def test_from_params_in_trainer(self):
# This is more of an integration test, making sure that a bunch of pieces fit together
# correctly, but it matters most for this learning rate scheduler, so we're testing it here.
params = Params({
"num_epochs": 5,
"learning_rate_scheduler": {
"type": "slanted_triangular",
"gradual_unfreezing": True,
"discriminative_fine_tuning": True,
"decay_factor": 0.5,
},
})
# The method called in the logic below only checks the length of this list, not its
# contents, so this should be safe.
instances = [
1,
] * 40 # noqa: E231, flake doesn't like what black does with this list
optim = self._get_optimizer()
trainer = TrainerBase.from_params(
model=self.model,
optimizer=Lazy(lambda **kwargs: optim),
serialization_dir=self.TEST_DIR,
params=params,
iterator=BasicIterator(batch_size=10),
train_data=instances,
)
assert isinstance(trainer._learning_rate_scheduler, SlantedTriangular)
# This is what we wrote this test for: to be sure that num_epochs is passed correctly, and
# that num_steps_per_epoch is computed and passed correctly. This logic happens inside of
# `Trainer.from_partial_objects`.
assert trainer._learning_rate_scheduler.num_epochs == 5
assert trainer._learning_rate_scheduler.num_steps_per_epoch == 4
# And we'll do one more to make sure that we can override num_epochs in the scheduler if we
# really want to. Not sure why you would ever want to in this case; this is just testing
# the functionality.
params = Params({
"num_epochs": 5,
"learning_rate_scheduler": {
"type": "slanted_triangular",
"num_epochs": 3,
"gradual_unfreezing": True,
"discriminative_fine_tuning": True,
"decay_factor": 0.5,
},
})
trainer = TrainerBase.from_params(
model=self.model,
optimizer=Lazy(lambda **kwargs: optim),
serialization_dir=self.TEST_DIR,
params=params,
iterator=BasicIterator(batch_size=10),
train_data=instances,
)
assert trainer._learning_rate_scheduler.num_epochs == 3
def setUp(self):
super().setUp()
params = Params({
"model": {
"type": "simple_tagger",
"text_field_embedder": {
"token_embedders": {
"tokens": {
"type": "embedding",
"embedding_dim": 5
}
}
},
"encoder": {
"type": "lstm",
"input_size": 5,
"hidden_size": 7,
"num_layers": 2
},
},
"dataset_reader": {
"type": "sequence_tagging"
},
"train_data_path":
str(self.FIXTURES_ROOT / "data" / "sequence_tagging.tsv"),
"validation_data_path":
str(self.FIXTURES_ROOT / "data" / "sequence_tagging.tsv"),
"iterator": {
"type": "basic",
"batch_size": 2
},
"trainer": {
"cuda_device": -1,
"num_epochs": 2,
"optimizer": "adam"
},
})
all_datasets = datasets_from_params(params)
vocab = Vocabulary.from_params(
params.pop("vocabulary", {}),
instances=(instance for dataset in all_datasets.values()
for instance in dataset),
)
model = Model.from_params(vocab=vocab, params=params.pop("model"))
iterator = DataIterator.from_params(params.pop("iterator"))
iterator.index_with(vocab)
train_data = all_datasets["train"]
trainer_params = params.pop("trainer")
serialization_dir = os.path.join(self.TEST_DIR,
"test_search_learning_rate")
self.trainer = TrainerBase.from_params(
model=model,
serialization_dir=serialization_dir,
iterator=iterator,
train_data=train_data,
params=trainer_params,
validation_data=None,
validation_iterator=None,
)
def setUp(self):
super().setUp()
param_file = self.FIXTURES_ROOT / "simple_tagger" / "experiment_with_regularization.json"
self.set_up_model(param_file,
self.FIXTURES_ROOT / "data" / "sequence_tagging.tsv")
params = Params.from_file(param_file)
self.reader = DatasetReader.from_params(params["dataset_reader"])
self.iterator = DataIterator.from_params(params["iterator"])
self.trainer = TrainerBase.from_params(
model=self.model,
serialization_dir=self.TEST_DIR,
iterator=self.iterator,
train_data=self.dataset,
params=params.get("trainer"),
)
def find_learning_rate_model(
params: Params,
serialization_dir: str,
start_lr: float = 1e-5,
end_lr: float = 10,
num_batches: int = 100,
linear_steps: bool = False,
stopping_factor: float = None,
force: bool = False,
) -> None:
"""
Runs learning rate search for given `num_batches` 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.
start_lr : ``float``
Learning rate to start the search.
end_lr : ``float``
Learning rate upto which search is done.
num_batches : ``int``
Number of mini-batches to run Learning rate finder.
linear_steps : ``bool``
Increase learning rate linearly if False exponentially.
stopping_factor : ``float``
Stop the search when the current loss exceeds the best loss recorded by
multiple of stopping factor. If ``None`` search proceeds till the ``end_lr``
force : ``bool``
If True and the serialization directory already exists, everything in it will
be removed prior to finding the learning rate.
"""
create_serialization_dir(params,
serialization_dir,
recover=False,
force=force)
prepare_environment(params)
cuda_device = params.params.get("trainer").get("cuda_device", -1)
check_for_gpu(cuda_device)
distributed_params = params.params.get("distributed", None)
# See https://github.com/allenai/allennlp/issues/3658
assert not distributed_params, "find-lr is not compatible with DistributedDataParallel."
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", {}),
instances=(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"))
iterator = DataIterator.from_params(params.pop("iterator"))
iterator.index_with(vocab)
train_data = all_datasets["train"]
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)
trainer_choice = trainer_params.pop("type", "default")
if trainer_choice != "default":
raise ConfigurationError(
"currently find-learning-rate only works with the default Trainer")
trainer: Trainer = TrainerBase.from_params( # type: ignore
model=model,
serialization_dir=serialization_dir,
iterator=iterator,
train_data=train_data,
validation_data=None,
params=trainer_params,
validation_iterator=None,
)
logger.info(
f"Starting learning rate search from {start_lr} to {end_lr} in {num_batches} iterations."
)
learning_rates, losses = search_learning_rate(
trainer,
start_lr=start_lr,
end_lr=end_lr,
num_batches=num_batches,
linear_steps=linear_steps,
stopping_factor=stopping_factor,
)
logger.info(f"Finished learning rate search.")
losses = _smooth(losses, 0.98)
_save_plot(learning_rates, losses,
os.path.join(serialization_dir, "lr-losses.png"))
