cuML - GPU Machine Learning Algorithms

cuML is a suite of libraries that implement machine learning algorithms and mathematical primitives functions that share compatible APIs with other RAPIDS projects.

cuML enables data scientists, researchers, and software engineers to run traditional tabular ML tasks on GPUs without going into the details of CUDA programming. In most cases, cuML’s Python API matches the API from scikit-learn.

For large datasets, these GPU-based implementations can complete 10-50x faster than their CPU equivalents. For details on performance, see the cuML Benchmarks Notebook.

As an example, the following Python snippet loads input and computes DBSCAN clusters, all on GPU, using cuDF:

import cudf
from cuml.cluster import DBSCAN

# Create and populate a GPU DataFrame
gdf_float = cudf.DataFrame()
gdf_float['0'] = [1.0, 2.0, 5.0]
gdf_float['1'] = [4.0, 2.0, 1.0]
gdf_float['2'] = [4.0, 2.0, 1.0]

# Setup and fit clusters
dbscan_float = DBSCAN(eps=1.0, min_samples=1)
dbscan_float.fit(gdf_float)

print(dbscan_float.labels_)

Output:

0    0
1    1
2    2
dtype: int32

cuML also features multi-GPU and multi-node-multi-GPU operation, using Dask, for a growing list of algorithms. The following Python snippet reads input from a CSV file and performs a NearestNeighbors query across a cluster of Dask workers, using multiple GPUs on a single node:

Initialize a LocalCUDACluster configured with UCXX for fast transport of CUDA arrays

# Initialize UCX for high-speed transport of CUDA arrays
from dask_cuda import LocalCUDACluster

# Create a Dask single-node CUDA cluster w/ one worker per device
cluster = LocalCUDACluster(protocol="ucx",
                           enable_tcp_over_ucx=True,
                           enable_nvlink=True,
                           enable_infiniband=False)

Load data and perform k-Nearest Neighbors search. cuml.dask estimators also support Dask.Array as input:


from dask.distributed import Client
client = Client(cluster)

# Read CSV file in parallel across workers
import dask_cudf
df = dask_cudf.read_csv("/path/to/csv")

# Fit a NearestNeighbors model and query it
from cuml.dask.neighbors import NearestNeighbors
nn = NearestNeighbors(n_neighbors = 10, client=client)
nn.fit(df)
neighbors = nn.kneighbors(df)

For additional examples, browse our complete API documentation, or check out our example walkthrough notebooks. Finally, you can find complete end-to-end examples in the notebooks-contrib repo.

Supported Algorithms

Category	Algorithm	Notes
Clustering	Density-Based Spatial Clustering of Applications with Noise (DBSCAN)	Multi-node multi-GPU via Dask
	Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN)
	K-Means	Multi-node multi-GPU via Dask
	Single-Linkage Agglomerative Clustering
	Spectral Clustering
Dimensionality Reduction	Principal Components Analysis (PCA)	Multi-node multi-GPU via Dask
	Incremental PCA
	Truncated Singular Value Decomposition (tSVD)	Multi-node multi-GPU via Dask
	Uniform Manifold Approximation and Projection (UMAP)	Multi-node multi-GPU Inference via Dask
	Random Projection
	t-Distributed Stochastic Neighbor Embedding (TSNE)
	Spectral Embedding
Linear Models for Regression or Classification	Linear Regression (OLS)	Multi-node multi-GPU via Dask
	Linear Regression with Lasso or Ridge Regularization	Multi-node multi-GPU via Dask
	ElasticNet Regression
	LARS Regression	(experimental)
	Logistic Regression	Multi-node multi-GPU via Dask-GLM demo
	Naive Bayes	Multi-node multi-GPU via Dask
	Stochastic Gradient Descent (SGD), Coordinate Descent (CD), and Quasi-Newton (QN) (including L-BFGS and OWL-QN) solvers for linear models
Nonlinear Models for Regression or Classification	Random Forest (RF) Classification	Experimental multi-node multi-GPU via Dask
	Random Forest (RF) Regression	Experimental multi-node multi-GPU via Dask
	Inference for decision tree-based models	Forest Inference Library (FIL)
	K-Nearest Neighbors (KNN) Classification	Multi-node multi-GPU via Dask+UCXX, uses Faiss for Nearest Neighbors Query.
	K-Nearest Neighbors (KNN) Regression	Multi-node multi-GPU via Dask+UCXX, uses Faiss for Nearest Neighbors Query.
	Support Vector Machine Classifier (SVC)
	Epsilon-Support Vector Regression (SVR)
Preprocessing	Standardization, or mean removal and variance scaling / Normalization / Encoding categorical features / Discretization / Imputation of missing values / Polynomial features generation / and coming soon custom transformers and non-linear transformation	Based on Scikit-Learn preprocessing
Time Series	Holt-Winters Exponential Smoothing
	Auto-regressive Integrated Moving Average (ARIMA)	Supports seasonality (SARIMA)
Model Explanation	SHAP Kernel Explainer	Based on SHAP
	SHAP Permutation Explainer	Based on SHAP
Execution device interoperability		Run estimators interchangeably from host/cpu or device/gpu with minimal code change demo
Other	K-Nearest Neighbors (KNN) Search	Multi-node multi-GPU via Dask+UCXX, uses Faiss for Nearest Neighbors Query.

Installation

See the RAPIDS Release Selector for the command line to install either nightly or official release cuML packages via conda, pip, or Docker.

Build/Install from Source

See the build guide.

Scikit-learn Compatibility

cuML is compatible with scikit-learn version 1.4 or higher.

Contributing

Please see our guide for contributing to cuML.

References

The RAPIDS team has a number of blogs with deeper technical dives and examples. You can find them here on Medium.

For additional details on the technologies behind cuML, as well as a broader overview of the Python Machine Learning landscape, see Machine Learning in Python: Main developments and technology trends in data science, machine learning, and artificial intelligence (2020) by Sebastian Raschka, Joshua Patterson, and Corey Nolet.

Please consider citing this when using cuML in a project. You can use the citation BibTeX:

@article{raschka2020machine,
  title={Machine Learning in Python: Main developments and technology trends in data science, machine learning, and artificial intelligence},
  author={Raschka, Sebastian and Patterson, Joshua and Nolet, Corey},
  journal={arXiv preprint arXiv:2002.04803},
  year={2020}
}

Contact

Find out more details on the RAPIDS site

Open GPU Data Science

The RAPIDS suite of open source software libraries aim to enable execution of end-to-end data science and analytics pipelines entirely on GPUs. It relies on NVIDIA® CUDA® primitives for low-level compute optimization, but exposing that GPU parallelism and high-bandwidth memory speed through user-friendly Python interfaces.

Reference Issues/PRs

What does this implement/fix? Explain your changes.

Any other comments?