def aggregate_logging_outputs(self, logging_outputs, criterion):
"""[deprecated] Aggregate logging outputs from data parallel training."""
utils.deprecation_warning(
"The aggregate_logging_outputs API is deprecated. "
"Please use the reduce_metrics API instead.")
with metrics.aggregate() as agg:
self.reduce_metrics(logging_outputs, criterion)
return agg.get_smoothed_values()
def _reduce_and_log_stats(self,
logging_outputs,
sample_size,
grad_norm=None):
if grad_norm is not None:
metrics.log_speed("ups", 1., priority=100, round=2)
metrics.log_scalar("gnorm", grad_norm, priority=400, round=3)
if self.args.clip_norm > 0:
metrics.log_scalar(
"clip",
torch.where(
grad_norm > self.args.clip_norm,
grad_norm.new_tensor(100),
grad_norm.new_tensor(0),
),
priority=500,
round=1,
)
with metrics.aggregate() as agg:
if logging_outputs is not None:
self.task.reduce_metrics(logging_outputs, self.get_criterion())
del logging_outputs
# extra warning for criterions that don't properly log a loss value
if "loss" not in agg:
if "loss" not in self._warn_once:
self._warn_once.add("loss")
logger.warning(
"Criterion.reduce_metrics did not log a 'loss' value, "
"which may break some functionality")
metrics.log_scalar("loss", -1)
# support legacy interface
logging_output = agg.get_smoothed_values()
logging_output["sample_size"] = sample_size
for key_to_delete in ["ppl", "wps", "wpb", "bsz"]:
if key_to_delete in logging_output:
del logging_output[key_to_delete]
return logging_output
def validate(args, trainer, task, epoch_itr, subsets):
"""Evaluate the model on the validation set(s) and return the losses."""
if args.fixed_validation_seed is not None:
# set fixed seed for every validation
utils.set_torch_seed(args.fixed_validation_seed)
valid_losses = []
for subset in subsets:
logger.info('begin validation on "{}" subset'.format(subset))
# Initialize data iterator
itr = trainer.get_valid_iterator(subset).next_epoch_itr(shuffle=False)
progress = progress_bar.progress_bar(
itr,
log_format=args.log_format,
log_interval=args.log_interval,
epoch=epoch_itr.epoch,
prefix=f"valid on '{subset}' subset",
tensorboard_logdir=(
args.tensorboard_logdir if distributed_utils.is_master(args) else None
),
default_log_format=("tqdm" if not args.no_progress_bar else "simple"),
)
# create a new root metrics aggregator so validation metrics
# don't pollute other aggregators (e.g., train meters)
with metrics.aggregate(new_root=True) as agg:
for sample in progress:
trainer.valid_step(sample)
# log validation stats
stats = get_valid_stats(args, trainer, agg.get_smoothed_values())
progress.print(stats, tag=subset, step=trainer.get_num_updates())
valid_losses.append(stats[args.best_checkpoint_metric])
return valid_losses
def main(args, override_args=None):
utils.import_user_module(args)
assert args.max_tokens is not None or args.max_sentences is not None, \
'Must specify batch size either with --max-tokens or --max-sentences'
use_fp16 = args.fp16
use_cuda = torch.cuda.is_available() and not args.cpu
if use_cuda:
torch.cuda.set_device(args.device_id)
if override_args is not None:
overrides = vars(override_args)
overrides.update(eval(getattr(override_args, 'model_overrides', '{}')))
else:
overrides = None
# Load ensemble
logger.info('loading model(s) from {}'.format(args.path))
models, model_args, task = checkpoint_utils.load_model_ensemble_and_task(
[args.path],
arg_overrides=overrides,
suffix=getattr(args, "checkpoint_suffix", ""),
)
model = models[0]
# Move models to GPU
for model in models:
if use_fp16:
model.half()
if use_cuda:
model.cuda()
# Print args
logger.info(model_args)
# Build criterion
criterion = task.build_criterion(model_args)
criterion.eval()
for subset in args.valid_subset.split(','):
try:
task.load_dataset(subset, combine=False, epoch=1)
dataset = task.dataset(subset)
except KeyError:
raise Exception('Cannot find dataset: ' + subset)
# Initialize data iterator
itr = task.get_batch_iterator(
dataset=dataset,
max_tokens=args.max_tokens,
max_sentences=args.max_sentences,
max_positions=utils.resolve_max_positions(
task.max_positions(),
*[m.max_positions() for m in models],
),
ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=args.required_batch_size_multiple,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
num_workers=args.num_workers,
).next_epoch_itr(shuffle=False)
progress = progress_bar.progress_bar(
itr,
log_format=args.log_format,
log_interval=args.log_interval,
prefix=f"valid on '{subset}' subset",
default_log_format=('tqdm' if not args.no_progress_bar else 'simple'),
)
log_outputs = []
for i, sample in enumerate(progress):
sample = utils.move_to_cuda(sample) if use_cuda else sample
_loss, _sample_size, log_output = task.valid_step(sample, model, criterion)
progress.log(log_output, step=i)
log_outputs.append(log_output)
if args.distributed_world_size > 1:
log_outputs = distributed_utils.all_gather_list(
log_outputs,
max_size=getattr(args, 'all_gather_list_size', 16384),
)
log_outputs = list(chain.from_iterable(log_outputs))
with metrics.aggregate() as agg:
task.reduce_metrics(log_outputs, criterion)
log_output = agg.get_smoothed_values()
progress.print(log_output, tag=subset, step=i)
def train(args, trainer, task, epoch_itr):
"""Train the model for one epoch and return validation losses."""
# Initialize data iterator
itr = epoch_itr.next_epoch_itr(
fix_batches_to_gpus=args.fix_batches_to_gpus,
shuffle=(epoch_itr.next_epoch_idx > args.curriculum),
)
update_freq = (
args.update_freq[epoch_itr.epoch - 1]
if epoch_itr.epoch <= len(args.update_freq)
else args.update_freq[-1]
)
itr = iterators.GroupedIterator(itr, update_freq)
progress = progress_bar.progress_bar(
itr,
log_format=args.log_format,
log_interval=args.log_interval,
epoch=epoch_itr.epoch,
tensorboard_logdir=(
args.tensorboard_logdir if distributed_utils.is_master(args) else None
),
default_log_format=("tqdm" if not args.no_progress_bar else "simple"),
)
trainer.begin_epoch(epoch_itr.epoch)
valid_subsets = args.valid_subset.split(",")
should_stop = False
num_updates = trainer.get_num_updates()
for i, samples in enumerate(progress):
with metrics.aggregate("train_inner"), torch.autograd.profiler.record_function(
"train_step-%d" % i
):
log_output = trainer.train_step(samples)
if log_output is not None: # not OOM, overflow, ...
# log mid-epoch stats
num_updates = trainer.get_num_updates()
if num_updates % args.log_interval == 0:
stats = get_training_stats(metrics.get_smoothed_values("train_inner"))
progress.log(stats, tag="train_inner", step=num_updates)
# reset mid-epoch stats after each log interval
# the end-of-epoch stats will still be preserved
metrics.reset_meters("train_inner")
end_of_epoch = not itr.has_next()
valid_losses, should_stop = validate_and_save(
args, trainer, task, epoch_itr, valid_subsets, end_of_epoch
)
if should_stop:
break
# log end-of-epoch stats
logger.info("end of epoch {} (average epoch stats below)".format(epoch_itr.epoch))
stats = get_training_stats(metrics.get_smoothed_values("train"))
progress.print(stats, tag="train", step=num_updates)
# reset epoch-level meters
metrics.reset_meters("train")
return valid_losses, should_stop