Tune XGBoost classification on GPU with Optuna

Open in Databricks

This notebook demonstrates hyperparameter optimization for an XGBoost classification model on a single GPU using Optuna and Databricks AI Runtime.

note

This example requires the Databricks AI environment version 6 or above.

Connect to serverless GPU

1. Click the Connect dropdown in the notebook toolbar.
2. Select Serverless GPU.
3. Open the Environment side panel.
4. Set Accelerator to 1xA10.
5. Select base environment AI v6 or above.

Configure MLflow tracking

Following the Databricks MLflow end-to-end example pattern, this notebook tracks Optuna tuning runs and the final XGBoost model with MLflow. Each tuning trial is logged as a nested run, and the final model is logged as a deployable MLflow model artifact.

Python

import mlflow
from mlflow.models import infer_signature
import xgboost as xgb

mlflow.xgboost.autolog(log_models=False)
print("MLflow XGBoost autologging enabled for parameters and metrics.")

Load and prepare data

We use the Breast Cancer Wisconsin dataset from scikit-learn, a binary classification task with 30 numerical features and 569 samples.

Python

import numpy as np
import pandas as pd
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
import xgboost as xgb

# Load dataset as a DataFrame so the logged MLflow model keeps feature names
data = load_breast_cancer(as_frame=True)
X = data.data
y = data.target

# Hold out a test set used only for the final evaluation, then split the remainder
# into train and validation sets. Tuning uses the validation set, so hyperparameters
# are never selected on the data used to report the final metrics.
X_trainval_df, X_test_df, y_trainval, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42, stratify=y
)
X_train_df, X_val_df, y_train, y_val = train_test_split(
    X_trainval_df, y_trainval, test_size=0.25, random_state=42, stratify=y_trainval
)

# Convert to XGBoost DMatrix format for efficient training
dtrain = xgb.DMatrix(X_train_df, label=y_train)
dval = xgb.DMatrix(X_val_df, label=y_val)
dtest = xgb.DMatrix(X_test_df, label=y_test)

print(f"Train: {X_train_df.shape[0]}, Validation: {X_val_df.shape[0]}, Test: {X_test_df.shape[0]}")
print(f"Features: {X_train_df.shape[1]}")
print(f"Classes: {np.unique(y)}")
display(X_train_df.head())

Define the Optuna objective function

Optuna searches over XGBoost hyperparameters using GPU-accelerated training (device: "cuda"). Following the Databricks MLflow example structure, each trial is also captured as a nested MLflow run, and pruning stops unpromising trials early.

Python

import optuna
from sklearn.metrics import log_loss
from optuna_integration import XGBoostPruningCallback

def objective(trial):
    params = {
        "tree_method": "hist",
        "device": "cuda",
        "objective": "binary:logistic",
        "eval_metric": "logloss",
        "max_depth": trial.suggest_int("max_depth", 3, 10),
        "learning_rate": trial.suggest_float("learning_rate", 1e-3, 0.3, log=True),
        "subsample": trial.suggest_float("subsample", 0.5, 1.0),
        "colsample_bytree": trial.suggest_float("colsample_bytree", 0.5, 1.0),
        "min_child_weight": trial.suggest_int("min_child_weight", 1, 10),
        "gamma": trial.suggest_float("gamma", 1e-8, 1.0, log=True),
        "reg_alpha": trial.suggest_float("reg_alpha", 1e-8, 10.0, log=True),
        "reg_lambda": trial.suggest_float("reg_lambda", 1e-8, 10.0, log=True),
    }

    n_estimators = trial.suggest_int("n_estimators", 50, 500)

    with mlflow.start_run(nested=True, run_name=f"optuna-trial-{trial.number}"):
        mlflow.set_tags({
            "phase": "tuning",
            "trial_number": trial.number,
            "model_family": "xgboost",
            "device": "cuda",
        })
        mlflow.log_param("n_estimators", n_estimators)

        # Evaluate and prune on the validation set; the test set is reserved for the final evaluation.
        pruning_callback = XGBoostPruningCallback(trial, "validation-logloss")
        model = xgb.train(
            params,
            dtrain,
            num_boost_round=n_estimators,
            evals=[(dval, "validation")],
            callbacks=[pruning_callback],
            verbose_eval=False,
        )

        preds = model.predict(dval)
        trial_logloss = log_loss(y_val, preds)
        mlflow.log_metric("trial_logloss", trial_logloss)
        return trial_logloss

Run hyperparameter optimization

We create an Optuna study to minimize validation log loss and run 50 trials. A parent MLflow run tracks the tuning session, while each trial is logged as a nested child run.

Python

optuna.logging.set_verbosity(optuna.logging.WARNING)

with mlflow.start_run(run_name="optuna-xgboost-gpu-classification") as tuning_run:
    mlflow.set_tags({
        "phase": "hyperparameter_tuning",
        "model_family": "xgboost",
        "task": "binary_classification",
        "device": "cuda",
        "optimizer": "optuna",
    })
    mlflow.log_param("n_trials", 50)

    study = optuna.create_study(direction="minimize", study_name="xgboost-gpu-tuning")
    study.optimize(objective, n_trials=50, show_progress_bar=True)

    mlflow.log_metric("best_trial_logloss", study.best_trial.value)
    mlflow.log_params({f"best_{key}": value for key, value in study.best_trial.params.items()})
    tuning_run_id = tuning_run.info.run_id

print(f"\nBest trial logloss: {study.best_trial.value:.6f}")
print("Best hyperparameters:")
for key, value in study.best_trial.params.items():
    print(f"  {key}: {value}")
print(f"\nMLflow tuning run_id: {tuning_run_id}")

Train final model with best parameters

Train the final XGBoost model using the best hyperparameters found by Optuna, evaluate it on the held-out test set, and log a deployable MLflow model artifact with signature and input example.

Python

from sklearn.metrics import accuracy_score, classification_report, confusion_matrix, log_loss
import matplotlib.pyplot as plt

# Build final parameters from the best trial
best_params = study.best_trial.params.copy()
n_estimators = best_params.pop("n_estimators")

best_params.update({
    "tree_method": "hist",
    "device": "cuda",
    "objective": "binary:logistic",
    "eval_metric": "logloss",
})

# Refit on train + validation (all non-test data) with the best hyperparameters,
# then evaluate once on the held-out test set.
X_fit_df = pd.concat([X_train_df, X_val_df])
y_fit = pd.concat([y_train, y_val])
dfit = xgb.DMatrix(X_fit_df, label=y_fit)

with mlflow.start_run(run_name="best-xgboost-gpu-model") as final_run:
    mlflow.set_tags({
        "phase": "final_model",
        "model_family": "xgboost",
        "task": "binary_classification",
        "device": "cuda",
    })
    mlflow.log_param("n_estimators", n_estimators)
    mlflow.log_params(best_params)

    # Train the final model on all non-test data
    final_model = xgb.train(
        best_params,
        dfit,
        num_boost_round=n_estimators,
        evals=[(dtest, "test")],
        verbose_eval=False,
    )

    # Predict and evaluate on the held-out test set
    y_pred_proba = final_model.predict(dtest)
    y_pred = (y_pred_proba > 0.5).astype(int)

    accuracy = accuracy_score(y_test, y_pred)
    test_logloss = log_loss(y_test, y_pred_proba)
    report_text = classification_report(y_test, y_pred, target_names=data.target_names)
    report_dict = classification_report(
        y_test, y_pred, target_names=data.target_names, output_dict=True
    )
    conf_matrix = confusion_matrix(y_test, y_pred)

    mlflow.log_metric("test_accuracy", accuracy)
    mlflow.log_metric("test_logloss", test_logloss)
    mlflow.log_dict(report_dict, "classification_report.json")

    fig, ax = plt.subplots(figsize=(4, 4))
    image = ax.imshow(conf_matrix, cmap="Blues")
    ax.set_title("Confusion Matrix")
    ax.set_xlabel("Predicted label")
    ax.set_ylabel("True label")
    ax.set_xticks([0, 1])
    ax.set_yticks([0, 1])
    ax.set_xticklabels(data.target_names)
    ax.set_yticklabels(data.target_names)
    for row_idx in range(conf_matrix.shape[0]):
        for col_idx in range(conf_matrix.shape[1]):
            ax.text(col_idx, row_idx, conf_matrix[row_idx, col_idx], ha="center", va="center")
    fig.colorbar(image, ax=ax)
    plt.tight_layout()
    mlflow.log_figure(fig, "confusion_matrix.png")
    plt.close(fig)

    signature = infer_signature(X_test_df, y_pred_proba)
    model_info = mlflow.xgboost.log_model(
        final_model,
        name="model",
        signature=signature,
        input_example=X_fit_df.head(3),
    )
    final_run_id = final_run.info.run_id

print(f"Test Accuracy: {accuracy:.4f}")
print(f"Test Log Loss: {test_logloss:.6f}\n")
print("Classification Report:")
print(report_text)
print("\nConfusion Matrix:")
print(conf_matrix)
print(f"\nFinal MLflow run_id: {final_run_id}")
print(f"Model URI: {model_info.model_uri}")

Visualize optimization results

Optuna provides built-in visualization tools to understand the optimization process and hyperparameter importance. These complement the MLflow runs logged during tuning and final model training.

Python

import matplotlib.pyplot as plt
from optuna.visualization.matplotlib import plot_optimization_history, plot_param_importances

plt.figure(figsize=(10, 4))
plot_optimization_history(study, target_name="Log Loss")
plt.tight_layout()
plt.show()

plt.figure(figsize=(10, 4))
plot_param_importances(study)
plt.tight_layout()
plt.show()

Next steps

• Databricks MLflow end-to-end example
• XGBoost GPU documentation
• Optuna documentation
• Databricks Serverless GPU compute
• Register model in Unity Catalog

Example notebook

Tune XGBoost classification on GPU with Optuna

Open notebook in new tabOpen in Databricks