cuANN

Artificial Neural Network for CUDA

Version 0.2

warning this isn't working properly! It was a learning experiment on how to code with CUDA and how MLPN work. If I ever get to re-writting this, I will update the readme, otherwise use at your own peril!!!

A simple Feed Forward Multilayered Neural Network, using CUDA and Thrust. I started working on this project, after I realised that (at the time of writting this) no simple CUDA-based ANN library exists for C++. Some that do exist use Python, rely on closed-source libraries, are obscure, have been abandoned or are CLI programs, rather than libraries.

cuANN is written in a minimal way, and templated where possible, in order for you to change and adapt it to your needs. It is inspired by FANN, which has for a long time been the most widely used ANN library. However, thanks to GPGPU, cuANN aspires to help you use large Neural Networks on your NVIDIA GPU.

This version (0.2) is written with CUDA Compute 3.0 in mind, and is tested on a GTX660. It has not been optimized yet, and most of the cycles are spent on cudaMemcpy and cudaMalloc. Most of the samples when run on my GTX660 use on average 25% to 35%. Next version (0.3) will hopefully be better optimised.

Building

In order to build:

mkdir build
cd build
cmake ..
make -j8

The dependencies are:

    Boost >= 1.49
    CUDA >= 6.5
    Thrust >= 1.7
    CMake >= 2.8
    g++   >= 4.8

cuANN uses C++11 features, so make sure you have a modern enough compiler (g++ or clang). If you wish to change the CUDA Compute to an earlier version, do so by editing the CMakeLists.txt file.

Examples

You can browse the simple exclusive or (XOR) example under samples/, or the more complex Abelone. The data used to train them is the same as with libFANN.

NOTE: CUDA Compute compatability.

I have tested cuANN using CUDA compute 3.0. Newer versions and GPUs should work fine, however I can't guarantee it will work for older cards. WARNING if you wish to use an older GPU, change the CMakeLists.txt and use the correct -gencode flag for your card's architecture and sm code.

NOTE: Neural Networks using Back-Propagation are bizzare machines: they may fail to learn, or may take too long to learn.

This greatly depends on the network architecture (input nodes, hidden nodes, hidden layers, etc) as well as the learning rate epsilon and the momentum alpha.