rapidsai/cuml

Features

• CUDA-Accelerated Libraries - Provides a comprehensive suite of machine learning tools and data preprocessing operations accelerated by NVIDIA CUDA.
• Distributed Learning - Provides a framework for training and performing inference on machine learning models across multiple GPUs and compute nodes.
• Distributed GPU Computing - Scales machine learning workloads across multiple GPUs and compute nodes to process datasets exceeding single-device memory.
• Feature Scale Normalization - Adjusts the range and distribution of numeric features using standardization, min-max, and robust scaling.
• GPU-Accelerated Machine Learning Libraries - Implements a wide array of tabular machine learning algorithms optimized for high performance on NVIDIA GPUs.
• Inference Scaling - Distributes both training and inference for algorithms across multiple nodes and hardware accelerators.
• Clustering Algorithms - Groups unlabeled data points using high-performance implementations of K-Means, DBSCAN, HDBSCAN, and Agglomerative Clustering.
• Large Scale Clustering - Provides high-performance implementations of K-Means and DBSCAN to group unlabeled data points in massive datasets.
• GPU Resource Management - Controls device selection, memory allocation, and stream synchronization to optimize GPU utilization for machine learning workloads.
• Machine Learning Classification - Builds predictive classification models using accelerated algorithms such as Random Forests, SVMs, Naive Bayes, and Logistic Regression.
• GPU-Accelerated Estimators - Accelerates scikit-learn and other machine learning workflows on GPUs with no changes required to the source code.
• GPU-Accelerated Training - Executes classification, regression, and clustering algorithms on NVIDIA GPUs to drastically reduce processing time.
• Regression Training - Fits linear and nonlinear regression models using hardware acceleration to predict continuous values.
• Estimator Interface Adaptors - Implements standard estimator interfaces to ensure seamless compatibility with common science tools and pipelines.
• Scikit-Learn Implementations - Implements a scikit-learn compatible estimator interface to allow seamless replacement of CPU models with GPU versions.
• Categorical Encodings - Transforms categorical features into numeric arrays using one-hot, binarization, or target encoding methods.
• Distributed Computing - Distributes algorithm computation and data across multiple hardware devices to handle massive datasets.
• GPU-Accelerated Processing - Processes large datasets exceeding single-device memory by spreading computation across multiple GPUs and nodes.
• Missing Value Imputation - Fills gaps in datasets using univariate imputation to complete missing data points.
• Tabular Data Preprocessing - Provides hardware-accelerated cleaning and transformation of structured tabular data for machine learning.
• Dimensionality Reduction - Implements accelerated PCA, t-SNE, and UMAP to compress high-dimensional datasets into fewer components.
• Tabular Data Preprocessing - Provides hardware-accelerated scaling, encoding, and imputation for tabular data to optimize machine learning workflows.
• CUDA Compute Kernels - Implements custom CUDA kernels to offload heavy mathematical computations to GPU cores for massive parallelization.
• GPU Linear Algebra Libraries - Leverages highly optimized CUDA libraries for GPU-accelerated matrix operations and linear algebra.
• Nearest Neighbor Searches - Identifies the closest data points in multi-dimensional feature spaces using hardware acceleration.
• Hardware Resource Allocation - Provides capabilities for selecting and assigning specific GPU devices and managing memory allocation.
• Tree Model Accelerators - Provides specialized acceleration for the execution speed and deployment of decision tree and forest-based models.
• Hardware Acceleration - Executes clustering, regression, classification, and dimensionality reduction algorithms on accelerators to minimize processing time.
• ML Performance Profilers - Provides tools for analyzing performance across CPUs and accelerators to optimize large-scale machine learning workloads.
• Model Serialization - Provides tools and formats for saving and loading machine learning model states and training metadata for persistence.
• Accelerated Forest Ensembles - Accelerates the training and prediction of random forests and sparse forests using hardware acceleration to increase throughput.
• DLPack Array Exchanges - Utilizes standard array interfaces and protocols for seamless zero-copy data exchange between frameworks.
• Text Feature Extraction - Converts raw text documents into numeric matrices using token counts, hashing, and weighted frequency calculations.
• Zero-Copy Memory Mappings - Uses shared memory pointers and standard array interfaces to move data between libraries without expensive serialization.
• Dataframe Format Conversion - Processes data directly from various in-memory dataframe and tensor formats without requiring manual conversion.
• Dataframe Interoperability - Enables seamless data movement between processing libraries by accepting various input formats like arrays and dataframes.
• Time Series Modeling - Forecasts and models temporal data using hardware-accelerated ARIMA and Holt-Winters exponential smoothing methods.
• Python-C++ Bindings - Provides high-performance Python bindings to execute critical C++ and CUDA logic for machine learning.
• AI & Machine Learning - GPU-accelerated machine learning library
• General Machine Learning - GPU-accelerated machine learning algorithms.
• Machine Learning - GPU-accelerated machine learning algorithms.
• Machine Learning Frameworks - GPU-accelerated machine learning algorithms.
• Machine Learning Packages - GPU-accelerated machine learning algorithms.
• Computation and Optimization - Suite of GPU-accelerated machine learning algorithms.
• Model Optimization - GPU-accelerated implementations of traditional machine learning algorithms.

Star history