GPU-accelerated model training with cuML and XGBoost: Worked Example — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)

GPU-Accelerated Model Training with cuML and XGBoost: Worked Example

This worked example demonstrates how to leverage RAPIDS cuML and XGBoost libraries for GPU-accelerated machine learning model training, focusing on a realistic classification problem. We will walk through data preparation, model training, evaluation, and interpretation using GPU acceleration to significantly speed up the workflow.

Scenario

Suppose we have a dataset containing customer information and want to predict whether a customer will subscribe to a service (binary classification). The dataset includes numerical and categorical features.

Step 1: Data Preparation on GPU

First, load the dataset into a GPU DataFrame using cudf, RAPIDS’ GPU-accelerated DataFrame library.

• Import cudf and load CSV data directly to GPU memory.
• Handle missing values and encode categorical variables using GPU-accelerated methods.

Example:

Code Snippet

import cudfdf = cudf.read_csv('customer_data.csv')df['category_encoded'] = df['category'].astype('category').cat.codes

Step 2: Splitting Data

Use cuml.model_selection.train_test_split to split data into training and testing sets directly on the GPU.

Code Snippet

from cuml.model_selection import train_test_splitX = df.drop(['target'], axis=1)y = df['target']X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

Step 3: Training a Classification Model with cuML

Train a GPU-accelerated logistic regression or random forest classifier using cuML.

• Initialize the model (e.g., cuml.LogisticRegression()).
• Fit the model on the training data.

Code Snippet

from cuml.linear_model import LogisticRegressionmodel = LogisticRegression()model.fit(X_train, y_train)

Step 4: Training with GPU-Accelerated XGBoost

For gradient boosting, use the GPU-enabled xgboost API by specifying the tree_method='gpu_hist' parameter.

• Convert data to xgboost.DMatrix format.
• Set parameters for binary classification and GPU usage.
• Train the model with xgb.train().

Code Snippet

import xgboost as xgbdtrain = xgb.DMatrix(X_train.to_pandas(), label=y_train.to_pandas())dtest = xgb.DMatrix(X_test.to_pandas(), label=y_test.to_pandas())params = {'objective': 'binary:logistic', 'tree_method': 'gpu_hist', 'eval_metric': 'auc'}bst = xgb.train(params, dtrain, num_boost_round=100)

Step 5: Model Evaluation

Evaluate the trained models using GPU-accelerated predictions and compute performance metrics such as accuracy, AUC, and confusion matrix.

• Predict on test data.
• Calculate metrics with RAPIDS or transfer results to CPU if needed.

Code Snippet

y_pred = model.predict(X_test)from cuml.metrics import accuracy_scoreaccuracy = accuracy_score(y_test, y_pred)

For XGBoost:

y_pred_prob = bst.predict(dtest)y_pred_label = (y_pred_prob > 0.5).astype(int)

Step 6: Interpretation of Results

Interpret the confusion matrix and other metrics to assess model generalization:

• True Positives (TP): Correct positive predictions
• False Positives (FP): Incorrect positive predictions
• True Negatives (TN): Correct negative predictions
• False Negatives (FN): Incorrect negative predictions

Use these to calculate precision, recall, and F1-score, which provide deeper insight into model performance beyond accuracy.

Summary

This example highlights how RAPIDS cuML and GPU-accelerated XGBoost enable efficient training and evaluation of machine learning models on GPUs. By keeping data and computation on the GPU, data scientists can achieve significant speedups compared to CPU-based workflows, which is critical for large datasets and iterative model tuning.

For more detailed documentation and examples, visit the official RAPIDS AI website: https://rapids.ai/ and XGBoost GPU documentation: https://xgboost.readthedocs.io/en/stable/gpu/index.html.

More in this topic

Model evaluation and generalization assessment — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)GPU-accelerated model training with cuML and XGBoost: Common Mistakes — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)Performance metrics and confusion matrix interpretation — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)Regression, classification, and clustering techniques — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)Machine Learning With RAPIDS — NVIDIA-Certified Associate: Accelerated Data ScienceGPU-accelerated model training with cuML and XGBoost: Practice Questions — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)GPU-accelerated model training with cuML and XGBoost: Quick Reference — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)Hyperparameter tuning and cross-validation — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)GPU-accelerated model training with cuML and XGBoost — Machine Learning With RAPIDS (NVIDIA-Certified Associate: Accelerated Data Science)