def train_model(params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False) -> 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.
Returns
-------
best_model: ``Model``
The model with the best epoch weights.
"""
prepare_environment(params)
create_serialization_dir(params, serialization_dir, recover)
prepare_global_logging(serialization_dir, file_friendly_logging)
check_for_gpu(params.get('trainer').get('cuda_device', -1))
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
all_datasets = datasets_from_params(params)
datasets_for_vocab_creation = set(params.pop("datasets_for_vocab_creation", all_datasets))
for dataset in datasets_for_vocab_creation:
if dataset not in all_datasets:
raise ConfigurationError(f"invalid 'dataset_for_vocab_creation' {dataset}")
logger.info("From dataset instances, %s will be considered for vocabulary creation.",
", ".join(datasets_for_vocab_creation))
vocab = Vocabulary.from_params(
params.pop("vocabulary", {}),
(instance for key, dataset in all_datasets.items()
for instance in dataset
if key in datasets_for_vocab_creation)
)
model = Model.from_params(vocab=vocab, params=params.pop('model'))
# Initializing the model can have side effect of expanding the vocabulary
vocab.save_to_files(os.path.join(serialization_dir, "vocabulary"))
iterator = DataIterator.from_params(params.pop("iterator"))
iterator.index_with(vocab)
validation_iterator_params = params.pop("validation_iterator", None)
if validation_iterator_params:
validation_iterator = DataIterator.from_params(validation_iterator_params)
validation_iterator.index_with(vocab)
else:
validation_iterator = None
train_data = all_datasets['train']
validation_data = all_datasets.get('validation')
test_data = all_datasets.get('test')
trainer_params = params.pop("trainer")
no_grad_regexes = trainer_params.pop("no_grad", ())
for name, parameter in model.named_parameters():
if any(re.search(regex, name) for regex in no_grad_regexes):
parameter.requires_grad_(False)
frozen_parameter_names, tunable_parameter_names = \
get_frozen_and_tunable_parameter_names(model)
logger.info("Following parameters are Frozen (without gradient):")
for name in frozen_parameter_names:
logger.info(name)
logger.info("Following parameters are Tunable (with gradient):")
for name in tunable_parameter_names:
logger.info(name)
trainer_choice = trainer_params.pop_choice("type",
Trainer.list_available(),
default_to_first_choice=True)
trainer = Trainer.by_name(trainer_choice).from_params(model=model,
serialization_dir=serialization_dir,
iterator=iterator,
train_data=train_data,
validation_data=validation_data,
params=trainer_params,
validation_iterator=validation_iterator)
evaluate_on_test = params.pop_bool("evaluate_on_test", False)
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
# Now tar up results
archive_model(serialization_dir, files_to_archive=params.files_to_archive)
logger.info("Loading the best epoch weights.")
best_model_state_path = os.path.join(serialization_dir, 'best.th')
best_model_state = torch.load(best_model_state_path)
best_model = model
best_model.load_state_dict(best_model_state)
if test_data and evaluate_on_test:
logger.info("The model will be evaluated using the best epoch weights.")
test_metrics = evaluate(
best_model, test_data, validation_iterator or iterator,
cuda_device=trainer._cuda_devices[0] # pylint: disable=protected-access
)
for key, value in test_metrics.items():
metrics["test_" + key] = value
elif test_data:
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)
return best_model
def train_model(
params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
force: bool = False,
node_rank: int = 0,
include_package: List[str] = None,
batch_weight_key: str = "",
) -> 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 ``Model.from_archive``.
force : ``bool``, optional (default=False)
If ``True``, we will overwrite the serialization directory if it already exists.
node_rank : ``int``, optional
Rank of the current node in distributed training
include_package : ``List[str]``, optional
In distributed mode, extra packages mentioned will be imported in trainer workers.
batch_weight_key : ``str``, optional (default="")
If non-empty, name of metric used to weight the loss on a per-batch basis.
# Returns
best_model : ``Model``
The model with the best epoch weights.
"""
training_util.create_serialization_dir(params, serialization_dir, recover,
force)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
distributed_params = params.params.pop("distributed", None)
# If distributed isn't in the config and the config contains strictly
# one cuda device, we just run a single training process.
if distributed_params is None:
model = _train_worker(
process_rank=0,
params=params,
serialization_dir=serialization_dir,
file_friendly_logging=file_friendly_logging,
include_package=include_package,
batch_weight_key=batch_weight_key,
)
archive_model(serialization_dir)
return model
# Otherwise, we are running multiple processes for training.
else:
# We are careful here so that we can raise a good error if someone
# passed the wrong thing - cuda_devices are required.
device_ids = distributed_params.pop("cuda_devices", None)
multi_device = isinstance(device_ids, list) and len(device_ids) > 1
num_nodes = distributed_params.pop("num_nodes", 1)
if not (multi_device or num_nodes > 1):
raise ConfigurationError(
"Multiple cuda devices/nodes need to be configured to run distributed training."
)
check_for_gpu(device_ids)
master_addr = distributed_params.pop("master_address", "127.0.0.1")
master_port = distributed_params.pop("master_port", 29500)
num_procs = len(device_ids)
world_size = num_nodes * num_procs
logging.info(
f"Switching to distributed training mode since multiple GPUs are configured"
f"Master is at: {master_addr}:{master_port} | Rank of this node: {node_rank} | "
f"Number of workers in this node: {num_procs} | Number of nodes: {num_nodes} | "
f"World size: {world_size}")
# Creating `Vocabulary` objects from workers could be problematic since
# the data iterators in each worker will yield only `rank` specific
# instances. Hence it is safe to construct the vocabulary and write it
# to disk before initializing the distributed context. The workers will
# load the vocabulary from the path specified.
if params.get("vocabulary", Params({})).get("type",
"") != "from_files":
vocab = training_util.make_vocab_from_params(
params.duplicate(), serialization_dir)
params["vocabulary"] = {
"type": "from_files",
"directory": os.path.join(serialization_dir, "vocabulary"),
"padding_token": vocab._padding_token,
"oov_token": vocab._oov_token,
}
mp.spawn(
_train_worker,
args=(
params.duplicate(),
serialization_dir,
file_friendly_logging,
include_package,
batch_weight_key,
node_rank,
master_addr,
master_port,
world_size,
device_ids,
),
nprocs=num_procs,
)
archive_model(serialization_dir)
model = Model.load(params, serialization_dir)
return model
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_model(
params: Params,
serialization_dir: Union[str, PathLike],
recover: bool = False,
force: bool = False,
node_rank: int = 0,
include_package: List[str] = None,
dry_run: bool = False,
file_friendly_logging: bool = False,
) -> Optional[Model]:
"""
Trains the model specified in the given [`Params`](../common/params.md#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.
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 `Model.from_archive`.
force : `bool`, optional (default=`False`)
If `True`, we will overwrite the serialization directory if it already exists.
node_rank : `int`, optional
Rank of the current node in distributed training
include_package : `List[str]`, optional
In distributed mode, extra packages mentioned will be imported in trainer workers.
dry_run : `bool`, optional (default=`False`)
Do not train a model, but create a vocabulary, show dataset statistics and other training
information.
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.
# Returns
best_model : `Optional[Model]`
The model with the best epoch weights or `None` if in dry run.
"""
common_logging.FILE_FRIENDLY_LOGGING = file_friendly_logging
training_util.create_serialization_dir(params, serialization_dir, recover,
force)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
distributed_params = params.params.pop("distributed", None)
# If distributed isn't in the config and the config contains strictly
# one cuda device, we just run a single training process.
if distributed_params is None:
model = _train_worker(
process_rank=0,
params=params,
serialization_dir=serialization_dir,
include_package=include_package,
dry_run=dry_run,
file_friendly_logging=file_friendly_logging,
)
if not dry_run:
archive_model(serialization_dir)
return model
# Otherwise, we are running multiple processes for training.
else:
# We are careful here so that we can raise a good error if someone
# passed the wrong thing - cuda_devices are required.
device_ids = distributed_params.pop("cuda_devices", None)
multi_device = isinstance(device_ids, list) and len(device_ids) > 1
num_nodes = distributed_params.pop("num_nodes", 1)
if not (multi_device or num_nodes > 1):
raise ConfigurationError(
"Multiple cuda devices/nodes need to be configured to run distributed training."
)
check_for_gpu(device_ids)
master_addr = distributed_params.pop("master_address", "127.0.0.1")
master_port = distributed_params.pop("master_port", 29500)
num_procs = len(device_ids)
world_size = num_nodes * num_procs
# Creating `Vocabulary` objects from workers could be problematic since
# the data loaders in each worker will yield only `rank` specific
# instances. Hence it is safe to construct the vocabulary and write it
# to disk before initializing the distributed context. The workers will
# load the vocabulary from the path specified.
vocab_dir = os.path.join(serialization_dir, "vocabulary")
if recover:
vocab = Vocabulary.from_files(vocab_dir)
else:
vocab = training_util.make_vocab_from_params(
params.duplicate(),
serialization_dir,
print_statistics=dry_run)
params["vocabulary"] = {
"type": "from_files",
"directory": vocab_dir,
"padding_token": vocab._padding_token,
"oov_token": vocab._oov_token,
}
logging.info(
"Switching to distributed training mode since multiple GPUs are configured | "
f"Master is at: {master_addr}:{master_port} | Rank of this node: {node_rank} | "
f"Number of workers in this node: {num_procs} | Number of nodes: {num_nodes} | "
f"World size: {world_size}")
mp.spawn(
_train_worker,
args=(
params.duplicate(),
serialization_dir,
include_package,
dry_run,
node_rank,
master_addr,
master_port,
world_size,
device_ids,
file_friendly_logging,
),
nprocs=num_procs,
)
if dry_run:
return None
else:
archive_model(serialization_dir)
model = Model.load(params, serialization_dir)
return model
def train_model(params: Params,
serialization_dir: str,
results_fn: str,
file_friendly_logging: bool = False,
recover: bool = False,
force: bool = False) -> Tuple[Model, Dict[str, Any]]:
prepare_environment(params)
create_serialization_dir(params, serialization_dir, recover, force)
prepare_global_logging(serialization_dir, file_friendly_logging)
cuda_device = params.params.get('trainer').get('cuda_device', -1)
if isinstance(cuda_device, list):
for device in cuda_device:
check_for_gpu(device)
else:
check_for_gpu(cuda_device)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
all_datasets = datasets_from_params(params)
datasets_for_vocab_creation = set(
params.pop("datasets_for_vocab_creation", all_datasets))
for dataset in datasets_for_vocab_creation:
if dataset not in all_datasets:
raise ConfigurationError(
f"invalid 'dataset_for_vocab_creation' {dataset}")
logger.info(
"From dataset instances, %s will be considered for vocabulary creation.",
", ".join(datasets_for_vocab_creation))
vocab = Vocabulary.from_params(
params.pop("vocabulary", {}),
(instance for key, dataset in all_datasets.items()
for instance in dataset if key in datasets_for_vocab_creation))
model = Model.from_params(vocab=vocab, params=params.pop('model'))
# Initializing the model can have side effect of expanding the vocabulary
vocab.save_to_files(os.path.join(serialization_dir, "vocabulary"))
iterator = DataIterator.from_params(params.pop("iterator"))
iterator.index_with(vocab)
validation_iterator_params = params.pop("validation_iterator", None)
if validation_iterator_params:
validation_iterator = DataIterator.from_params(
validation_iterator_params)
validation_iterator.index_with(vocab)
else:
validation_iterator = None
held_out_iterator_params = params.pop("held_out_iterator", None)
if held_out_iterator_params:
held_out_iterator = DataIterator.from_params(held_out_iterator_params)
held_out_iterator.index_with(vocab)
else:
held_out_iterator = None
train_data = all_datasets['train']
held_out_train_data = all_datasets.get('held_out_train')
validation_data = all_datasets.get('validation')
test_data = all_datasets.get('test')
trainer_params = params.pop("trainer")
no_grad_regexes = trainer_params.pop("no_grad", ())
for name, parameter in model.named_parameters():
if any(re.search(regex, name) for regex in no_grad_regexes):
parameter.requires_grad_(False)
frozen_parameter_names, tunable_parameter_names = \
get_frozen_and_tunable_parameter_names(model)
logger.info("Following parameters are Frozen (without gradient):")
for name in frozen_parameter_names:
logger.info(name)
logger.info("Following parameters are Tunable (with gradient):")
for name in tunable_parameter_names:
logger.info(name)
trainer_choice = trainer_params.pop_choice("type",
Trainer.list_available(),
default_to_first_choice=True)
trainer = Trainer.by_name(trainer_choice).from_params(
model=model,
serialization_dir=serialization_dir,
iterator=iterator,
train_data=train_data,
held_out_train_data=held_out_train_data,
validation_data=validation_data,
params=trainer_params,
validation_iterator=validation_iterator,
held_out_iterator=held_out_iterator)
evaluate_on_test = params.pop_bool("evaluate_on_test", False)
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
# Now tar up results
archive_model(serialization_dir, files_to_archive=params.files_to_archive)
logger.info("Loading the best epoch weights.")
best_model_state_path = os.path.join(serialization_dir, 'best.th')
best_model_state = torch.load(best_model_state_path)
best_model = model
best_model.load_state_dict(best_model_state)
if test_data and evaluate_on_test:
logger.info(
"The model will be evaluated using the best epoch weights.")
test_metrics = evaluate(
best_model,
test_data,
validation_iterator or iterator,
cuda_device=trainer._cuda_devices[0] # pylint: disable=protected-access
)
for key, value in test_metrics.items():
metrics["test_" + key] = value
elif test_data:
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(results_dir, results_fn), metrics, log=True)
return best_model, metrics
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,
node_rank: int = 0,
include_package: List[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`.
node_rank : ``int``, optional
Rank of the current node in distributed training
include_package : ``List[str]``, optional
In distributed mode, extra packages mentioned will be imported in trainer workers.
# Returns
best_model : ``Model``
The model with the best epoch weights.
"""
create_serialization_dir(params, serialization_dir, recover, force)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
distributed_params = params.params.pop("distributed", None)
# If distributed isn't in the config and the config contains strictly
# one cuda device, we just run a single training process.
if distributed_params is None:
model = _train_worker(
process_rank=0,
params=params,
serialization_dir=serialization_dir,
file_friendly_logging=file_friendly_logging,
recover=recover,
cache_directory=cache_directory,
cache_prefix=cache_prefix,
include_package=include_package,
)
archive_model(serialization_dir,
files_to_archive=params.files_to_archive)
return model
# Otherwise, we are running multiple processes for training.
else:
# We are careful here so that we can raise a good error if someone
# passed the wrong thing - cuda_devices are required.
device_ids = distributed_params.pop("cuda_devices", None)
multi_device = isinstance(device_ids, list) and len(device_ids) > 1
num_nodes = distributed_params.pop("num_nodes", 1)
if not (multi_device or num_nodes > 1):
raise ConfigurationError(
"Multiple cuda devices/nodes need to be configured to run distributed training."
)
check_for_gpu(device_ids)
master_addr = distributed_params.pop("master_address", "127.0.0.1")
master_port = distributed_params.pop("master_port", 29500)
num_procs = len(device_ids)
world_size = num_nodes * num_procs
os.environ["MASTER_ADDR"] = master_addr
os.environ["MASTER_PORT"] = str(master_port)
os.environ["WORLD_SIZE"] = str(world_size)
logging.info(
f"Switching to distributed training mode since multiple GPUs are configured"
f"Master is at: {master_addr}:{master_port} | Rank of this node: {node_rank} | "
f"Number of workers in this node: {num_procs} | Number of nodes: {num_nodes} | "
f"World size: {world_size}")
# Creating `Vocabulary` objects from workers could be problematic since the data iterators
# in each worker will yield only `rank` specific instances. Hence it is safe to construct
# the vocabulary and write it to disk before initializing the distributed context. The workers
# will load the vocabulary from the path specified.
make_vocab_from_params(params.duplicate(), serialization_dir)
params["vocabulary"] = {
"directory_path": os.path.join(serialization_dir, "vocabulary"),
"extend": False, # vocab extension would have been done above
}
mp.spawn(
_train_worker,
args=(
params.duplicate(),
serialization_dir,
file_friendly_logging,
recover,
cache_directory,
cache_prefix,
include_package,
node_rank,
master_addr,
master_port,
world_size,
device_ids,
),
nprocs=num_procs,
)
archive_model(serialization_dir,
files_to_archive=params.files_to_archive)
model = Model.load(params, serialization_dir)
return model
def train_model(params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
force: bool = False,
debate_mode: List[str] = ('f'),
judge_filename: str = None,
update_judge: bool = False,
eval_mode: bool = False,
reward_method: str = None,
detach_value_head: bool = False,
breakpoint_level: int = 0,
search_outputs_path: str = None,
accumulation_steps: int = 1,
multi_gpu: bool = False,
choice_mode: str = None,
qa_loss_weight: float = 0.,
influence_reward: bool = False,
theory_of_mind: bool = False,
num_pred_rounds: int = -1,
x_order_prob: float = 0.,
require_action: bool = False,
single_shot: bool = False) -> 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.
debate_mode : ``List[str]``
List of debate turns (e.g. aa, ar, rr, Ar) => capitalization implies search agent
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.
judge_filename : ``str``, optional (default=None)
Path to judge config or pre-trained judge model. If config, judge trained during debate. Necessary parameter
if running in debate mode.
update_judge : ``bool``, optional (default=False)
Boolean whether or not to update Judge model during debate training.
eval_mode : ``bool``, optional (default=False)
Boolean whether or not to run in eval-only mode, on test data. Does not update/train any of the models.
reward_method : ``str``, optional (default=False)
Choice of reward function (RL) or loss function (Supervised Learning) for training debate agents
detach_value_head : ``bool``, optional (default=False)
Boolean whether or not to detatch value function gradient updates from the policy network. This prevents
value function gradients from affecting policy network parameters.
breakpoint_level : ``int`` optional (default=0)
Debugging option to set breakpoint sensitivity (0 - no breakpoints).
id_to_search_filename : ``str`` optional (default=None)
Path to file with search predictions for each agent - necessary for supervised training
accumulation_steps : ``int`` (default=1)
Number of gradient steps to accumulate over before performing an update. Poor-man's batching for instances where
number of examples per batch is small (limited GPU memory)
multi_gpu : ``bool`` (default=False)
Boolean whether or not to run models/training in model parallel mode. Requires specifying GPU allocations for
trainer, judge, and debaters in the training config file (see training_config/bidaf.race.size=0.5.gpu=2.jsonnet
for example usage).
Returns
-------
best_model: ``Model``
The model with the best epoch weights.
"""
assert (
not single_shot
) or eval_mode, 'Using single shot prediction outside eval_mode not yet supported.'
assert (not single_shot) or (num_pred_rounds == -1), \
'Using single shot prediction for a specific number of rounds is not yet supported.'
# Get number of debate turns, and assert that not performing judge-only training
num_no_qa_turns = sum([(('l' in debate_turn) or ('w' in debate_turn))
for debate_turn in debate_mode])
if (qa_loss_weight > 0) and (num_no_qa_turns == 0):
warnings.warn(
'Unused argument qa_loss_weight in debate mode ' +
str(debate_mode) +
'. If this was unintentional, please remove the -q flag.',
UserWarning)
not_using_trained_debater = len(
set('ablwⅰⅱⅲⅳ').intersection(''.join(debate_mode))) == 0
if (judge_filename is not None) and not_using_trained_debater:
warnings.warn(
'Unnecessary to have debaters in debate mode ' + str(debate_mode) +
'. If this was unintentional, please remove the -j flag.',
UserWarning)
prepare_environment(params)
create_serialization_dir(params, serialization_dir, recover, force)
prepare_global_logging(serialization_dir, file_friendly_logging)
# Check that all Desired CUDA Devices exist => trainer => cuda_devices should contain list of required devices
cuda_device = params.params.get('trainer').get('cuda_device', -1)
check_for_gpu(cuda_device)
# Build Allocation Dictionary (to be passed to all future functions)
if multi_gpu:
gpu_allocations, allocation_dict = params.params.pop(
'gpu_allocations', {}), {}
assert len(gpu_allocations
) == 3, 'Must set gpu_allocations in config if multi-gpu'
for k in ['debate', 'judge', 'trainer']:
assert gpu_allocations[
k] in cuda_device, "Desired GPU not available... current: %s" % str(
cuda_device)
allocation_dict[k] = gpu_allocations[k]
else:
allocation_dict = {}
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.
params['dataset_reader'][
'debate_mode'] = debate_mode # If debate_mode requires sample duplicates
pieces = TrainerPieces.from_params(params,
serialization_dir,
cuda_device,
recover,
judge_filename=judge_filename,
update_judge=update_judge,
eval_mode=eval_mode,
reward_method=reward_method,
detach_value_head=detach_value_head,
allocation_dict=allocation_dict,
qa_loss_weight=qa_loss_weight,
influence_reward=influence_reward,
theory_of_mind=theory_of_mind) # pylint: disable=no-member
trainer = Trainer.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
debate_mode=debate_mode,
iterator=pieces.iterator,
train_data=pieces.train_dataset,
validation_data=pieces.validation_dataset,
params=pieces.params,
validation_iterator=pieces.validation_iterator,
eval_mode=eval_mode,
breakpoint_level=breakpoint_level,
search_outputs_path=search_outputs_path,
accumulation_steps=accumulation_steps,
allocation_dict=allocation_dict,
choice_mode=choice_mode,
num_pred_rounds=num_pred_rounds,
x_order_prob=x_order_prob,
require_action=require_action,
single_shot=single_shot)
evaluation_iterator = pieces.validation_iterator or pieces.iterator
evaluation_dataset = pieces.test_dataset
else:
assert (len(debate_mode)
== 1) and (debate_mode[0]
== 'f'), 'TrainerBase untested for debate training.'
trainer = TrainerBase.from_params(params, serialization_dir, recover)
evaluation_iterator = 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)) and not eval_mode:
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,
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.")
# Now tar up results
if not eval_mode:
archive_model(serialization_dir,
files_to_archive=params.files_to_archive)
dump_metrics(os.path.join(serialization_dir, "metrics.json"),
metrics,
log=True)
else:
dump_metrics(os.path.join(
serialization_dir,
"metrics.eval.d=" + '-'.join(debate_mode) + ".json"),
metrics,
log=True)
# We count on the trainer to have the model with best weights
return trainer.model
def train_model(params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
force: bool = False) -> 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.
Returns
-------
best_model: ``Model``
The model with the best epoch weights.
"""
prepare_environment(params)
create_serialization_dir(params, serialization_dir, recover, force)
prepare_global_logging(serialization_dir, file_friendly_logging)
cuda_device = params.params.get('trainer').get('cuda_device', -1)
if isinstance(cuda_device, list):
for device in cuda_device:
check_for_gpu(device)
else:
check_for_gpu(cuda_device)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
all_datasets = datasets_from_params(params)
datasets_for_vocab_creation = set(params.pop("datasets_for_vocab_creation", all_datasets))
for dataset in datasets_for_vocab_creation:
if dataset not in all_datasets:
raise ConfigurationError(f"invalid 'dataset_for_vocab_creation' {dataset}")
logger.info("From dataset instances, %s will be considered for vocabulary creation.",
", ".join(datasets_for_vocab_creation))
vocab = Vocabulary.from_params(
params.pop("vocabulary", {}),
(instance for key, dataset in all_datasets.items()
for instance in dataset
if key in datasets_for_vocab_creation)
)
model = Model.from_params(vocab=vocab, params=params.pop('model'))
# Initializing the model can have side effect of expanding the vocabulary
vocab.save_to_files(os.path.join(serialization_dir, "vocabulary"))
iterator = DataIterator.from_params(params.pop("iterator"))
iterator.index_with(vocab)
validation_iterator_params = params.pop("validation_iterator", None)
if validation_iterator_params:
validation_iterator = DataIterator.from_params(validation_iterator_params)
validation_iterator.index_with(vocab)
else:
validation_iterator = None
train_data = all_datasets['train']
validation_data = all_datasets.get('validation')
test_data = all_datasets.get('test')
trainer_params = params.pop("trainer")
no_grad_regexes = trainer_params.pop("no_grad", ())
for name, parameter in model.named_parameters():
if any(re.search(regex, name) for regex in no_grad_regexes):
parameter.requires_grad_(False)
frozen_parameter_names, tunable_parameter_names = \
get_frozen_and_tunable_parameter_names(model)
logger.info("Following parameters are Frozen (without gradient):")
for name in frozen_parameter_names:
logger.info(name)
logger.info("Following parameters are Tunable (with gradient):")
for name in tunable_parameter_names:
logger.info(name)
trainer_choice = trainer_params.pop_choice("type",
Trainer.list_available(),
default_to_first_choice=True)
trainer = Trainer.by_name(trainer_choice).from_params(model=model,
serialization_dir=serialization_dir,
iterator=iterator,
train_data=train_data,
validation_data=validation_data,
params=trainer_params,
validation_iterator=validation_iterator)
evaluate_on_test = params.pop_bool("evaluate_on_test", False)
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
# Now tar up results
archive_model(serialization_dir, files_to_archive=params.files_to_archive)
logger.info("Loading the best epoch weights.")
best_model_state_path = os.path.join(serialization_dir, 'best.th')
best_model_state = torch.load(best_model_state_path)
best_model = model
best_model.load_state_dict(best_model_state)
if test_data and evaluate_on_test:
logger.info("The model will be evaluated using the best epoch weights.")
test_metrics = evaluate(
best_model, test_data, validation_iterator or iterator,
cuda_device=trainer._cuda_devices[0] # pylint: disable=protected-access
)
for key, value in test_metrics.items():
metrics["test_" + key] = value
elif test_data:
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)
return best_model
def train_model(params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
del_models: bool = False,
del_vocab: bool = False,
convert: bool = False) -> 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.
del_models : ``bool``, optional (default=False)
If ``True``, we will delete existing models and logs if they already exist.
del_vocab : ``bool``, optional (default=False)
If ``True``, we will delete existing vocabulary if it already exists.
Returns
-------
best_model: ``Model``
The model with the best epoch weights.
"""
cuda_device = params.params.get('trainer').get('cuda_device', -1)
if cuda_device >= 0:
check_for_gpu(cuda_device)
torch.cuda.set_device(cuda_device)
# Sometimes we might change the config a bit but still want to continue training
# if recover:
# create_serialization_dir(
# params, serialization_dir, recover, del_models)
if del_models:
for path in glob(f'{serialization_dir}/*'):
if os.path.isfile(path) and not path.endswith('config.yaml'):
os.remove(path)
log_path = f'{serialization_dir}/log'
if os.path.isdir(log_path):
shutil.rmtree(log_path)
if del_vocab:
vocab_path = f'{serialization_dir}/vocabulary'
if os.path.isdir(vocab_path):
shutil.rmtree(vocab_path)
prepare_global_logging(serialization_dir, file_friendly_logging)
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, serialization_dir, recover) # pylint: disable=no-member
trainer = Trainer.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
iterator=pieces.iterator,
train_data=pieces.corpus.train,
validation_data=pieces.corpus.valid,
params=pieces.params,
validation_iterator=pieces.validation_iterator)
evaluation_iterator = pieces.validation_iterator or pieces.iterator
evaluation_dataset = pieces.corpus.test
batch_weight_key = pieces.batch_weight_key
elif trainer_type == 'trainer_fp16_single':
params.get("trainer").pop('type')
# Special logic to instantiate backward-compatible trainer.
pieces = TrainerPieces.from_params(params, serialization_dir, recover) # pylint: disable=no-member
trainer = TrainerF16SingleTask.from_params(
model=pieces.model,
serialization_dir=serialization_dir,
files_to_archive=params.files_to_archive,
iterator=pieces.iterator,
train_data=pieces.corpus.train,
validation_data=pieces.corpus.valid,
params=pieces.params,
validation_iterator=pieces.validation_iterator)
evaluation_iterator = pieces.validation_iterator or pieces.iterator
evaluation_dataset = pieces.corpus.test
batch_weight_key = pieces.batch_weight_key
else:
trainer = TrainerBase.from_params(params, serialization_dir, recover)
# TODO(joelgrus): handle evaluation in the general case
evaluation_iterator = evaluation_dataset = None
params.assert_empty('base train command')
if convert:
logging.info('In conversion mode.')
trainer._save_checkpoint(epoch=0)
create_model_archive(serialization_dir, params)
sys.exit(0)
try:
metrics = trainer.train()
except (KeyboardInterrupt, RuntimeError):
# if we have completed an epoch, try to create a model archive.
logging.info("Training stopped. Attempting to create "
"a model archive using the current best epoch weights.")
create_model_archive(serialization_dir, params)
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=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.")
# 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_model(params: Params,
serialization_dir: str,
file_friendly_logging: bool = False,
recover: bool = False,
force: bool = False) -> 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.
Returns
-------
best_model: ``Model``
The model with the best epoch weights.
"""
prepare_environment(params)
create_serialization_dir(params, serialization_dir, recover, force)
prepare_global_logging(serialization_dir, file_friendly_logging)
cuda_device = params.params.get('trainer').get('cuda_device', -1)
if isinstance(cuda_device, list):
for device in cuda_device:
check_for_gpu(device)
else:
check_for_gpu(cuda_device)
params.to_file(os.path.join(serialization_dir, CONFIG_NAME))
all_datasets = datasets_from_params(params)
datasets_for_vocab_creation = set(params.pop("datasets_for_vocab_creation", all_datasets))
for dataset in datasets_for_vocab_creation:
if dataset not in all_datasets:
raise ConfigurationError(f"invalid 'dataset_for_vocab_creation' {dataset}")
logger.info("From dataset instances, %s will be considered for vocabulary creation.",
", ".join(datasets_for_vocab_creation))
if recover and os.path.exists(os.path.join(serialization_dir, "vocabulary")):
vocab = Vocabulary.from_files(os.path.join(serialization_dir, "vocabulary"))
else:
vocab = Vocabulary.from_params(
params.pop("vocabulary", {}),
(instance for key, dataset in all_datasets.items()
for instance in dataset
if key in datasets_for_vocab_creation)
)
model = Model.from_params(vocab=vocab, params=params.pop('model'))
# Initializing the model can have side effect of expanding the vocabulary
vocab.save_to_files(os.path.join(serialization_dir, "vocabulary"))
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
