Trainers

TabularTrainer

TabularTrainer(
    model: TabularModel, dp_budget: DpBudget | None = None
)

Trainer for a TabularModel.

Parameters:

Name	Type	Description	Default
model	TabularModel	A TabularModel instance.	required
dp_budget	DpBudget | None	The (eps, delta)-budget for differentially private (DP) training. If None (the default), the training will not be differentially private. Available only for single table datasets.	None

train

train(
    dataset: TabularDataset,
    batch_size: int | AutoBatch,
    n_epochs: int | None = None,
    n_steps: int | None = None,
    lr: float = 0.0,
    valid: Validation | None = None,
    hooks: Sequence[TrainHook] = (),
    accumulate_grad: int = 1,
    dp_step: DpStep | None = None,
    world_size: int = 0,
) -> None

Train the TabularModel with the input TabularDataset.

Parameters:

Name	Type	Description	Default
dataset	TabularDataset	The training data, as a TabularDataset object.	required
batch_size	int | AutoBatch	The size of a batch of data during training. It is possible to specify an instance of AutoBatch to automatically estimate the optimal batch size.	required
n_epochs	int | None	The number of training epochs. One and only one of n_epochs and n_steps must be provided.	None
n_steps	int | None	The number of training steps. One and only one of n_epochs and n_steps must be provided.	None
lr	float	The learning rate. If it is 0 the optimal value for the learning rate is automatically determined.	0.0
valid	Validation | None	A Validation object. If None, no validation is performed.	None
hooks	Sequence[TrainHook]	A sequence of custom TrainHook objects.	()
accumulate_grad	int	The number of gradient accumulation steps. If equal to 1, the weights are updated at each step.	1
dp_step	DpStep | None	Data for differentially private step. Must be provided if and only if the trainer has a DP-budget.	None
world_size	int	Number of GPUs where to distribute the training. If 0, the training is performed on a single device, on the current device of the TabularTrainer object.	0

save

save(path: Path | str) -> None

Save the TabularTrainer to a checkpoint at the given path.

Parameters:

Name	Type	Description	Default
path	Path | str	The path where to save the checkpoint.	required

load classmethod

load(path: Path | str) -> TabularTrainer

Load the TabularTrainer from the checkpoint at the given path.

Parameters:

Name	Type	Description	Default
path	Path | str	The path to the loaded checkpoint.	required

Returns:

Type	Description
TabularTrainer	The loaded TabularTrainer.

TextTrainer

TextTrainer(model: TextModel)

Trainer for a TextModel.

Parameters:

Name	Type	Description	Default
model	TextModel	A TextModel instance.	required

train

train(
    dataset: TextDataset,
    batch_size: int | AutoBatch,
    n_epochs: int | None = None,
    n_steps: int | None = None,
    lr: float = 0.0,
    valid: Validation | None = None,
    hooks: Sequence[TrainHook] = (),
    accumulate_grad: int = 1,
    world_size: int = 0,
) -> None

Train the TextModel with the input TextDataset.

Parameters:

Name	Type	Description	Default
dataset	TextDataset	The training data, as a TextDataset object.	required
batch_size	int | AutoBatch	The size of a batch of data during training. It is possible to specify an instance of AutoBatch to automatically estimate the optimal batch size.	required
n_epochs	int | None	The number of training epochs. One and only one of n_epochs and n_steps must be provided.	None
n_steps	int | None	The number of training steps. One and only one of n_epochs and n_steps must be provided.	None
lr	float	The learning rate. If it is 0 the optimal value for the learning rate is automatically determined.	0.0
valid	Validation | None	A Validation object. If None, no validation is performed.	None
hooks	Sequence[TrainHook]	A sequence of custom TrainHook objects.	()
accumulate_grad	int	The number of gradient accumulation steps. If equal to 1, the weights are updated at each step.	1
world_size	int	Number of GPUs where to distribute the training. If 0, the training is performed on a single device, on the current device of the TextTrainer object.	0

save

save(path: Path | str) -> None

Save the TextTrainer to a checkpoint at the given path.

Parameters:

Name	Type	Description	Default
path	Path | str	The path where to save the checkpoint.	required

load classmethod

load(path: Path | str) -> TextTrainer

Load the TextTrainer from the checkpoint at the given path.

Parameters:

Name	Type	Description	Default
path	Path | str	The path to the loaded checkpoint.	required

Returns:

Type	Description
TextTrainer	The loaded TextTrainer.

AutoBatch pydantic-config

AutoBatch(
    memory: PositiveInt,
    min_batch_size: PositiveInt = 1,
    max_batch_size: PositiveInt = 256,
    increment_factor: Annotated[float, Field(gt=1)] = sqrt(
        2
    ),
    min_steps_epoch: Annotated[int, Field(ge=1)] = 50,
)

Configuration for automatic batch size estimation, given the available (CPU or GPU) memory. The optimal batch size is obtained by estimating the memory required for a training step, for increasing values of the batch size. The selected batch size is the largest that fits within the provided memory constraints.

Parameters:

Name	Type	Description	Default
memory	PositiveInt	The available memory in MB.	required
min_batch_size	PositiveInt	The minimum batch size.	1
max_batch_size	PositiveInt	The maximum batch size.	256
increment_factor	Annotated[float, Field(gt=1)]	The multiplicative factor used to compute the increasing batch sizes.	sqrt(2)
min_steps_epoch	Annotated[int, Field(ge=1)]	The minimum number of steps per epoch to ensure when computing the maximum batch size tested.	50