def train(hidden_features: int = 100, dropout: float = 0.05,
bias: bool = True, negative_slope: float = 0.05,
seed: int = 42, device: str = 'cpu', batch_size: int = 5, num_epochs: int = 50,
out_dir: Path = Path('../out_dir'), monitor: ('filesystem', 'tensorboard') = 'tensorboard'):
""" train iris classifier
Args:
hidden_features: the size of hidden layers
dropout: the dropout ratio
bias: whether or not use the bias in hidden layers
negative_slope: the ratio of negative part
seed: the random seed number
device: device id
batch_size: the size of each batch
num_epochs: the total numbers of epochs
out_dir: the root path of output
monitor: the type of monitor
"""
options = locals()
experiment_dir = out_dir / experiment_hash(**options)
ensure_output_dir(experiment_dir)
options_dump(experiment_dir, **options)
log_system.notice(f'experiment_dir => {experiment_dir}')
manual_seed(seed)
log_system.notice(f'seed => {seed}')
train, test = prepare_iris_dataset(batch_size)
estimator = IrisEstimator(
in_features=4, dropout=dropout, num_classes=3, hidden_features=hidden_features,
negative_slope=negative_slope, bias=bias
)
optimizer = optim.Adam(estimator.parameters())
monitor = get_monitor(monitor)(log_dir=experiment_dir)
to_device(device, estimator)
schedule = EpochalSchedule(estimator, optimizer, monitor)
schedule.register_extension(Periodic(Moment.AFTER_ITERATION, iteration=5))(CommitScalarByMean(
'criterion', 'acc', chapter='train',
))
schedule.register_extension(Periodic(Moment.AFTER_BACKWARD, iteration=1))(ClipGradNorm(max_norm=4.))
schedule.register_extension(Periodic(Moment.AFTER_EPOCH, epoch=1))(Pipeline(
Evaluation(data_loader=test, chapter='test'),
CommitScalarByMean('criterion', 'acc', chapter='test'),
))
return schedule.run(train, num_epochs)
def glove_(tensor: Tensor, file: Path, vocab: Vocab, encoding: str = 'utf-8') -> None:
with torch.no_grad():
for count, (token, vectors) in enumerate(
token_vectors_stream(file, vocab, encoding), start=1):
tensor[vocab(token)] = f32_tensor(vectors)
return log_system.notice(f'loaded {count} tokens from {file}')
def setattr_with_notice(self, name: str, value) -> None:
setattr(self, name, value)
return log_system.notice(
f'{self.__class__.__name__}.{name} := {value.__class__.__name__}')
def __call__(self, schedule: 'Schedule') -> None:
value = getattr(schedule, self.key)
tm = datetime.now() - getattr(schedule, self.attr)
log_system.notice(
f'[{self.key} {value}] finished, time elapsed => {tm}')
return delattr(schedule, self.attr)
def __call__(self, schedule: 'Schedule') -> None:
value = getattr(schedule, self.key)
tm = datetime.now()
log_system.notice(f'[{self.key} {value}] start => {tm}')
return setattr(schedule, self.attr, tm)