def from_params(
cls, # type: ignore
params: Params,
serialization_dir: str,
recover: bool = False):
# pylint: disable=arguments-differ
typ3 = params.get("trainer", {}).pop("type", "default")
if typ3 == "default":
# Special logic to keep old from_params behavior.
from allennlp.training.trainer import Trainer
from allennlp.training.trainer_pieces import TrainerPieces
pieces = TrainerPieces.from_params(params, serialization_dir,
recover) # pylint: disable=no-member
return Trainer.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
iterator=pieces.iterator,
train_data=pieces.train_dataset,
validation_data=pieces.validation_dataset,
params=pieces.params,
validation_iterator=pieces.validation_iterator)
else:
klass = TrainerBase.by_name(typ3)
# Explicit check to prevent recursion.
is_overriden = klass.from_params.__func__ != TrainerBase.from_params.__func__ # type: ignore
assert is_overriden, f"Class {klass.__name__} must override `from_params`."
return klass.from_params(params, serialization_dir, recover)
def from_params(
cls, # type: ignore
params: Params,
serialization_dir: str,
recover: bool = False):
# pylint: disable=arguments-differ
pieces = TrainerPieces.from_params(params, serialization_dir, recover) # pylint: disable=no-member
return NoOpTrainer(serialization_dir, pieces.model)
def from_params( # type: ignore
cls,
params: Params,
serialization_dir: str,
recover: bool = False,
cache_directory: str = None,
cache_prefix: str = None,
) -> "CallbackTrainer":
pieces = TrainerPieces.from_params(params, serialization_dir, recover,
cache_directory, cache_prefix)
model = pieces.model
params = pieces.params
validation_iterator = pieces.validation_iterator or pieces.iterator
shuffle = params.pop_bool("shuffle", True)
num_epochs = params.pop_int("num_epochs", 20)
cuda_device = parse_cuda_device(params.pop("cuda_device", -1))
if isinstance(cuda_device, list):
model_device = cuda_device[0]
else:
model_device = cuda_device
if model_device >= 0:
# Moving model to GPU here so that the optimizer state gets constructed on
# the right device.
model = model.cuda(model_device)
parameters = [[n, p] for n, p in model.named_parameters()
if p.requires_grad]
optimizer = Optimizer.from_params(parameters, params.pop("optimizer"))
callbacks_params = params.pop("callbacks", [])
callbacks: List[Callback] = [
Callback.from_params(
params=callback_params,
model=model,
optimizer=optimizer,
instances=pieces.train_dataset,
iterator=pieces.iterator,
shuffle=shuffle,
validation_data=pieces.validation_dataset,
validation_iterator=validation_iterator,
serialization_dir=serialization_dir,
) for callback_params in callbacks_params
]
params.assert_empty(cls.__name__)
return cls(
model,
pieces.train_dataset,
pieces.iterator,
optimizer,
num_epochs=num_epochs,
shuffle=shuffle,
serialization_dir=serialization_dir,
cuda_device=cuda_device,
callbacks=callbacks,
)
def from_params( # type: ignore
cls,
params: Params,
serialization_dir: str,
recover: bool = False,
cache_directory: str = None,
cache_prefix: str = None,
):
pieces = TrainerPieces.from_params(params, serialization_dir, recover)
return NoOpTrainer(serialization_dir, pieces.model)
def from_params(
cls, # type: ignore
params: Params,
serialization_dir: str,
recover: bool = False):
from allennlp.training.trainer_pieces import TrainerPieces
pieces = TrainerPieces.from_params(params, serialization_dir, recover) # pylint: disable=no-member
return SplitUncertainModelTrainer.this_from_params(
model=pieces.model,
serialization_dir=serialization_dir,
iterator=pieces.iterator,
train_data=pieces.train_dataset,
validation_data=pieces.validation_dataset,
params=pieces.params,
validation_iterator=pieces.validation_iterator)
def from_params(
cls, # type: ignore
params: Params,
serialization_dir: str,
recover: bool = False,
cache_directory: str = None,
cache_prefix: str = None):
print(params)
pieces = TrainerPieces.from_params(
params, # pylint: disable=no-member
serialization_dir,
recover,
cache_directory,
cache_prefix)
return _from_params(cls, pieces.model, serialization_dir,
pieces.iterator, pieces.train_dataset,
pieces.validation_dataset, pieces.params,
pieces.validation_iterator)
def from_params( # type: ignore
cls,
params: Params,
serialization_dir: str,
recover: bool = False,
cache_directory: str = None,
cache_prefix: str = None,
):
typ3 = params.get("trainer", {}).pop("type", "default")
if typ3 == "default":
# Special logic to keep old from_params behavior.
from allennlp.training.trainer import Trainer
from allennlp.training.trainer_pieces import TrainerPieces
pieces = TrainerPieces.from_params(params, serialization_dir,
recover, cache_directory,
cache_prefix)
return Trainer.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
iterator=pieces.iterator,
train_data=pieces.train_dataset,
validation_data=pieces.validation_dataset,
params=pieces.params,
validation_iterator=pieces.validation_iterator,
)
else:
klass = TrainerBase.by_name(typ3)
# Explicit check to prevent recursion.
is_overriden = (
klass.from_params.__func__ !=
TrainerBase.from_params.__func__ # type: ignore
)
assert is_overriden, f"Class {klass.__name__} must override `from_params`."
return klass.from_params(params, serialization_dir, recover,
cache_directory, cache_prefix)
def train_model(params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
force: bool = False,
cache_directory: str = None,
cache_prefix: str = None) -> Model:
"""
Trains the model specified in the given :class:`Params` object, using the data and training
parameters also specified in that object, and saves the results in ``serialization_dir``.
Parameters
----------
params : ``Params``
A parameter object specifying an AllenNLP Experiment.
serialization_dir : ``str``
The directory in which to save results and logs.
file_friendly_logging : ``bool``, optional (default=False)
If ``True``, we add newlines to tqdm output, even on an interactive terminal, and we slow
down tqdm's output to only once every 10 seconds.
recover : ``bool``, optional (default=False)
If ``True``, we will try to recover a training run from an existing serialization
directory. This is only intended for use when something actually crashed during the middle
of a run. For continuing training a model on new data, see the ``fine-tune`` command.
force : ``bool``, optional (default=False)
If ``True``, we will overwrite the serialization directory if it already exists.
cache_directory : ``str``, optional
For caching data pre-processing. See :func:`allennlp.training.util.datasets_from_params`.
cache_prefix : ``str``, optional
For caching data pre-processing. See :func:`allennlp.training.util.datasets_from_params`.
Returns
-------
best_model: ``Model``
The model with the best epoch weights.
"""
create_serialization_dir(params, serialization_dir, recover, force)
stdout_handler = prepare_global_logging(serialization_dir,
file_friendly_logging)
prepare_environment(params)
cuda_device = params.params.get('trainer').get('cuda_device', -1)
check_for_gpu(cuda_device)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
evaluate_on_test = params.pop_bool("evaluate_on_test", False)
trainer_type = params.get("trainer", {}).get("type", "default")
if trainer_type == "default":
# Special logic to instantiate backward-compatible trainer.
pieces = TrainerPieces.from_params(
params, # pylint: disable=no-member
serialization_dir,
recover,
cache_directory,
cache_prefix)
trainer = Trainer.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
iterator=pieces.iterator,
train_data=pieces.train_dataset,
validation_data=pieces.validation_dataset,
params=pieces.params,
validation_iterator=pieces.validation_iterator)
evaluation_iterator = pieces.validation_iterator or pieces.iterator
evaluation_dataset = pieces.test_dataset
else:
if evaluate_on_test:
raise ValueError(
"--evaluate-on-test only works with the default Trainer. "
"If you're using the CallbackTrainer you can use a callback "
"to evaluate at Events.TRAINING_END; otherwise you'll have "
"to run allennlp evaluate separately.")
trainer = TrainerBase.from_params(params, serialization_dir, recover,
cache_directory, cache_prefix)
evaluation_dataset = None
params.assert_empty('base train command')
try:
metrics = trainer.train()
except KeyboardInterrupt:
# if we have completed an epoch, try to create a model archive.
if os.path.exists(os.path.join(serialization_dir, _DEFAULT_WEIGHTS)):
logging.info(
"Training interrupted by the user. Attempting to create "
"a model archive using the current best epoch weights.")
archive_model(serialization_dir,
files_to_archive=params.files_to_archive)
raise
# Evaluate
if evaluation_dataset and evaluate_on_test:
logger.info(
"The model will be evaluated using the best epoch weights.")
test_metrics = evaluate(
trainer.model,
evaluation_dataset,
evaluation_iterator,
cuda_device=trainer._cuda_devices[0], # pylint: disable=protected-access,
# TODO(brendanr): Pass in an arg following Joel's trainer refactor.
batch_weight_key="")
for key, value in test_metrics.items():
metrics["test_" + key] = value
elif evaluation_dataset:
logger.info(
"To evaluate on the test set after training, pass the "
"'evaluate_on_test' flag, or use the 'allennlp evaluate' command.")
cleanup_global_logging(stdout_handler)
# Now tar up results
archive_model(serialization_dir, files_to_archive=params.files_to_archive)
dump_metrics(os.path.join(serialization_dir, "metrics.json"),
metrics,
log=True)
# We count on the trainer to have the model with best weights
return trainer.model
def _train_worker(
process_rank: int,
params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
cache_directory: str = None,
cache_prefix: str = None,
include_package: List[str] = None,
node_rank: int = 0,
master_addr: str = "127.0.0.1",
master_port: int = 29500,
world_size: int = 1,
distributed_device_ids: List[str] = None,
) -> Optional[Model]:
"""
Helper to train the configured model/experiment. In distributed mode, this is spawned as a
worker process. In a single GPU experiment, this returns the ``Model`` object and in distributed
training, nothing is returned.
# Parameters
process_rank : ``int``
The process index that is initialized using the GPU device id.
params : ``Params``
A parameter object specifying an AllenNLP Experiment.
serialization_dir : ``str``
The directory in which to save results and logs.
file_friendly_logging : ``bool``, optional (default=False)
If ``True``, we add newlines to tqdm output, even on an interactive terminal, and we slow
down tqdm's output to only once every 10 seconds.
recover : ``bool``, optional (default=False)
If ``True``, we will try to recover a training run from an existing serialization
directory. This is only intended for use when something actually crashed during the middle
of a run. For continuing training a model on new data, see the ``fine-tune`` command.
cache_directory : ``str``, optional
For caching data pre-processing. See :func:`allennlp.training.util.datasets_from_params`.
cache_prefix : ``str``, optional
For caching data pre-processing. See :func:`allennlp.training.util.datasets_from_params`.
include_package : ``List[str]``, optional
In distributed mode, since this function would have been spawned as a separate process,
the extra imports need to be done again. NOTE: This does not have any effect in single
GPU training.
node_rank : ``int``, optional
Rank of the node
world_size : ``int``, optional
The number of processes involved in distributed training.
# Returns
best_model : ``Model``
The model with the best epoch weights.
"""
prepare_global_logging(serialization_dir,
file_friendly_logging,
rank=process_rank,
world_size=world_size)
prepare_environment(params)
distributed = world_size > 1
# not using `allennlp.common.util.is_master` as the process group is yet to be initialized
master = process_rank == 0
evaluate_on_test = params.pop_bool("evaluate_on_test", False)
if distributed:
# Since the worker is spawned and not forked, the extra imports
# need to be done again.
if include_package is not None:
for package_name in include_package:
import_submodules(package_name)
num_procs_per_node = len(distributed_device_ids)
# The Unique identifier of the worker process among all the processes in the
# distributed training group is computed here. This is used while initializing
# the process group using `init_process_group`
global_rank = node_rank * num_procs_per_node + process_rank
# In distributed training, the configured device is always going to be a list.
# The corresponding gpu id for the particular worker is obtained by picking the id
# from the device list with the rank as index
gpu_id = distributed_device_ids[process_rank] # type: ignore
# Till now, "cuda_device" might not be set in the trainer params.
# But a worker trainer needs to only know about its specific GPU id.
params["trainer"]["cuda_device"] = gpu_id
params["trainer"]["world_size"] = world_size
params["trainer"]["distributed"] = True
torch.cuda.set_device(gpu_id)
dist.init_process_group(
backend="nccl",
init_method=f"tcp://{master_addr}:{master_port}",
world_size=world_size,
rank=global_rank,
)
logging.info(f"Process group of world size {world_size} initialized "
f"for distributed training in worker {global_rank}")
trainer_type = params.get("trainer", {}).get("type", "default")
if trainer_type == "default":
# Special logic to instantiate backward-compatible trainer.
pieces = TrainerPieces.from_params(params, serialization_dir, recover,
cache_directory, cache_prefix)
trainer = Trainer.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
iterator=pieces.iterator,
train_data=pieces.train_dataset,
validation_data=pieces.validation_dataset,
params=pieces.params,
validation_iterator=pieces.validation_iterator,
)
evaluation_iterator = pieces.validation_iterator or pieces.iterator
evaluation_dataset = pieces.test_dataset
else:
if evaluate_on_test:
raise ValueError(
"--evaluate-on-test only works with the default Trainer. "
"If you're using the CallbackTrainer you can use a callback "
"to evaluate at Events.TRAINING_END; otherwise you'll have "
"to run allennlp evaluate separately.")
trainer = TrainerBase.from_params(params, serialization_dir, recover,
cache_directory, cache_prefix)
evaluation_dataset = None
params.assert_empty("base train command")
try:
if distributed: # let the setup get ready for all the workers
dist.barrier()
metrics = trainer.train()
except KeyboardInterrupt:
# if we have completed an epoch, try to create a model archive.
if master and os.path.exists(
os.path.join(serialization_dir, _DEFAULT_WEIGHTS)):
logging.info(
"Training interrupted by the user. Attempting to create "
"a model archive using the current best epoch weights.")
archive_model(serialization_dir,
files_to_archive=params.files_to_archive)
raise
if master:
if evaluation_dataset and evaluate_on_test:
logger.info(
"The model will be evaluated using the best epoch weights.")
test_metrics = evaluate(
trainer.model,
evaluation_dataset,
evaluation_iterator,
cuda_device=trainer.cuda_device,
# TODO(brendanr): Pass in an arg following Joel's trainer refactor.
batch_weight_key="",
)
for key, value in test_metrics.items():
metrics["test_" + key] = value
elif evaluation_dataset:
logger.info(
"To evaluate on the test set after training, pass the "
"'evaluate_on_test' flag, or use the 'allennlp evaluate' command."
)
dump_metrics(os.path.join(serialization_dir, "metrics.json"),
metrics,
log=True)
if not distributed:
return trainer.model
return None # to make mypy happy
def from_params( # type: ignore
cls,
params: Params,
serialization_dir: str,
recover: bool = False,
local_rank: int = 0,
) -> "Trainer":
from allennlp.training.trainer import Trainer
from allennlp.training.trainer_pieces import TrainerPieces
config = dict(as_flat_dict(params.as_dict()))
pieces = TrainerPieces.from_params(params, serialization_dir, recover)
model = pieces.model
serialization_dir = serialization_dir
iterator = pieces.iterator
train_data = pieces.train_dataset
validation_data = pieces.validation_dataset
params = pieces.params
validation_iterator = pieces.validation_iterator
patience = params.pop_int("patience", None)
validation_metric = params.pop("validation_metric", "-loss")
shuffle = params.pop_bool("shuffle", True)
num_epochs = params.pop_int("num_epochs", 20)
cuda_device = parse_cuda_device(params.pop("cuda_device", -1))
grad_norm = params.pop_float("grad_norm", None)
grad_clipping = params.pop_float("grad_clipping", None)
lr_scheduler_params = params.pop("learning_rate_scheduler", None)
momentum_scheduler_params = params.pop("momentum_scheduler", None)
check_for_gpu(cuda_device)
if cuda_device >= 0:
# Moving model to GPU here so that the optimizer state gets constructed on
# the right device.
model = model.cuda(cuda_device)
parameters = [[n, p] for n, p in model.named_parameters() if p.requires_grad]
optimizer = Optimizer.from_params(parameters, params.pop("optimizer"))
if "moving_average" in params:
moving_average = MovingAverage.from_params(
params.pop("moving_average"), parameters=parameters
)
else:
moving_average = None
if lr_scheduler_params:
lr_scheduler = LearningRateScheduler.from_params(optimizer, lr_scheduler_params)
else:
lr_scheduler = None
if momentum_scheduler_params:
momentum_scheduler = MomentumScheduler.from_params(optimizer, momentum_scheduler_params)
else:
momentum_scheduler = None
if "checkpointer" in params:
if (
"keep_serialized_model_every_num_seconds" in params
or "num_serialized_models_to_keep" in params
):
raise ConfigurationError(
"Checkpointer may be initialized either from the 'checkpointer' key or from the "
"keys 'num_serialized_models_to_keep' and 'keep_serialized_model_every_num_seconds'"
" but the passed config uses both methods."
)
checkpointer = Checkpointer.from_params(params.pop("checkpointer"))
else:
num_serialized_models_to_keep = params.pop_int("num_serialized_models_to_keep", 20)
keep_serialized_model_every_num_seconds = params.pop_int(
"keep_serialized_model_every_num_seconds", None
)
checkpointer = Checkpointer(
serialization_dir=serialization_dir,
num_serialized_models_to_keep=num_serialized_models_to_keep,
keep_serialized_model_every_num_seconds=keep_serialized_model_every_num_seconds,
)
model_save_interval = params.pop_float("model_save_interval", None)
summary_interval = params.pop_int("summary_interval", 100)
histogram_interval = params.pop_int("histogram_interval", None)
should_log_parameter_statistics = params.pop_bool("should_log_parameter_statistics", True)
should_log_learning_rate = params.pop_bool("should_log_learning_rate", False)
log_batch_size_period = params.pop_int("log_batch_size_period", None)
distributed = params.pop_bool("distributed", False)
world_size = params.pop_int("world_size", 1)
num_gradient_accumulation_steps = params.pop("num_gradient_accumulation_steps", 1)
lang_mean_dir = params.pop("ft_lang_mean_dir", None)
if lang_mean_dir:
try:
assert model._lang_means is not None
lang_mean = get_lang_mean(lang_mean_dir)
model.add_ft_lang_mean_to_lang_means(lang_mean)
except (AttributeError, AssertionError) as e:
pass
writer = None
wandb_config = params.pop("wandb", None)
if wandb_config is not None:
writer = WandBWriter(config, model, wandb_config)
params.assert_empty(cls.__name__)
return cls(
model,
optimizer,
iterator,
train_data,
validation_data,
patience=patience,
validation_metric=validation_metric,
validation_iterator=validation_iterator,
shuffle=shuffle,
num_epochs=num_epochs,
serialization_dir=serialization_dir,
cuda_device=cuda_device,
grad_norm=grad_norm,
grad_clipping=grad_clipping,
learning_rate_scheduler=lr_scheduler,
momentum_scheduler=momentum_scheduler,
checkpointer=checkpointer,
model_save_interval=model_save_interval,
summary_interval=summary_interval,
histogram_interval=histogram_interval,
should_log_parameter_statistics=should_log_parameter_statistics,
should_log_learning_rate=should_log_learning_rate,
log_batch_size_period=log_batch_size_period,
moving_average=moving_average,
distributed=distributed,
local_rank=local_rank,
world_size=world_size,
num_gradient_accumulation_steps=num_gradient_accumulation_steps,
writer=writer,
)
def from_params(
cls, # type: ignore
params: Params,
serialization_dir: str,
recover: bool,
cache_directory: str,
cache_prefix: str) -> 'MatchingTrainer':
# pylint: disable=arguments-differ
pieces = TrainerPieces.from_params(
params, # pylint: disable=no-member
serialization_dir,
recover,
cache_directory,
cache_prefix)
model = pieces.model
iterator = pieces.iterator
train_data = pieces.train_dataset
validation_data = pieces.validation_dataset
params = pieces.params
validation_iterator = pieces.validation_iterator
patience = params.pop_int("patience", None)
validation_metric = params.pop("validation_metric", "-loss")
shuffle = params.pop_bool("shuffle", True)
num_epochs = params.pop_int("num_epochs", 20)
cuda_device = parse_cuda_device(params.pop("cuda_device", -1))
grad_norm = params.pop_float("grad_norm", None)
grad_clipping = params.pop_float("grad_clipping", None)
lr_scheduler_params = params.pop("learning_rate_scheduler", None)
momentum_scheduler_params = params.pop("momentum_scheduler", None)
retrieve_text = params.pop("retrieve_text", True)
if isinstance(cuda_device, list):
model_device = cuda_device[0]
else:
model_device = cuda_device
if model_device >= 0:
# Moving model to GPU here so that the optimizer state gets constructed on
# the right device.
model = model.cuda(model_device)
parameters = [[n, p] for n, p in model.named_parameters()
if p.requires_grad]
optimizer = Optimizer.from_params(parameters, params.pop("optimizer"))
if "moving_average" in params:
moving_average = MovingAverage.from_params(
params.pop("moving_average"), parameters=parameters)
else:
moving_average = None
if lr_scheduler_params:
lr_scheduler = LearningRateScheduler.from_params(
optimizer, lr_scheduler_params)
else:
lr_scheduler = None
if momentum_scheduler_params:
momentum_scheduler = MomentumScheduler.from_params(
optimizer, momentum_scheduler_params)
else:
momentum_scheduler = None
if 'checkpointer' in params:
if 'keep_serialized_model_every_num_seconds' in params or \
'num_serialized_models_to_keep' in params:
raise ConfigurationError(
"Checkpointer may be initialized either from the 'checkpointer' key or from the "
"keys 'num_serialized_models_to_keep' and 'keep_serialized_model_every_num_seconds'"
" but the passed config uses both methods.")
checkpointer = Checkpointer.from_params(params.pop("checkpointer"))
else:
num_serialized_models_to_keep = params.pop_int(
"num_serialized_models_to_keep", 20)
keep_serialized_model_every_num_seconds = params.pop_int(
"keep_serialized_model_every_num_seconds", None)
checkpointer = Checkpointer(
serialization_dir=serialization_dir,
num_serialized_models_to_keep=num_serialized_models_to_keep,
keep_serialized_model_every_num_seconds=
keep_serialized_model_every_num_seconds)
model_save_interval = params.pop_float("model_save_interval", None)
summary_interval = params.pop_int("summary_interval", 100)
histogram_interval = params.pop_int("histogram_interval", None)
should_log_parameter_statistics = params.pop_bool(
"should_log_parameter_statistics", True)
should_log_learning_rate = params.pop_bool("should_log_learning_rate",
False)
log_batch_size_period = params.pop_int("log_batch_size_period", None)
params.assert_empty(cls.__name__)
return cls(
model,
optimizer,
iterator,
train_data,
validation_data,
patience=patience,
retrieve_text=retrieve_text,
validation_metric=validation_metric,
validation_iterator=validation_iterator,
shuffle=shuffle,
num_epochs=num_epochs,
serialization_dir=serialization_dir,
cuda_device=cuda_device,
grad_norm=grad_norm,
grad_clipping=grad_clipping,
learning_rate_scheduler=lr_scheduler,
momentum_scheduler=momentum_scheduler,
checkpointer=checkpointer,
model_save_interval=model_save_interval,
summary_interval=summary_interval,
histogram_interval=histogram_interval,
should_log_parameter_statistics=should_log_parameter_statistics,
should_log_learning_rate=should_log_learning_rate,
log_batch_size_period=log_batch_size_period,
moving_average=moving_average)
return self.name
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--action",
type=Action,
choices=list(Action),
required=True)
parser.add_argument("--config", required=True)
parser.add_argument("--serialization-dir", required=True)
parser.add_argument("--batch-count", type=int, default=0)
parser.add_argument("--assume-multiprocess-types", action="store_true")
args = parser.parse_args()
params = Params.from_file(args.config)
pieces = TrainerPieces.from_params(params, args.serialization_dir)
raw_generator = pieces.iterator(pieces.train_dataset,
num_epochs=1,
shuffle=True)
if args.action is Action.log:
log_iterable(raw_generator, args.assume_multiprocess_types)
elif args.action is Action.time:
time_iterable(raw_generator, args.batch_count)
elif args.action is Action.first:
time_to_first(raw_generator)
else:
raise Exception(f"Unaccounted for action {action}")
def from_params( # type: ignore
cls,
params: Params,
serialization_dir: str,
recover: bool = False,
) -> "CallbackTrainer":
pieces = TrainerPieces.from_params(params, serialization_dir, recover)
model = pieces.model
params = pieces.params
validation_iterator = pieces.validation_iterator or pieces.iterator
shuffle = params.pop_bool("shuffle", True)
num_epochs = params.pop_int("num_epochs", 20)
cuda_device = parse_cuda_device(params.pop("cuda_device", -1))
check_for_gpu(cuda_device)
if cuda_device >= 0:
# Moving model to GPU here so that the optimizer state gets constructed on
# the right device.
model = model.cuda(cuda_device)
parameters = [[n, p] for n, p in model.named_parameters()
if p.requires_grad]
optimizer = Optimizer.from_params(parameters, params.pop("optimizer"))
callbacks_params = params.pop("callbacks", [])
callbacks: List[Callback] = [
Callback.from_params(
params=callback_params,
model=model,
optimizer=optimizer,
instances=pieces.train_dataset,
iterator=pieces.iterator,
shuffle=shuffle,
validation_data=pieces.validation_dataset,
validation_iterator=validation_iterator,
serialization_dir=serialization_dir,
) for callback_params in callbacks_params
]
distributed = params.pop_bool("distributed", False)
world_size = params.pop_int("world_size", 1)
if distributed:
rank = cuda_device
else:
rank = 0
params.assert_empty(cls.__name__)
return cls(
model,
pieces.train_dataset,
pieces.iterator,
optimizer,
num_epochs=num_epochs,
shuffle=shuffle,
serialization_dir=serialization_dir,
cuda_device=cuda_device,
callbacks=callbacks,
distributed=distributed,
rank=rank,
world_size=world_size,
)
def train_model(
params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
force: bool = False,
cache_directory: str = None,
cache_prefix: str = None,
) -> Model:
create_serialization_dir(params, serialization_dir, recover, force)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
stdout_handler = prepare_global_logging(serialization_dir,
file_friendly_logging)
prepare_environment(params)
cuda_device = params.params.get("trainer").get("cuda_device", -1)
check_for_gpu(cuda_device)
evaluate_on_test = params.pop_bool("evaluate_on_test", False)
trainer_type = params.get("trainer", {}).get("type", "default")
if True:
# Special logic to instantiate backward-compatible trainer.
pieces = TrainerPieces.from_params(params, serialization_dir, recover,
cache_directory, cache_prefix)
logger.info("Using MultiTrainer")
from lm.trainining.MultiTaskTrainer import MultiTaskTrainer
# MultiTrainer
trainer = MultiTaskTrainer.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
iterator=pieces.iterator,
train_data=pieces.train_dataset,
validation_data=pieces.validation_dataset,
params=pieces.params,
validation_iterator=pieces.validation_iterator,
)
evaluation_iterator = pieces.validation_iterator or pieces.iterator
evaluation_dataset = pieces.test_dataset
else:
if evaluate_on_test:
raise ValueError(
"--evaluate-on-test only works with the default Trainer. "
"If you're using the CallbackTrainer you can use a callback "
"to evaluate at Events.TRAINING_END; otherwise you'll have "
"to run allennlp evaluate separately.")
"""
The only main difference
"""
print("Using MultuTrainer")
logger.info("Using MultiTrainer")
trainer = MultiTrainer.from_params(params, serialization_dir, recover,
cache_directory, cache_prefix)
evaluation_dataset = None
params.assert_empty("base train command")
try:
metrics = trainer.train()
except KeyboardInterrupt:
# if we have completed an epoch, try to create a model archive.
if os.path.exists(os.path.join(serialization_dir, _DEFAULT_WEIGHTS)):
logging.info(
"Training interrupted by the user. Attempting to create "
"a model archive using the current best epoch weights.")
archive_model(serialization_dir,
files_to_archive=params.files_to_archive)
raise
# Evaluate
if evaluation_dataset and evaluate_on_test:
logger.info(
"The model will be evaluated using the best epoch weights.")
test_metrics = evaluate(
trainer.model,
evaluation_dataset,
evaluation_iterator,
cuda_device=trainer._cuda_devices[0],
# TODO(brendanr): Pass in an arg following Joel's trainer refactor.
batch_weight_key="",
)
for key, value in test_metrics.items():
metrics["test_" + key] = value
elif evaluation_dataset:
logger.info(
"To evaluate on the test set after training, pass the "
"'evaluate_on_test' flag, or use the 'allennlp evaluate' command.")
cleanup_global_logging(stdout_handler)
# Now tar up results
archive_model(serialization_dir, files_to_archive=params.files_to_archive)
dump_metrics(os.path.join(serialization_dir, "metrics.json"),
metrics,
log=True)
# We count on the trainer to have the model with best weights
return trainer.model
def from_params(
cls, # type: ignore
params: Params,
serialization_dir: str,
recover: bool = False,
cache_directory: str = None,
cache_prefix: str = None) -> 'Trainer':
# pylint: disable=arguments-differ
# We have to call TrainerPieces.from_params since we are using our own Trainer
pieces = TrainerPieces.from_params(params, serialization_dir, recover)
model = pieces.model
serialization_dir = serialization_dir
iterator = pieces.iterator
train_data = pieces.train_dataset
validation_data = pieces.validation_dataset
validation_iterator = pieces.validation_iterator
params = pieces.params
patience = params.pop_int("patience", None)
validation_metric = params.pop("validation_metric", "-loss")
shuffle = params.pop_bool("shuffle", True)
accumulation_steps = params.pop("accumulation_steps", 0)
opt_level = params.pop("opt_level", "O1")
num_epochs = params.pop_int("num_epochs", 20)
cuda_device = parse_cuda_device(params.pop("cuda_device", -1))
grad_norm = params.pop_float("grad_norm", None)
grad_clipping = params.pop_float("grad_clipping", None)
lr_scheduler_params = params.pop("learning_rate_scheduler", None)
momentum_scheduler_params = params.pop("momentum_scheduler", None)
half_precision = params.pop("half_precision", False)
warmup_proportion = params.pop("warmup_proportion", None)
pretrained_model = params.pop("pretrained_model", None)
if pretrained_model:
logger.info('Loading pretrained model from', pretrained_model)
model = load_archive(pretrained_model).model
model._discriminative_loss_weight = 1 # TODO: fix this hack
if isinstance(cuda_device, list):
model_device = cuda_device[0]
else:
model_device = cuda_device
if model_device >= 0:
# Moving model to GPU here so that the optimizer state gets constructed on
# the right device.
model = model.cuda(model_device)
parameters = [[n, p] for n, p in model.named_parameters()
if p.requires_grad]
optimizer = Optimizer.from_params(parameters, params.pop("optimizer"))
if "moving_average" in params:
moving_average = MovingAverage.from_params(
params.pop("moving_average"), parameters=parameters)
else:
moving_average = None
if lr_scheduler_params:
lr_scheduler = LearningRateScheduler.from_params(
optimizer, lr_scheduler_params)
else:
lr_scheduler = None
if momentum_scheduler_params:
momentum_scheduler = MomentumScheduler.from_params(
optimizer, momentum_scheduler_params)
else:
momentum_scheduler = None
if 'checkpointer' in params:
if 'keep_serialized_model_every_num_seconds' in params or \
'num_serialized_models_to_keep' in params:
raise ConfigurationError(
"Checkpointer may be initialized either from the 'checkpointer' key or from the "
"keys 'num_serialized_models_to_keep' and 'keep_serialized_model_every_num_seconds'"
" but the passed config uses both methods.")
checkpointer = Checkpointer.from_params(params.pop("checkpointer"))
else:
num_serialized_models_to_keep = params.pop_int(
"num_serialized_models_to_keep", 20)
keep_serialized_model_every_num_seconds = params.pop_int(
"keep_serialized_model_every_num_seconds", None)
checkpointer = Checkpointer(
serialization_dir=serialization_dir,
num_serialized_models_to_keep=num_serialized_models_to_keep,
keep_serialized_model_every_num_seconds=
keep_serialized_model_every_num_seconds)
model_save_interval = params.pop_float("model_save_interval", None)
summary_interval = params.pop_int("summary_interval", 100)
histogram_interval = params.pop_int("histogram_interval", None)
should_log_parameter_statistics = params.pop_bool(
"should_log_parameter_statistics", True)
should_log_learning_rate = params.pop_bool("should_log_learning_rate",
False)
log_batch_size_period = params.pop_int("log_batch_size_period", None)
params.assert_empty(cls.__name__)
return cls(
model,
optimizer,
iterator,
train_data,
validation_data,
patience=patience,
validation_metric=validation_metric,
validation_iterator=validation_iterator,
shuffle=shuffle,
accumulation_steps=accumulation_steps,
opt_level=opt_level,
num_epochs=num_epochs,
serialization_dir=serialization_dir,
cuda_device=cuda_device,
grad_norm=grad_norm,
grad_clipping=grad_clipping,
learning_rate_scheduler=lr_scheduler,
momentum_scheduler=momentum_scheduler,
checkpointer=checkpointer,
model_save_interval=model_save_interval,
summary_interval=summary_interval,
histogram_interval=histogram_interval,
should_log_parameter_statistics=should_log_parameter_statistics,
should_log_learning_rate=should_log_learning_rate,
log_batch_size_period=log_batch_size_period,
moving_average=moving_average,
half_precision=half_precision,
warmup_proportion=warmup_proportion)