oneAPI Data Analytics Library

oneAPI Data Analytics Library (oneDAL) is a powerful machine learning library that helps speed up big data analysis. oneDAL solvers are also used in Intel Distribution for Python for scikit-learn optimization.

oneAPI Data Analytics Library is an extension of Intel® Data Analytics Acceleration Library (Intel® DAAL).

oneDAL is part of oneAPI. The current branch implements version 1.1 of oneAPI Specification.

Python API
- Distributed multi-node mode
Scikit-learn patching
oneDAL Apache Spark MLlib samples
Installation
Documentation
Support
Technical Preview Features
oneDAL and Intel® DAAL

Build your high-performance data science application with oneDAL

oneDAL uses all capabilities of Intel® hardware, which allows you to get a significant performance boost for the classic machine learning algorithms.

We provide highly optimized algorithmic building blocks for all stages of data analytics: preprocessing, transformation, analysis, modeling, validation, and decision making.

oneDAL also provides Data Parallel C++ (DPC++) API extensions to the traditional C++ interfaces.

The size of the data is growing exponentially as does the need for high-performance and scalable frameworks to analyze all this data and benefit from it. Besides superior performance on a single node, oneDAL also provides distributed computation mode that shows excellent results for strong and weak scaling:

oneDAL K-Means fit, strong scaling result	oneDAL K-Means fit, weak scaling results

Technical details: FPType: float32; HW: Intel Xeon Processor E5-2698 v3 @2.3GHz, 2 sockets, 16 cores per socket; SW: Intel® DAAL (2019.3), MPI4Py (3.0.0), Intel® Distribution Of Python (IDP) 3.6.8; Details available in the article https://arxiv.org/abs/1909.11822

Refer to our examples and documentation for more information about our API.

Python API

oneDAL has a Python API that is provided as a standalone Python library called daal4py.

The example below shows how daal4py can be used to calculate K-Means clusters:

import numpy as np
import pandas as pd
import daal4py as d4p

data = pd.read_csv("local_kmeans_data.csv", dtype = np.float32)

init_alg = d4p.kmeans_init(nClusters = 10,
                           fptype = "float",
                           method = "randomDense")

centroids = init_alg.compute(data).centroids
alg = d4p.kmeans(nClusters = 10, maxIterations = 50, fptype = "float",
                 accuracyThreshold = 0, assignFlag = False)
result = alg.compute(data, centroids)

Distributed multi-node mode

Data scientists often require different tools for analysis of regular and big data. daal4py offers various processing models, which makes it easy to enable distributed multi-node mode.

import numpy as np
import pandas as pd
import daal4py as d4p

d4p.daalinit() # <-- Initialize SPMD mode
data = pd.read_csv("local_kmeans_data.csv", dtype = np.float32)

init_alg = d4p.kmeans_init(nClusters = 10,
                           fptype = "float",
                           method = "randomDense",
                           distributed = True) # <-- change model to distributed

centroids = init_alg.compute(data).centroids

alg = d4p.kmeans(nClusters = 10, maxIterations = 50, fptype = "float",
                 accuracyThreshold = 0, assignFlag = False,
                 distributed = True)  # <-- change model to distributed

result = alg.compute(data, centroids)

For more details browse daal4py documentation.

Scikit-learn patching

You can speed up Scikit-learn using Intel(R) Extension for Scikit-learn*.

Intel(R) Extension for Scikit-learn* speeds up scikit-learn beyond by providing drop-in patching. Acceleration is achieved through the use of the oneAPI Data Analytics Library that allows for fast usage of the framework suited for Data Scientists or Machine Learning users.



Technical details: HW: c5.24xlarge AWS EC2 Instance using an Intel Xeon Platinum 8275CL with 2 sockets and 24 cores per socket; SW: scikit-learn version 0.24.2, scikit-learn-intelex version 2021.2.3, Python 3.8

Intel(R) Extension for Scikit-learn* provides an option to replace some scikit-learn methods by oneDAL solvers, which makes it possible to get a performance gain without any code changes. You can patch the stock scikit-learn by using the following command-line flag:

python -m sklearnex my_application.py

Patches can also be enabled programmatically:

from sklearn.svm import SVC
from sklearn.datasets import load_digits
from time import time

svm_sklearn = SVC(kernel="rbf", gamma="scale", C=0.5)

digits = load_digits()
X, y = digits.data, digits.target

start = time()
svm_sklearn = svm_sklearn.fit(X, y)
end = time()
print(end - start) # output: 0.141261...
print(svm_sklearn.score(X, y)) # output: 0.9905397885364496

from sklearnex import patch_sklearn
patch_sklearn() # <-- apply patch
from sklearn.svm import SVC

svm_sklearnex = SVC(kernel="rbf", gamma="scale", C=0.5)

start = time()
svm_sklearnex = svm_sklearnex.fit(X, y)
end = time()
print(end - start) # output: 0.032536...
print(svm_sklearnex.score(X, y)) # output: 0.9905397885364496

For more details browse Intel(R) Extension for Scikit-learn* documentation.

oneDAL Apache Spark MLlib samples

oneDAL provides Scala and Java interfaces that match Apache Spark MlLib API and use oneDAL solvers under the hood. This implementation allows you to get a 3-18X increase in performance compared to the default Apache Spark MLlib.

Technical details: FPType: double; HW: 7 x m5.2xlarge AWS instances; SW: Intel DAAL 2020 Gold, Apache Spark 2.4.4, emr-5.27.0; Spark config num executors 12, executor cores 8, executor memory 19GB, task cpus 8

Check the samples tab for more details.

Installation

You can download the specific version of oneDAL or install from sources.

Examples

Beside C++ and Python API, oneDAL also provides APIs for DPC++ and Java:

Documentation

Refer to GitHub Wiki to browse the full list of oneDAL and daal4py resources.

Support

Ask questions and engage in discussions with oneDAL developers, contributers, and other users through the following channels:

You may reach out to project maintainers privately at onedal.maintainers@intel.com.

Security

To report a vulnerability, refer to Intel vulnerability reporting policy.

Contribute

Report issues and make feature requests using GitHub Issues.

We welcome community contributions, so check our contributing guidelines to learn more.

Feedback

Use GitHub Wiki to provide feedback about oneDAL.

Samples

Samples are examples of how oneDAL can be used in different applications:

Technical Preview Features

Technical preview features are introduced to gain early feedback from developers. A technical preview feature is subject to change in the future releases. Using a technical preview feature in a production code base is therefore strongly discouraged.

In C++ APIs, technical preview features are located in daal::preview and oneapi::dal::preview namespaces. In Java APIs, technical preview features are located in packages that have the com.intel.daal.preview name prefix.

The preview features list:

Graph Analytics:
- Undirected graph without edge and vertex weights (undirected_adjacency_vector_graph), where vertex indices can only be of type int32
- Directed graph with and without edge weights (directed_adjacency_vector_graph), where vertex indices can only be of type int32, edge weights can be of type int32 or double
  - Jaccard Similarity Coefficients for all pairs of vertices, a batch algorithm that processes the graph by blocks
- Local and Global Triangle Counting
- Single Source Shortest Paths (SSSP)
- Subgraph isomorphism algorithm for induced and non-induced subgraphs in undirected graphs (integer vertex attributes are supported, edge attributes are not supported).

oneDAL and Intel® DAAL