Serving GBT models from other libraries

Introduction

The Extension for Scikit-learn* can be used to accelerate computation of predictions (also known as “inference”) from GBT (gradient-boosted decision trees) models produced by other libraries such as XGBoost, by converting those models to class daal4py.mb.GBTDAALModel which offers faster methods daal4py.mb.GBTDAALModel.predict() and daal4py.mb.GBTDAALModel.predict_proba().

Example speedups in predictions compared to prediction method in XGBoost on Intel Xeon Platinum 8275CL (see blog post).

Model objects from other GBT libraries can be converted to this daal4py class through function daal4py.mb.convert_model in the model builders (mb) submodule, which can also convert logistic regression models (Serving Logistic Regression from Other Libraries). See About daal4py for more information about the daal4py module.

Currently, model builders from daal4py can work with GBT model objects produced by the following libraries, both in their base booster classes and their scikit-learn-compatible classes:

• XGBoost

• LightGBM

• CatBoost

Acceleration is achieved by a smart arrangement of tree representations in memory which is optimized for the way in which modern CPUs interact with RAM, along with leveraging of instruction set extensions for Intel hardware; which can result in very significant speed ups of model predictions (inference) without any loss of numerical precision compared to the original libraries that produced the models.

In addition to regular predictions, daal4py can also accelerate SHAP computations, for both feature contributions and feature interactions.

Example

Example converting an XGBoost model:

import numpy as np
import xgboost as xgb
import daal4py
from sklearn.datasets import make_regression

X, y = make_regression(n_samples=100, n_features=10, random_state=123)
dm = xgb.DMatrix(X, y)
xgb_model = xgb.train(
    params={"objective": "reg:squarederror"},
    dtrain=dm,
    num_boost_round=10
)

d4p_model = daal4py.mb.convert_model(xgb_model)

np.testing.assert_allclose(
    xgb_model.predict(dm),
    d4p_model.predict(X),
    rtol=1e-6,
)

np.testing.assert_allclose(
    xgb_model.predict(dm, pred_contribs=True),
    d4p_model.predict(X, pred_contribs=True),
    rtol=1e-6,
)

np.testing.assert_allclose(
    xgb_model.predict(dm, pred_interactions=True),
    d4p_model.predict(X, pred_interactions=True),
    rtol=1e-6,
)

More examples:

• XGBoost model conversion

• SHAP value prediction from an XGBoost model

• LightGBM model conversion

• CatBoost model conversion

Blog posts:

• Faster XGBoost, LightGBM, and CatBoost Inference on the CPU

• Improving the Performance of XGBoost and LightGBM Inference

Notes about computed results

• The shapes of SHAP contributions and interactions are consistent with the XGBoost results. In contrast, the SHAP Python package drops bias terms, resulting in SHAP contributions (SHAP interactions) with one fewer column (one fewer column and row) per observation.

• Predictions for regression objectives are computed in the link scale only (what XGBoost calls “margin” and LightGBM calls “raw”).

Limitations

• Models with categorical features are not supported.

• Multioutput models are not supported.

• SHAP values can be calculated for regression models only.

• Objectives that are not for regression nor classification (e.g. ranking) are not supported.

Documentation

See the section about model builders in the daal4py API reference for full documentation.