Xgboost

Get an XGBoost component for your pipeline.

A Component wrapped around XGBoost with two possible default spaces "large" and "performant" or you own custom space=.

See amltk.pipeline.xgboost.xgboost_component

def xgboost_component(kind, name=None, space='large', config=None)

Create a component with an XGBoost estimator.

PARAMETER	DESCRIPTION
kind	The kind of estimator to create, either "classifier" or "regressor". TYPE: Literal['classifier', 'regressor']
name	The name of the step in the pipeline. TYPE: str | None DEFAULT: None
space	The space to use for hyperparameter optimization. Choose from "large", "performant" or provide a custom space. Todo Currently "performant" links to "large" and are the same. Warning The default space is by no means optimal, please adjust it to your needs. You can find the link to the XGBoost parameters here: xgboost.readthedocs.io/en/stable/parameter.html TYPE: Any | Literal['large', 'performant'] DEFAULT: 'large'
config	The keyword arguments to pass to the XGBoost estimator when it will be created. These will be hard set on the estimator and removed from the default space if no space is provided. TYPE: Mapping[str, Any] | None DEFAULT: None

RETURNS	DESCRIPTION
Component	A pipeline with an XGBoost estimator.

Source code in src/amltk/pipeline/xgboost.py

def xgboost_component(
    kind: Literal["classifier", "regressor"],
    name: str | None = None,
    space: Any | Literal["large", "performant"] = "large",
    config: Mapping[str, Any] | None = None,
) -> Component:
    """Create a component with an XGBoost estimator.

    Args:
        kind: The kind of estimator to create, either "classifier" or "regressor".
        name: The name of the step in the pipeline.
        space: The space to use for hyperparameter optimization. Choose from
            "large", "performant" or provide a custom space.

            !!! todo

                Currently `#!python "performant"` links to `#!python "large"` and are
                the same.

            !!! warning "Warning"

                The default space is by no means optimal, please adjust it to your
                needs. You can find the link to the XGBoost parameters here:

                https://xgboost.readthedocs.io/en/stable/parameter.html


        config: The keyword arguments to pass to the XGBoost estimator when it will be
            created. These will be hard set on the estimator and removed from the
            default space if no space is provided.

    Returns:
        A pipeline with an XGBoost estimator.
    """
    estimator_types = {
        "classifier": XGBClassifier,
        "regressor": XGBRegressor,
    }
    config = config or {}
    estimator_type = estimator_types.get(kind)
    if estimator_type is None:
        raise ValueError(
            f"Unknown kind: {kind}, please choose from {list(estimator_types.keys())}",
        )

    if name is None:
        name = str(estimator_type.__name__)

    if isinstance(space, str):
        device = config.get("device", "cpu")
        tree_method = config.get("tree_method", "auto")
        is_classifier = estimator_type is XGBClassifier
        _spaces = {
            "large": xgboost_large_space,
            "performant": xgboost_performant_space,
        }
        space_f = _spaces[space]
        space = space_f(
            is_classifier=is_classifier,
            tree_method=tree_method,
            device=device,
        )
        for key in config:
            space.pop(key, None)
    elif isinstance(space, Mapping):
        overlap = set(space.keys()).intersection(config)
        if any(overlap):
            raise ValueError(
                f"Space and kwargs overlap: {overlap}, please remove one of them",
            )

    return Component(name=name, item=estimator_type, config=config, space=space)

def xgboost_large_space(*, is_classifier, tree_method, device)

Create a large space for XGBoost hyperparameter optimization.

Source code in src/amltk/pipeline/xgboost.py

def xgboost_large_space(
    *,
    is_classifier: bool,
    tree_method: str,
    device: str,
) -> dict[str, Any]:
    """Create a large space for XGBoost hyperparameter optimization."""
    # TODO: Do we want a general space kind
    # For now we use ConfigSpace where needed
    from ConfigSpace import Float, Integer

    space = {
        "eta": Float("eta", (1e-3, 1), default=0.3, log=True),
        "min_split_loss": Float("min_split_loss", (0, 20), default=0),
        "max_depth": Integer("max_depth", (1, 20), default=6),
        "min_child_weight": Float("min_child_weight", (0, 20), default=1),
        "max_delta_step": Float("max_delta_step", (0, 10), default=0),
        "subsample": Float("subsample", (1e-5, 1), default=1),
        "colsample_bytree": Float("colsample_bytree", (1e-5, 1), default=1),
        "colsample_bylevel": Float("colsample_bylevel", (1e-5, 1), default=1),
        "colsample_bynode": Float("colsample_bynode", (1e-5, 1), default=1),
        "reg_lambda": Float("reg_lambda", (1e-5, 1e3), default=1),
        "reg_alpha": Float("reg_alpha", (0, 1e3), default=0),
    }

    if tree_method == "hist" and ("cuda" in device or "gpu" in device):
        space["sampling_method"] = ["uniform", "gradient_based"]

    if tree_method in ("hist", "approx"):
        space["max_bin"] = Integer("max_bin", (2, 512), default=256)

    if is_classifier:
        space["scale_pos_weight"] = Float("scale_pos_weight", (1e-5, 1e5), default=1)

    return space

def xgboost_performant_space(*, is_classifier, tree_method, device)

Create a performant space for XGBoost hyperparameter optimization.

Source code in src/amltk/pipeline/xgboost.py

def xgboost_performant_space(
    *,
    is_classifier: bool,
    tree_method: str,
    device: str,
) -> dict[str, Any]:
    """Create a performant space for XGBoost hyperparameter optimization."""
    warnings.warn(
        "This space is not yet optimized for performance"
        " and is subject to change in the future.",
        FutureWarning,
        stacklevel=2,
    )
    return xgboost_large_space(
        is_classifier=is_classifier,
        tree_method=tree_method,
        device=device,
    )