Self-hosted HTTP

Looking for usage documentation? Check out Classification, Regression and KV cache.

`hosted.TabPFNClassifier`

TabPFN classifier backed by a self-hosted inference endpoint.

hosted.TabPFNClassifier(
    endpoint_url: str,
    api_key: Optional[str] = None,
    extra_headers: Optional[Dict[str, str]] = None,
    model_id: Optional[str] = None,
    model_path: Optional[str] = None,
    model_version: Optional[str] = None,
    fit_mode: Optional[Literal["fit_preprocessors", "low_memory", "fit_with_cache", "batched"]] = None,
    n_estimators: int | None = None,
    softmax_temperature: float | None = None,
    balance_probabilities: bool = False,
    average_before_softmax: bool | None = None,
    inference_precision: Literal["autocast", "auto"] | None = None,
    random_state: int | None = 0,
    inference_config: Dict[str, Any] | None = None,
    n_preprocessing_jobs: int = 4,
    memory_saving_mode: bool | Literal["auto"] | None = None,
    categorical_features_indices: List[int] | None = None,
    timeout_s: float = 300.0,
    payload_format: Literal["json", "parquet"] = "json",
    use_kv_cache: bool = False,
)

Parameters

Parameter	Type	Default	Description
`endpoint_url`	`str`	Required	—
`api_key`	`Optional[str]`	`None`	—
`extra_headers`	`Optional[Dict[str, str]]`	`None`	—
`model_id`	`Optional[str]`	`None`	—
`model_path`	`Optional[str]`	`None`	—
`model_version`	`Optional[str]`	`None`	—
`fit_mode`	`Optional[Literal[“fit_preprocessors”, “low_memory”, “fit_with_cache”, “batched”]]`	`None`	—
`n_estimators`	`int \| None`	`None`	—
`softmax_temperature`	`float \| None`	`None`	—
`balance_probabilities`	`bool`	`False`	—
`average_before_softmax`	`bool \| None`	`None`	—
`inference_precision`	`Literal[“autocast”, “auto”] \| None`	`None`	—
`random_state`	`int \| None`	`0`	—
`inference_config`	`Dict[str, Any] \| None` `InferenceConfig` options	`None`	—
`n_preprocessing_jobs`	`int`	`4`	—
`memory_saving_mode`	`bool \| Literal[“auto”] \| None`	`None`	—
`categorical_features_indices`	`List[int] \| None`	`None`	—
`timeout_s`	`float`	`300.0`	—
`payload_format`	`Literal[“json”, “parquet”]`	`"json"`	—
`use_kv_cache`	`bool`	`False`	—

Example

from tabpfn_client.hosted import TabPFNClassifier
clf = TabPFNClassifier(
    endpoint_url="https://<your-endpoint>/predict",
    api_key="<optional-bearer-token>",
)
clf.fit(X_train, y_train)
clf.predict(X_test)
clf.predict_proba(X_test)

`hosted.TabPFNClassifier.fit`

View source

Store the training data; the endpoint fits on the next predict* call. Clears any model_id_ captured from an earlier fit.

hosted.TabPFNClassifier.fit(
    X: Any,
    y: Any,
) -> "_HostedBase"

Parameters

Parameter	Type	Default	Description
`X`	`Any`	Required	Training features.
`y`	`Any`	Required	Training targets.

Returns

Type	Description
`_HostedBase`	The estimator itself.

Raises ValueError If X and y have different numbers of samples.

`hosted.TabPFNClassifier.predict`

View source

Predict class labels for X.

hosted.TabPFNClassifier.predict(
    X: Any,
) -> np.ndarray

Parameters

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

Returns

Type	Description
`np.ndarray`	—

`hosted.TabPFNClassifier.predict_proba`

View source

Predict class probabilities for X.

hosted.TabPFNClassifier.predict_proba(
    X: Any,
) -> np.ndarray

Parameters

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

Returns

Type	Description
`np.ndarray`	—

`hosted.TabPFNRegressor`

View source

TabPFN regressor backed by a self-hosted inference endpoint.

hosted.TabPFNRegressor(
    endpoint_url: str,
    api_key: Optional[str] = None,
    extra_headers: Optional[Dict[str, str]] = None,
    model_id: Optional[str] = None,
    model_path: Optional[str] = None,
    model_version: Optional[str] = None,
    fit_mode: Optional[Literal["fit_preprocessors", "low_memory", "fit_with_cache", "batched"]] = None,
    n_estimators: int | None = None,
    softmax_temperature: float | None = None,
    balance_probabilities: bool = False,
    average_before_softmax: bool | None = None,
    inference_precision: Literal["autocast", "auto"] | None = None,
    random_state: int | None = 0,
    inference_config: Dict[str, Any] | None = None,
    n_preprocessing_jobs: int = 4,
    memory_saving_mode: bool | Literal["auto"] | None = None,
    categorical_features_indices: List[int] | None = None,
    timeout_s: float = 300.0,
    payload_format: Literal["json", "parquet"] = "json",
    use_kv_cache: bool = False,
)

Parameters

Parameter	Type	Default	Description
`endpoint_url`	`str`	Required	—
`api_key`	`Optional[str]`	`None`	—
`extra_headers`	`Optional[Dict[str, str]]`	`None`	—
`model_id`	`Optional[str]`	`None`	—
`model_path`	`Optional[str]`	`None`	—
`model_version`	`Optional[str]`	`None`	—
`fit_mode`	`Optional[Literal[“fit_preprocessors”, “low_memory”, “fit_with_cache”, “batched”]]`	`None`	—
`n_estimators`	`int \| None`	`None`	—
`softmax_temperature`	`float \| None`	`None`	—
`balance_probabilities`	`bool`	`False`	—
`average_before_softmax`	`bool \| None`	`None`	—
`inference_precision`	`Literal[“autocast”, “auto”] \| None`	`None`	—
`random_state`	`int \| None`	`0`	—
`inference_config`	`Dict[str, Any] \| None` `InferenceConfig` options	`None`	—
`n_preprocessing_jobs`	`int`	`4`	—
`memory_saving_mode`	`bool \| Literal[“auto”] \| None`	`None`	—
`categorical_features_indices`	`List[int] \| None`	`None`	—
`timeout_s`	`float`	`300.0`	—
`payload_format`	`Literal[“json”, “parquet”]`	`"json"`	—
`use_kv_cache`	`bool`	`False`	—

Example

from tabpfn_client.hosted import TabPFNRegressor
reg = TabPFNRegressor(
    endpoint_url="https://<your-endpoint>/predict",
    api_key="<optional-bearer-token>",
)
reg.fit(X_train, y_train)
reg.predict(X_test)
reg.predict(X_test, output_type="quantiles", quantiles=[0.1, 0.5, 0.9])

`hosted.TabPFNRegressor.fit`

View source

Store the training data; the endpoint fits on the next predict* call. Clears any model_id_ captured from an earlier fit.

hosted.TabPFNRegressor.fit(
    X: Any,
    y: Any,
) -> "_HostedBase"

Parameters

Parameter	Type	Default	Description
`X`	`Any`	Required	Training features.
`y`	`Any`	Required	Training targets.

Returns

Type	Description
`_HostedBase`	The estimator itself.

Raises ValueError If X and y have different numbers of samples.

`hosted.TabPFNRegressor.predict`

View source

Predict regression targets for X.

hosted.TabPFNRegressor.predict(
    X: Any,
    output_type: str = "mean",
    quantiles: Optional[list] = None,
) -> np.ndarray

Parameters

Parameter	Type	Default	Description
`X`	`Any`	Required	Test features.
`output_type`	`str`	`"mean"`	Which prediction to return, e.g. “mean”, “median” or “quantiles”.
`quantiles`	`Optional[list]`	`None`	Quantile levels to compute when `output_type` is “quantiles”.

Returns

Type	Description
`np.ndarray`	The endpoint’s prediction as an array.

Scikit-learn configuration and scoring

The adapters inherit get_params, set_params, and the corresponding estimator mixin’s score method. They do not inherit the main hosted client’s save_model, load_model, or model-version selection methods.

Reference

tabpfn

tabpfn-client

tabpfn-extensions

REST API

`hosted.TabPFNClassifier`

`hosted.TabPFNClassifier.fit`

`hosted.TabPFNClassifier.predict`

`hosted.TabPFNClassifier.predict_proba`

`hosted.TabPFNRegressor`

`hosted.TabPFNRegressor.fit`

`hosted.TabPFNRegressor.predict`

Scikit-learn configuration and scoring

​hosted.TabPFNClassifier

​hosted.TabPFNClassifier.fit

​hosted.TabPFNClassifier.predict

​hosted.TabPFNClassifier.predict_proba

​hosted.TabPFNRegressor

​hosted.TabPFNRegressor.fit

​hosted.TabPFNRegressor.predict

​Scikit-learn configuration and scoring

`hosted.TabPFNClassifier`

`hosted.TabPFNClassifier.fit`

`hosted.TabPFNClassifier.predict`

`hosted.TabPFNClassifier.predict_proba`

`hosted.TabPFNRegressor`

`hosted.TabPFNRegressor.fit`

`hosted.TabPFNRegressor.predict`

Scikit-learn configuration and scoring