sklearn bridge

concrete_fhe_toolkit.ml.sklearn_bridge

Convert trained scikit-learn models into toolkit FHE models.

These helpers kill the manual scaling ritual: train with sklearn on clear data as usual, convert once, and run encrypted inference with the toolkit. scikit-learn is imported lazily, so it stays an optional dependency.

For linear models, scale quantizes coefficients. Inputs stay in the training coordinate system by default. To encode real inputs as integers, pass a separate input_scale and feed round(value * input_scale). Tree conversion uses its scale parameter for input thresholds.

from_sklearn_linear

from_sklearn_linear(model: Any, *, scale: int = 100, input_scale: int = 1) -> Any

[Client-Side Helper] Convert a fitted sklearn linear model into an FHE model.

LogisticRegression (binary) becomes an :class:FHELogisticRegression; LinearRegression (and other regressors exposing coef_/intercept_) become an :class:FHELinearRegression whose predictions are scaled by scale * input_scale. With the default input_scale=1, use integer features in exactly the coordinate system used during training. If you trained on already quantized data, do not quantize those inputs again.

Parameters:

Name	Type	Description	Default
model	Any	A fitted sklearn estimator with coef_ and intercept_.	required
scale	int	The scaling factor applied to the fixed-point representation.	100
input_scale	int	The fixed-point scaling factor of the input. inference. Bias and output scale account for both factors.	1

Returns:

Type	Description
Any	A ready-to-compile toolkit model.

Example

from sklearn.linear_model import LogisticRegression
from concrete_fhe_toolkit.ml.sklearn_bridge import from_sklearn_linear

clf = LogisticRegression().fit(X_scaled, y)
fhe_model = from_sklearn_linear(clf, scale=100, input_scale=10)
# quantized_inputset/sample contain round(training_space_value * 10)
fhe_model.compile(quantized_inputset)
fhe_model.predict(quantized_sample)

Source code in src/concrete_fhe_toolkit/ml/sklearn_bridge.py

@client_side_helper
def from_sklearn_linear(
    model: Any, *, scale: int = 100, input_scale: int = 1
) -> Any:
    """[Client-Side Helper] Convert a fitted sklearn linear model into an FHE model.

    ``LogisticRegression`` (binary) becomes an
    :class:`FHELogisticRegression`; ``LinearRegression`` (and other
    regressors exposing ``coef_``/``intercept_``) become an
    :class:`FHELinearRegression` whose predictions are scaled by
    ``scale * input_scale``. With the default ``input_scale=1``, use integer
    features in exactly the coordinate system used during training. If you
    trained on already quantized data, do not quantize those inputs again.

    Args:
        model: A fitted sklearn estimator with ``coef_`` and ``intercept_``.
        scale: The scaling factor applied to the fixed-point representation.
        input_scale: The fixed-point scaling factor of the input.
            inference. Bias and output scale account for both factors.

    Returns:
        A ready-to-compile toolkit model.

    Example:
        ```python
        from sklearn.linear_model import LogisticRegression
        from concrete_fhe_toolkit.ml.sklearn_bridge import from_sklearn_linear

        clf = LogisticRegression().fit(X_scaled, y)
        fhe_model = from_sklearn_linear(clf, scale=100, input_scale=10)
        # quantized_inputset/sample contain round(training_space_value * 10)
        fhe_model.compile(quantized_inputset)
        fhe_model.predict(quantized_sample)
        ```
    """
    normalized_scale = validate_integer("scale", scale, minimum=1)
    source_scale = validate_integer("input_scale", input_scale, minimum=1)
    if not hasattr(model, "coef_") or not hasattr(model, "intercept_"):
        raise ValueError("model must be a fitted sklearn linear estimator")

    import numpy as np

    coefficients = np.atleast_2d(np.asarray(model.coef_))
    intercepts = np.atleast_1d(np.asarray(model.intercept_))
    if coefficients.shape[0] != 1:
        raise ValueError(
            "only binary/single-output linear models are supported; "
            "use one converted model per class for multi-class"
        )

    weights = [int(round(value * normalized_scale)) for value in coefficients[0]]
    bias = int(round(float(intercepts[0]) * normalized_scale * source_scale))

    is_classifier = hasattr(model, "classes_")
    if is_classifier:
        return FHELogisticRegression(
            weights, bias, 
            input_scale=source_scale, 
            output_scale=normalized_scale * source_scale
        )
    return FHELinearRegression(
        weights, bias, 
        input_scale=source_scale, 
        output_scale=normalized_scale * source_scale
    )

from_sklearn_tree

from_sklearn_tree(model: Any, *, scale: int = 1, leaf_scale: int = 1) -> Any

[Client-Side Helper] Convert a fitted sklearn decision tree into an :class:FHEDecisionTree.

Classification trees keep their majority-class integer leaves; regression trees get leaves quantized by leaf_scale. Thresholds are quantized by scale — quantize inference features with the same factor.

Parameters:

Name	Type	Description	Default
model	Any	A fitted sklearn decision tree model.	required
scale	int	The scaling factor applied to the fixed-point representation.	1
leaf_scale	int	Factor for quantizing leaf values. Defaults to 1.	1

Returns:

Name	Type	Description
Any	Any	A ready-to-compile FHEDecisionTree model.

Raises:

Type	Description
ValueError	If model is not a fitted sklearn decision tree.

Example

from sklearn.tree import DecisionTreeClassifier
from concrete_fhe_toolkit.ml.sklearn_bridge import from_sklearn_tree

clf = DecisionTreeClassifier(max_depth=3).fit(X_int, y)
fhe_tree = from_sklearn_tree(clf)
fhe_tree.compile(inputset)

Source code in src/concrete_fhe_toolkit/ml/sklearn_bridge.py

@client_side_helper
def from_sklearn_tree(model: Any, *, scale: int = 1, leaf_scale: int = 1) -> Any:
    """[Client-Side Helper] Convert a fitted sklearn decision tree into an :class:`FHEDecisionTree`.

    Classification trees keep their majority-class integer leaves;
    regression trees get leaves quantized by ``leaf_scale``. Thresholds are
    quantized by ``scale`` — quantize inference features with the same
    factor.

    Args:
        model (Any): A fitted sklearn decision tree model.
        scale (int, optional): The scaling factor applied to the fixed-point representation.
        leaf_scale (int, optional): Factor for quantizing leaf values. Defaults to 1.

    Returns:
        Any: A ready-to-compile FHEDecisionTree model.

    Raises:
        ValueError: If model is not a fitted sklearn decision tree.

    Example:
        ```python
        from sklearn.tree import DecisionTreeClassifier
        from concrete_fhe_toolkit.ml.sklearn_bridge import from_sklearn_tree

        clf = DecisionTreeClassifier(max_depth=3).fit(X_int, y)
        fhe_tree = from_sklearn_tree(clf)
        fhe_tree.compile(inputset)
        ```
    """
    normalized_scale = validate_integer("scale", scale, minimum=1)
    normalized_leaf_scale = validate_integer("leaf_scale", leaf_scale, minimum=1)
    if not hasattr(model, "tree_"):
        raise ValueError("model must be a fitted sklearn decision tree")
    tree_dict = _convert_tree_node(
        model.tree_, 0, scale=normalized_scale, leaf_scale=normalized_leaf_scale
    )
    return FHEDecisionTree(tree_dict)

from_sklearn_forest

from_sklearn_forest(model: Any, *, scale: int = 1) -> Any

[Client-Side Helper] Convert a fitted sklearn random-forest classifier into an :class:FHERandomForest (binary labels, majority vote).

Parameters:

Name	Type	Description	Default
model	Any	A fitted sklearn random forest model.	required
scale	int	The scaling factor applied to the fixed-point representation.	1

Returns:

Name	Type	Description
Any	Any	A ready-to-compile FHERandomForest model.

Raises:

Type	Description
ValueError	If model is not a fitted sklearn forest.

Example

from sklearn.ensemble import RandomForestClassifier
from concrete_fhe_toolkit.ml.sklearn_bridge import from_sklearn_forest

clf = RandomForestClassifier(n_estimators=5, max_depth=3).fit(X_int, y)
fhe_forest = from_sklearn_forest(clf)

Source code in src/concrete_fhe_toolkit/ml/sklearn_bridge.py

@client_side_helper
def from_sklearn_forest(model: Any, *, scale: int = 1) -> Any:
    """[Client-Side Helper] Convert a fitted sklearn random-forest classifier into an
    :class:`FHERandomForest` (binary labels, majority vote).

    Args:
        model (Any): A fitted sklearn random forest model.
        scale (int, optional): The scaling factor applied to the fixed-point representation.

    Returns:
        Any: A ready-to-compile FHERandomForest model.

    Raises:
        ValueError: If model is not a fitted sklearn forest.

    Example:
        ```python
        from sklearn.ensemble import RandomForestClassifier
        from concrete_fhe_toolkit.ml.sklearn_bridge import from_sklearn_forest

        clf = RandomForestClassifier(n_estimators=5, max_depth=3).fit(X_int, y)
        fhe_forest = from_sklearn_forest(clf)
        ```
    """
    if not hasattr(model, "estimators_"):
        raise ValueError("model must be a fitted sklearn forest")
    trees: List[Any] = [
        from_sklearn_tree(estimator, scale=scale).tree
        for estimator in model.estimators_
    ]
    return FHERandomForest(trees)