API Reference

save

def save(
    estimator: BaseEstimator,
    arrays_path: str | Path,
    state_path: str | Path,
    format: str | None = None,
) -> None: ...

Serialize a fitted estimator to a pair of files:

• arrays_path: safetensors file containing numpy arrays (weights, coefficients, etc.)
• state_path: JSON file containing hyperparameters and scalar fitted state.
• format: optional, "native" to use the library's own format for LightGBM/XGBoost models. Default uses columnar tensors.

load

def load(
    arrays_path: str | Path,
    state_path: str | Path,
) -> BaseEstimator: ...

Load a fitted estimator from files previously created by save. Reconstructs the full model hierarchy (pipelines, nested estimators) from the JSON config, restores all fitted arrays, and auto-detects the serialization format (e.g. columnar vs native for tree models). The returned estimator is ready for inference or further training.

serialize

def serialize(
    estimator: BaseEstimator,
    format: str | None = None,
) -> tuple[dict[str, Any], dict[str, np.ndarray]]: ...

Like save, but returns (state_dict, arrays_dict) in memory instead of writing to files.

deserialize

def deserialize(
    state: dict[str, Any],
    arrays: dict[str, np.ndarray],
) -> BaseEstimator: ...

Reconstruct a fitted estimator from the dicts returned by serialize.

get_model_params

def get_model_params(
    estimator: BaseEstimator,
) -> dict[str, Any]: ...

Recursively extract hyperparameters from an estimator. For composite models like Pipeline or TransformedTargetRegressor, it traverses the model hierarchy and returns a nested dict where each sub-model's parameters are grouped under its step name.

For example, given Pipeline([("scaler", StandardScaler()), ("model", Ridge())]), the keys "scaler" and "model" match the step names from the pipeline:

{"steps": {
    "scaler": {"with_mean": True, "type": "sklearn.preprocessing.StandardScaler", ...},
    "model":  {"alpha": 0.1, "type": "sklearn.linear_model.Ridge", ...}
}, ...}

set_model_params

def set_model_params(
    estimator: BaseEstimator,
    params: dict[str, Any],
) -> BaseEstimator: ...

Set hyperparameters on an estimator using the same nested structure returned by get_model_params. For composite models, it traverses the hierarchy and applies parameters to each sub-model. You only need to include the parameters you want to change.

skeights.set_model_params(pipe, {"steps": {"model": {"alpha": 0.5}}})

get_sklearn_public_path

def get_sklearn_public_path(
    cls: type,
) -> str: ...

Return the stable public import path for a scikit-learn class. Mostly useful internally when developing handlers, but exposed for external use. sklearn places classes in private submodules (e.g. sklearn.preprocessing._data.StandardScaler) but re-exports them from public packages (sklearn.preprocessing.StandardScaler). This function resolves the shortest public path, falling back to the full private path if no public re-export exists.

from sklearn.preprocessing import StandardScaler
skeights.get_sklearn_public_path(StandardScaler)
# "sklearn.preprocessing.StandardScaler"