codeplaysoftware/SYCL-ML

SYCL-ML is a C++ library, implementing classical machine learning algorithms using SYCL.

SYCL-ML

What is it?

SYCL-ML is a framework providing simple classical machine learning algorithms using SYCL. It is meant to be accelerated on any OpenCL device supporting SPIR or SPIR-V. The following links give more details on what SYCL is:

• https://www.khronos.org/sycl
• https://developer.codeplay.com/computecppce/latest/sycl-guide-introduction

What can it do?

Some linear algebra operations had to be implemented such as:

• Matrix inversion
• SVD decomposition
• QR decomposition

In terms of machine learning related algorithms it includes:

• Principal Component Analysis: used to reduce the dimensionality of a problem.
• Linear Classifier (see naive Bayes classifier): classify assuming all variables are equally as important.
• Gaussian Classifier: classify using the Gaussian distribution.
• Gaussian Mixture Model: based on the EM algorithm, uses multiple Gaussian distribution for each labels.
• Support Vector Machine: C-SVM with any kernel function.

SYCL-ML is a header only library which makes it easy to integrate.

More details on what the project implements and how it works can be found on our website.

TODO list

• Optimize SVD decomposition for faster PCA. The algorithm probably needs to be changed to compute eigenpairs differently.
• Optimize SVM for GPU. More recent papers on SVM for GPU should be experimented.
• Implement an LDA (or dimensionality reduction algorithms) which would be used as a preprocessing step similarly to a PCA.
• Implement a K-means (or other clustering algorithms) which could be used to improve the initialization of the EM.
• Add a proper way to select a SYCL device.

Prerequisites

SYCL-ML has been tested with:

• Ubuntu 16.04, amdgpu pro driver 17.40
• CMake 3.0
• g++ 5.4
• ComputeCpp 1.2.0

ComputeCpp can be downloaded from the CodePlay website. Once extracted, ComputeCpp path should be set as an environment variable to COMPUTECPP_DIR (usually /usr/local/computecpp). Alternatively, it can be given as an argument to cmake with -DComputeCpp_DIR=path/to/computecpp.

Building

Build all the targets with:

mkdir build
cd build
cmake ..
make

CMake will take care of downloading the Eigen dependency and MNIST dataset. On Unix it will automatically extract the MNIST dataset using gunzip.

It is recommended to run the tests before running the examples:

cd build/tests
ctest --output-on-failure

The documentation can be built with doxygen. It requires dot from the graphviz package. Simply run:

doxygen

Contributing

The project is under the Apache 2.0 license. Any contribution is welcome! Also feel free to raise an issue for any questions or suggestions.