def default_callbacks(
self,
validation_metric: str = "-loss",
patience: int = None,
max_checkpoints: int = 20,
checkpoint_every: int = None,
model_save_interval: float = None,
serialization_dir: str = "__DEFAULT__",
validation_data: Iterable[Instance] = None,
validation_iterator: DataIterator = None,
batch_size: int = 2,
):
if serialization_dir == "__DEFAULT__":
serialization_dir = self.TEST_DIR
checkpointer = Checkpointer(serialization_dir, checkpoint_every,
max_checkpoints)
tensorboard = TensorboardWriter(get_batch_num_total=lambda: None)
if validation_iterator is None:
validation_iterator = BasicIterator(batch_size=batch_size)
validation_iterator.index_with(self.vocab)
return [
LogToTensorboard(log_batch_size_period=10,
tensorboard=tensorboard),
Checkpoint(checkpointer, model_save_interval),
Validate(
validation_data=self.instances
if validation_data is None else validation_data,
validation_iterator=validation_iterator,
),
TrackMetrics(patience, validation_metric),
GradientNormAndClip(),
]
def from_partial_objects(
cls,
serialization_dir: str,
train_dataset_readers: Dict[str, DatasetReader],
train_file_paths: Dict[str, str],
model: Lazy[Model],
iterator: DataIterator,
mingler: DatasetMingler,
optimizer: Lazy[Optimizer],
num_epochs: int = 10,
) -> "MultiTaskTrainer":
datasets = {
name: reader.read(train_file_paths[name])
for name, reader in train_dataset_readers.items()
}
instances = (instance for dataset in datasets.values()
for instance in dataset)
vocab = Vocabulary.from_instances(instances=instances)
model = model.construct(vocab=vocab)
iterator.index_with(vocab)
parameters = [[n, p] for n, p in model.named_parameters()
if p.requires_grad]
optimizer_ = optimizer.construct(model_parameters=parameters)
return MultiTaskTrainer(model, serialization_dir, iterator, mingler,
optimizer_, datasets, num_epochs)
def default_callbacks(self,
validation_metric: str = "-loss",
patience: int = None,
max_checkpoints: int = 20,
checkpoint_every: int = None,
serialization_dir: str = "__DEFAULT__",
iterator: DataIterator = None,
validation_data: Iterable[Instance] = None,
validation_iterator: DataIterator = None,
batch_size: int = 2):
if serialization_dir == "__DEFAULT__":
serialization_dir = self.TEST_DIR
checkpointer = Checkpointer(serialization_dir,
checkpoint_every,
max_checkpoints)
tensorboard = TensorboardWriter(get_batch_num_total=lambda: None)
if iterator is None:
iterator = BasicIterator(batch_size=batch_size)
iterator.index_with(self.vocab)
return [
LogToTensorboard(log_batch_size_period=10, tensorboard=tensorboard),
Checkpoint(checkpointer),
Validate(validation_data=self.instances if validation_data is None else validation_data,
validation_iterator=iterator if validation_iterator is None else validation_iterator),
TrackMetrics(patience, validation_metric),
TrainSupervised(),
GenerateTrainingBatches(self.instances, iterator, True)
]
