Options

Options

mozuma.torch.options.TorchRunnerOptions dataclass

Options for PyTorch runners

Attributes:

Name	Type	Description
device	torch.device	Torch device
data_loader_options	dict	Options passed to torch.utils.dataloader.DataLoader.
tqdm_enabled	bool	Whether to print a tqdm progress bar

mozuma.torch.options.TorchMultiGPURunnerOptions dataclass

Options for PyTorch multi-gpu runners

Attributes:

Name	Type	Description
data_loader_options	dict	Options passed to torch.utils.dataloader.DataLoader.
tqdm_enabled	bool	Whether to print a tqdm progress bar. Default, False.
dist_options	dict	Options passed to ignite.distributed.Parallel.
seed	int	random state seed to set. Default, 543.

Note

data_loader_options's options batch_size and num_worker will be automatically scaled according to world_size and nprocs respectively. For more info visit auto_dataloader documentation.

Note

dist_options usually include backend and nproc_per_node parameters. For more info visit PyTorch Ignite's distributed documentation.

mozuma.torch.options.TorchTrainingOptions dataclass

Options for PyTorch training runners

Attributes:

Name	Type	Description
criterion	Union[Callable, torch.nn.Module]	the loss function to use during training.
optimizer	torch.optim.Optimizer	Optimization strategy to use during training.
num_epochs	int	number of epochs to train the model.
validate_every	int	run model's validation every validate_every epochs.
checkpoint_every	Optional[int]	store training checkpoint every checkpoint_every epochs.
metrics	Dict[str, Metric]	Dictionary where values are Ignite's metrics to compute during evaluation.
data_loader_options	dict	Options passed to torch.utils.dataloader.DataLoader.
tqdm_enabled	bool	Whether to print a tqdm progress bar. Default, False.
dist_options	dict	Options passed to ignite.distributed.Parallel.
seed	int	random state seed to set. Default, 543.
loggers_factory	Callable[[Engine, Engine, Engine], None] | None	Function to attach additional loggers to training runner and its evaluators internal (PyTorch Ignite) engines. The function receives three engines, one for training and two for evaluation, where the first is for the train set and second for the test set.

Note

data_loader_options's options batch_size and num_worker will be automatically scaled according to world_size and nprocs respectively. For more info visit auto_dataloader documentation.

Note

dist_options usually include backend and nproc_per_node parameters. For more info visit PyTorch Ignite's distributed documentation.