Batch and evaluation types

Looking for usage documentation? Check out Fine tuning.

`ClassifierBatch`

Batch data for classifier fine-tuning.

ClassifierBatch(
    X_context: list[torch.Tensor],
    X_query: list[torch.Tensor],
    y_context: list[torch.Tensor],
    y_query: torch.Tensor,
    cat_indices: list[list[int] | None] | list[list[list[int] | None]],
    configs: list[Any],
)

Fields

Name	Type	Default	Description
`X_context`	`list[torch.Tensor]`	Required	Preprocessed training features (list per estimator).
`X_query`	`list[torch.Tensor]`	Required	Preprocessed test features (list per estimator).
`y_context`	`list[torch.Tensor]`	Required	Preprocessed training targets (list per estimator).
`y_query`	`torch.Tensor`	Required	Raw test target tensor.
`cat_indices`	`list[list[int] \| None] \| list[list[list[int] \| None]]`	Required	Categorical feature indices (list per estimator).
`configs`	`list[Any]`	Required	Preprocessing configurations used for this batch.

`EvalResult`

View source

Container for evaluation results.

EvalResult(
    primary: float,
    secondary: dict[str, float] = dict(),
)

Fields

Name	Type	Default	Description
`primary`	`float`	Required	The primary metric used for early stopping decisions.
`secondary`	`dict[str, float]`	`dict()`	Additional metrics for logging purposes only.

`RegressorBatch`

View source

Batch data for regressor fine-tuning.

RegressorBatch(
    X_context: list[torch.Tensor],
    X_query: list[torch.Tensor],
    y_context: list[torch.Tensor],
    y_query: torch.Tensor,
    cat_indices: list[list[int] | None] | list[list[list[int] | None]],
    configs: list[Any],
    raw_space_bardist: FullSupportBarDistribution,
    znorm_space_bardist: FullSupportBarDistribution,
    X_query_raw: torch.Tensor,
    y_query_raw: torch.Tensor,
)

Fields

Name	Type	Default	Description
`X_context`	`list[torch.Tensor]`	Required	Preprocessed training features (list per estimator).
`X_query`	`list[torch.Tensor]`	Required	Preprocessed test features (list per estimator).
`y_context`	`list[torch.Tensor]`	Required	Preprocessed standardized training targets (list per estimator).
`y_query`	`torch.Tensor`	Required	Standardized test target tensor.
`cat_indices`	`list[list[int] \| None] \| list[list[list[int] \| None]]`	Required	Categorical feature indices (list per estimator).
`configs`	`list[Any]`	Required	Preprocessing configurations used for this batch.
`raw_space_bardist`	`FullSupportBarDistribution`	Required	Bar distribution in raw (original) target space.
`znorm_space_bardist`	`FullSupportBarDistribution`	Required	Bar distribution in z-normalized target space.
`X_query_raw`	`torch.Tensor`	Required	Original unprocessed test features.
`y_query_raw`	`torch.Tensor`	Required	Original unprocessed test targets.

`main_process_first`

View source

Run the with-block on the main process before all other ranks. Useful under torchrun for work that should happen once and be read from a shared cache afterwards, such as dataset downloads: the main process runs the block while the other ranks wait at a barrier, then the other ranks run it against the warm cache. Initializes the process group from the torchrun env vars if needed, and leaves it initialized so that a subsequent fit() reuses it. Call torch.distributed.destroy_process_group() at the end of your script. No-op when running with a single process.

main_process_first() -> Iterator[None]

Returns

Type	Description
`Iterator[None]`	—

Reference

tabpfn

tabpfn-client

tabpfn-extensions

REST API

Batch and evaluation types

`ClassifierBatch`

`EvalResult`

`RegressorBatch`

`main_process_first`

​ClassifierBatch

​EvalResult

​RegressorBatch

​main_process_first

`ClassifierBatch`

`EvalResult`

`RegressorBatch`

`main_process_first`