用 NVIDIA cuML 和 RAPIDS 实现机器学习工作流的完整教程

我们配置教程环境，定义数据集大小和基准测试工具，并验证是否有可用的 NVIDIA GPU。在必要时，我们安装并初始化 RAPIDS cuML，设置 CuPy 和可重现性控制，并创建同步计时和结果跟踪辅助工具。我们还演示了 cuml.accel，通过运行未修改的 scikit-learn 工作负载，并将其 CPU 执行与 GPU 加速执行进行比较。

我们直接使用原生 cuML API，并探索 GPU 驻留数据如何在 CuPy、cuDF 和 cuML 组件之间移动。我们检查设备指针以理解零拷贝互操作性，并使用 cuML 输出类型控制来管理结果是留在 GPU 上还是作为 NumPy 数组返回。我们还执行 GPU 本地的训练-测试拆分，以便数据在整个工作流程中都保留在设备上。

我们对 scikit-learn 和 cuML 实现的 PCA、K-Means、最近邻、逻辑回归、随机森林和 DBSCAN 进行基准测试。我们在 GPU 上生成数据集，同步 CUDA 操作以进行公平计时，并记录每个加速算法实现的加速比。我们还比较模型行为，并保留训练好的 cuML 随机森林，以便稍后在教程中重用。

我们使用 UMAP 和 t-SNE 构建无监督 GPU 流水线，将高维数据降维成二维嵌入。我们使用可信度指标评估 UMAP 配置，选择最优嵌入，并应用 HDBSCAN 识别簇和噪声点。然后，我们可视化生成的嵌入，并使用已知的真实簇标签比较其结构。

我们专注于在训练后加速基于树模型的推理。我们在 CPU 上训练一个 scikit-learn 随机森林，在支持的情况下将其加载到 cuML 森林推理库中，并针对当前 GPU 批量大小优化推理配置。我们比较 CPU 和 GPU 的预测时间，并验证预测概率在数值上保持一致。

我们使用 cuML 基于 GPU 的置换解释器来计算 Ridge 回归模型的 SHAP 值，并将这些解释与分析线性解进行验证。我们测试 SHAP 的可加性，并可视化特征重要性，以确认计算出的归因行为符合预期。我们还将 cuML 估算器与 scikit-learn 的 RandomizedSearchCV 结合，在 GPU 上拟合模型的同时执行交叉验证超参数优化。

我们使用 pickle 序列化训练好的 cuML 随机森林模型，恢复它，并验证模型经过序列化和反序列化后预测结果保持不变。我们汇总在整个教程中收集的 CPU 和 GPU 时间结果，并在对数图上可视化生成的加速效果。最后，我们按顺序运行每个教程部分，打印积累的实际经验，并报告完整工作流的总运行时间。

总的来说，我们全面了解了 NVIDIA cuML 如何将 GPU 加速集成到现有的 scikit-learn 工作流和完全 GPU 原生的机器学习管道中。我们比较了多个核心算法的计算性能，使用 CuPy 和 cuDF 高效管理设备驻留数据，评估无监督表示和聚类质量，加速树模型推理，并直接在 GPU 上生成可解释的 SHAP 解释。我们还展示了熟悉的 scikit-learn 工具，例如 RandomizedSearchCV，可以与 cuML 估算器协同工作，在受益于 GPU 执行的同时保留已建立的机器学习开发模式。

需要与我们合作以推广您的 GitHub 仓库或 Hugging Face 页面或产品发布或网络研讨会等？请通过以下链接联系： https://forms.gle/wbash1wF6efRj8G58

以实践者为先的 AI/ML 新闻与分析，每月有超过 100 万开发者和研究人员阅读。

We configure the tutorial environment, define dataset sizes and benchmarking utilities, and verify that an NVIDIA GPU is available. We install and initialize RAPIDS cuML when necessary, set up CuPy and reproducibility controls, and create synchronized timing and result-tracking helpers. We also demonstrate cuml.accel by running an unmodified scikit-learn workload and comparing its CPU execution with GPU-accelerated execution.

We work directly with the native cuML API and explore how GPU-resident data moves between CuPy, cuDF, and cuML components. We inspect device pointers to understand zero-copy interoperability and use cuML output-type controls to manage whether results remain on the GPU or return as NumPy arrays. We also perform a GPU-native train-test split so that our data remains on the device throughout the workflow.

We benchmark scikit-learn and cuML implementations of PCA, K-Means, nearest neighbors, logistic regression, random forests, and DBSCAN. We generate datasets on the GPU, synchronize CUDA operations for fair timing, and record the speedup each accelerated algorithm achieves. We also compare model behavior and retain the trained cuML random forest so that we can reuse it later in the tutorial.

We build an unsupervised GPU pipeline using UMAP and t-SNE to reduce high-dimensional data into two-dimensional embeddings. We evaluate UMAP configurations with the trustworthiness metric, select the strongest embedding, and apply HDBSCAN to identify clusters and noise points. We then visualize the resulting embeddings and compare their structures using the known ground-truth cluster labels.

We focus on accelerating inference for tree-based models after training. We train a scikit-learn random forest on the CPU, load it into the cuML Forest Inference Library when supported, and optimize the inference configuration for the current GPU batch size. We compare CPU and GPU prediction times and validate that the predicted probabilities remain numerically consistent.

We use cuML’s GPU-based permutation explainer to calculate SHAP values for a Ridge regression model and validate those explanations against the analytical linear solution. We test SHAP additivity and visualize feature importance to confirm that the computed attributions behave as expected. We also combine cuML estimators with scikit-learn’s RandomizedSearchCV to perform cross-validated hyperparameter optimization while fitting the model on the GPU.

We serialize a trained cuML random forest with pickle, restore it, and verify that its predictions remain unchanged after the round trip. We aggregate the CPU and GPU timing results collected throughout the tutorial and visualize the resulting speedups on a logarithmic chart. Finally, we run every tutorial section in sequence, print the accumulated practical takeaways, and report the total runtime of the complete workflow.

In conclusion, we implemented a comprehensive understanding of how NVIDIA cuML integrates GPU acceleration into both existing scikit-learn workflows and fully GPU-native machine learning pipelines. We compared computational performance across several core algorithms, managed device-resident data efficiently with CuPy and cuDF, evaluated unsupervised representations and clustering quality, accelerated tree-model inference, and generated interpretable SHAP explanations directly on the GPU. We also showed that familiar scikit-learn utilities such as RandomizedSearchCV can work alongside cuML estimators, preserving established machine learning development patterns while benefiting from GPU execution.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Practitioner-first AI/ML news and analysis, read by 1M+ developers and researchers every month.