist.get_rank() == 0 and (not config["with_clearml"]):
common.ProgressBar(desc=f"Evaluation ({tag})", persist=False).attach(evaluator)
return evaluator
def get_save_handler(config):
if config["with_clearml"]:
from ignite.contrib.handlers.clearml_logger import ClearMLSaver
return ClearMLSaver(dirname=config["output_path"])
return DiskSaver(config["output_path"], require_empty=False)
One notable difference with your command is that I am running this in a Jupyter Notebook in the following manner (no torch.distributed.launch in my code):
res = run(
model,
[train_dl, test_dl],
backend='gloo',
nproc_per_node=4,
)
The result is that the program hangs after printing the following log:
ignite.distributed.launcher.Parallel INFO: Spawn function '<function training at ...>' in 4 processes
Should I run this as a script with torch.distributed.launch?
Thanks for the prompt answer!
. Any other clues on what might be going on?
