Batched and differentiable inference

Looking for usage documentation? Check out Fine tuning.

`TabPFNClassifier.fit_from_preprocessed`

Used in Fine-Tuning. Fit the model to preprocessed inputs from torch dataloader inside a training loop a Dataset provided by get_preprocessed_datasets. This function always uses the “batched” fit_mode.

TabPFNClassifier.fit_from_preprocessed(
    X_preprocessed: list[torch.Tensor],
    y_preprocessed: list[torch.Tensor],
    cat_ix: list[list[list[int]]],
    configs: list[list[EnsembleConfig]],
    *,
    performance_options: PerformanceOptions,
    no_refit: bool = True,
) -> TabPFNClassifier

Parameters

Parameter	Type	Default	Description
`X_preprocessed`	`list[torch.Tensor]`	Required	The input features obtained from the preprocessed Dataset The list contains one item for each ensemble predictor. use tabpfn.utils.collate_for_tabpfn_dataset to use this function with batch sizes of more than one dataset (see examples/tabpfn_finetune.py)
`y_preprocessed`	`list[torch.Tensor]`	Required	The target variable obtained from the preprocessed Dataset
`cat_ix`	`list[list[list[int]]]`	Required	categorical indices obtained from the preprocessed Dataset
`configs`	`list[list[EnsembleConfig]]`	Required	Ensemble configurations obtained from the preprocessed Dataset
`performance_options`	`PerformanceOptions`	Required	Performance and memory options forwarded to the model on each forward call inside the resulting executor.
`no_refit`	`bool`	`True`	if `True`, the classifier will not be reinitialized when calling fit multiple times.

Returns

Type	Description
`TabPFNClassifier`	—

`TabPFNClassifier.predict_proba_batched`

View source

Predict probabilities for several independent datasets in one pass. Each (X_train, y_train, X_test) triple is preprocessed exactly as in fit() + predict_proba() (input validation, CPU and GPU preprocessing, same ensemble configs), then compatible model-input shapes are fused. Heterogeneous post-preprocessing shapes run in separate groups. All datasets must share the same set of classes (they are scored together with a single n_classes_) and the same raw array shapes. Fitted transforms may produce different model-input shapes; these are grouped internally without padding. This method does not modify the estimator: the per-dataset fits run on an internal clone, so self is unchanged on return (any prior fit is preserved).

TabPFNClassifier.predict_proba_batched(
    X_train_list: list[XType],
    y_train_list: list[YType],
    X_test_list: list[XType],
) -> np.ndarray

Parameters

Parameter	Type	Default	Description
`X_train_list`	`list[XType]`	Required	Training features, one array per dataset (all same shape).
`y_train_list`	`list[YType]`	Required	Training labels, one array per dataset.
`X_test_list`	`list[XType]`	Required	Test features, one array per dataset (all same shape).

Returns

Type	Description
`np.ndarray`	Probabilities of shape `(n_datasets, n_test, n_classes)`.

Raises ValueError If the input lists have unequal or zero length, the datasets do not all share the same set of classes, or the training (or test) arrays do not all share one shape. NotImplementedError If balance_probabilities or tuning_config is configured on the estimator — their state is per-dataset and cannot be applied correctly across a shared batch. Score those datasets individually with predict_proba. Also raised for inference_precision=torch.float64, which the fused forward does not support.

`TabPFNClassifier.fit_with_differentiable_input`

View source

Fit the model with differentiable input.

TabPFNClassifier.fit_with_differentiable_input(
    X: torch.Tensor,
    y: torch.Tensor,
) -> Self

Parameters

Parameter	Type	Default	Description
`X`	`torch.Tensor`	Required	The input data.
`y`	`torch.Tensor`	Required	The target variable.

Returns

Type	Description
`Self`	self

`TabPFNClassifier.logits_to_probabilities`

View source

Convert logits to probabilities using the classifier’s post-processing.

TabPFNClassifier.logits_to_probabilities(
    raw_logits: np.ndarray | torch.Tensor,
    *,
    softmax_temperature: float | None = None,
    average_before_softmax: bool | None = None,
    balance_probabilities: bool | None = None,
) -> torch.Tensor

Parameters

Parameter	Type	Default	Description
`raw_logits`	`np.ndarray \| torch.Tensor`	Required	Logits with shape (n_estimators, n_samples, n_classes) or (n_samples, n_classes). If the logits have three dimensions, they are averaged across the estimator dimension (dim=0).
`softmax_temperature`	`float \| None`	`None`	Optional override for temperature scaling.
`average_before_softmax`	`bool \| None`	`None`	Optional override for averaging order.
`balance_probabilities`	`bool \| None`	`None`	Optional override for probability balancing.

Returns

Type	Description
`torch.Tensor`	Probabilities with shape (n_samples, n_classes).

`TabPFNClassifier.forward`

View source

Forward pass returning predicted probabilities or logits for TabPFNClassifier Inference Engine. Used in Fine-Tuning and prediction. Called directly in FineTuning training loop or by predict() function with the use_inference_mode flag explicitly set to True. Iterates over outputs of InferenceEngine.

TabPFNClassifier.forward(
    X: list[torch.Tensor] | torch.Tensor,
    *,
    use_inference_mode: bool = False,
    return_logits: bool = False,
    return_raw_logits: bool = False,
) -> torch.Tensor

Parameters

Parameter	Type	Default	Description
`X`	`list[torch.Tensor] \| torch.Tensor`	Required	list[torch.Tensor] in fine-tuning, XType in normal predictions.
`use_inference_mode`	`bool`	`False`	Flag for inference mode., default at `False` since
`return_logits`	`bool`	`False`	If `True`, returns logits averaged across estimators. Otherwise, probabilities are returned.
`return_raw_logits`	`bool`	`False`	If `True`, returns the raw logits, without averaging estimators or temperature scaling.

Returns

Type	Description
`torch.Tensor`	The predicted probabilities or logits of the classes as a torch.Tensor. - If `use_inference_mode` is `True`: Shape (N_samples, N_classes) - If `use_inference_mode` is `False` (e.g., for training/fine-tuning): Shape (Batch_size, N_classes, N_samples), suitable for NLLLoss. - If `return_raw_logits` is `True`: Shape (n_estimators, n_samples, n_classes)

Reference

tabpfn

tabpfn-client

tabpfn-extensions

REST API

Batched and differentiable inference

`TabPFNClassifier.fit_from_preprocessed`

`TabPFNClassifier.predict_proba_batched`

`TabPFNClassifier.fit_with_differentiable_input`

`TabPFNClassifier.logits_to_probabilities`

`TabPFNClassifier.forward`

​TabPFNClassifier.fit_from_preprocessed

​TabPFNClassifier.predict_proba_batched

​TabPFNClassifier.fit_with_differentiable_input

​TabPFNClassifier.logits_to_probabilities

​TabPFNClassifier.forward

`TabPFNClassifier.fit_from_preprocessed`

`TabPFNClassifier.predict_proba_batched`

`TabPFNClassifier.fit_with_differentiable_input`

`TabPFNClassifier.logits_to_probabilities`

`TabPFNClassifier.forward`