Conditional randomization tests

Installation

pip install tabpfn-extensions

`tabpfn_crt`

View source

Conditional Randomization Test (CRT) using TabPFN. This function tests whether one or more features contain predictive information about the target variable y beyond the remaining covariates. For each feature X_j, the CRT compares: Observed statistic: Mean log predictive density of y given the observed X. Null statistic: Same quantity when X_j is replaced by samples drawn from p(X_j | X_-j).

tabpfn_crt(
    X: Any,
    y: Any,
    j: int | str | Sequence[int | str],
    *,
    B: int = 200,
    alpha: float = 0.05,
    test_size: float = 0.2,
    seed: int = 0,
    device: str | None = None,
    K: int = 100,
    max_unique_cat: int = 10,
    model_version: ModelVersion = ModelVersion.V2,
    use_kv_cache: bool = True,
) -> dict[str, Any] | dict[int | str, dict[str, Any]]

Parameters

Parameter	Type	Default	Description
`X`	`Any`	Required	—
`y`	`Any`	Required	—
`j`	`int \| str \| Sequence[int \| str]`	Required	—
`B`	`int`	`200`	—
`alpha`	`float`	`0.05`	—
`test_size`	`float`	`0.2`	—
`seed`	`int`	`0`	—
`device`	`str \| None`	`None`	—
`K`	`int`	`100`	—
`max_unique_cat`	`int`	`10`	—
`model_version`	`ModelVersion`	`ModelVersion.V2`	—
`use_kv_cache`	`bool`	`True`	—

Efficiency The predictive model p(y | X) is fit ONLY ONCE and reused across all tested features. Each feature requires fitting only the conditional model p(X_j | X_-j). Supports: • single feature testing • batch feature testing Returns

Type	Description
`Single feature → result dict`
`Multiple features → dict[feature → result dict]`

`coerce_X_y_to_numpy`

View source

Convert X and y to numpy arrays while preserving feature names if X is a DataFrame.

coerce_X_y_to_numpy(
    X: Any,
    y: Any,
) -> tuple[np.ndarray, np.ndarray, Sequence[str] | None]

Parameters

Parameter	Type	Default	Description
`X`	`Any`	Required	—
`y`	`Any`	Required	—

Returns

Type	Description
`tuple[np.ndarray, np.ndarray, Sequence[str] \| None]`	—

`is_categorical`

View source

Heuristic to determine whether a variable should be treated as categorical.

is_categorical(
    arr,
    max_unique = 10,
)

Parameters

Parameter	Type	Default	Description
`arr`	`array - like`	Required	Input array.
`max_unique`	`int`	`10`	Maximum number of unique values for categorical treatment.

Returns

Type	Description
`bool`	`True` if categorical, `False` otherwise.

`logp_from_full_output`

View source

Extract log predictive density from TabPFN ‘full’ prediction output.

logp_from_full_output(
    full_out,
    y_np,
)

Parameters

Parameter	Type	Default	Description
`full_out`	`dict`	Required	Output from TabPFN predict(…, output_type=“full”).
`y_np`	`array - like`	Required	Ground-truth targets.

Returns

Type	Description
`np.ndarray`	Log predictive density for each observation.

`logp_from_proba`

View source

Compute log p(y_true | x) from class probabilities. probs: shape (n, C) y_true: shape (n,) classes: model.classes_

logp_from_proba(
    probs,
    y_true,
    classes,
)

Parameters

Parameter	Type	Default	Description
`probs`	—	Required	—
`y_true`	—	Required	—
`classes`	—	Required	—

`resolve_feature_index`

View source

Resolve feature identifier j into integer index and optional name.

resolve_feature_index(
    j: FeatureSpec,
    feature_names: Sequence[str] | None,
    n_features: int,
) -> tuple[int, str | None]

Type aliases

FeatureSpec = int | str

Parameters

Parameter	Type	Default	Description
`j`	`FeatureSpec`	Required	—
`feature_names`	`Sequence[str] \| None`	Required	—
`n_features`	`int`	Required	—

Returns

Type	Description
`tuple[int, str \| None]`	—

Reference

tabpfn

tabpfn-client

tabpfn-extensions

REST API

Conditional randomization tests

Installation

`tabpfn_crt`

`coerce_X_y_to_numpy`

`is_categorical`

`logp_from_full_output`

`logp_from_proba`

`resolve_feature_index`

​Installation

​tabpfn_crt

​coerce_X_y_to_numpy

​is_categorical

​logp_from_full_output

​logp_from_proba

​resolve_feature_index

Installation

`tabpfn_crt`

`coerce_X_y_to_numpy`

`is_categorical`

`logp_from_full_output`

`logp_from_proba`

`resolve_feature_index`