Dask-ML provides machine learning utilities built on top of the scalable Dask platform. Dask already offers first-class integration with RAPIDS, and Dask-ML is no exception.

The Dask-ML hyperparameter search tools make it easy to take advantage of grid search, randomized search, or hyperband HPO algorithms. It particularly excels at incorporating cross-validation into the HPO process for more stable accuracy estimates and at allowing intelligent reuse of intermediate results from the Dask task graph.

This sample notebook shows how to use Ray Tune to optimize XGBoost and cuML Random Forest classifiers over a large dataset of airline arrival times. By design, it is very similar to the RAPIDS examples provided for other cloud and bring-your-own-cloud HPO offerings.

You need both Jupyter and RAPIDS 0.13 or later installed to begin. See https://rapids.ai/start.html for instructions. We recommend using the 21.06 packages for the latest updates. Dask-ML can be installed via conda or pip, following the instructions from: https://ml.dask.org/install.html.

Dask_cuML_Exploration and Dask_cuML_Exploration_Full provide a template for launching a dask cluster on top of your kubernetes environment, loading the NYC-Taxi dataset, and generating performance metrics for the available cuML Multi-Node Multi-GPU (MNMG) algorithms in your environment.

See: Dask-Kubernetes and cuML's API documentation for additional information.