ViNN - A modular OpenCL accelerated library for Deep Learning

ViNN is a cross platform MIT-licensed C++ library for training and evaluating artificial neural networks using OpenCL. ViNN parallelizes computationally intensive linear algebra routines using OpenCL, which provides significant performance gains over a single threaded CPU based implementation. Python bindings with NumPy integration ease developing new models and enable interoperability with existing tools and libraries.

The design goals for ViNN are

• correctness - use gradient checking, static analysis and strive for 100% unit test coverage
• modularity - for easy integration, extensibility and testing
• performance - optimize performance critical parts using OpenCL

Features

Linear Algebra

ViNN ships with a CPU/C++ based and an OpenCL based linear algebra backend that supports most common matrix operations needed for neural networks and machine learning applications.

Activation Functions

• Linear
• Sigmoid
• Softmax
• Hyperbolic Tangent

Cost Functions

• Squared error
• Cross entropy

Training

• Batch gradient descent with early stopping
• Stochastic/Minibatch gradient descent with early stopping
• L2 regularization

Result Measurements

• Confusion table for binary classification
• Multiclass performance measures (average, micro, macro)

Input/Output Formats

• CSV
• libsvm format

Developing ViNN

Depencies

Integrating with a project:

• c++ compiler with c++11 support (clang or gcc >= 4.8)
• OpenCL 1.1 driver and development headers
• Swig 3

Additional dependencies when building from source:

• Ruby
• Python 3, NumPy

Additional dependencies for committers:

• clang-format
• lcov

Mac OS X (10.10)

brew install lcov
brew install clang-format
brew install swig

On Ubuntu (15.04)

sudo apt-get install build-essential ocl-icd-libopencl1 opencl-headers ruby

On Debian Wheezy or Raspbian

sudo apt-get install gcc-4.8
sudo apt-get install g++-4.8
sudo apt-get install libgl1-mesa-dev

Building from Source

1. Create build directory & configure Debug build:

   mkdir build cd build

   On Ubuntu 15 and Mac OS X:

   cmake -DCMAKE_BUILD_TYPE=Debug ..

   On Debian Wheezy or Raspbian

   CC=gcc-4.8 CXX=g++-4.8 cmake -DCMAKE_BUILD_TYPE=Debug ..

2. Build library and unit tests

   make

3. Run all tests:

   make test # OR ctest

4. Run unit tests & generate coverage report

   make coverage

5. Format source code after making changes:

   make format

6. Build package

   cpack --config CPackConfig.cmake

7. Build Python bindings

   Source distribution

   cd build/bindings/python python setup.py sdist pip install --global-option=build_ext --global-option="--swig-opts=-I/Users/ville/projects/vinn/src -c++" dist/vinnpy-0.2.0.tar.gz

   Binary wheel distribution

   cd build/bindings/python python setup.py build_ext --swig-opts="-I/Users/ville/projects/vinn/src -c++" bdist_wheel