Home

Welcome to arac.

arac is a neural networking library that aims to combine performance and flexibility. It supports the composition of modern algorithms in nearly arbitrary topologies and ease of use via an API to the PyBrain framework written in Python while letting the workload being managed by a C++ core.

Features

arac lets you build networks as a direct acyclic graph of modules and connections. Modules can be simple things as sigmoid squashing units as well as the very powerful LSTM cells. Connections can be parametrized and thus be optimized by error back propagation.

Work in progress

This software is constantly enhanced. This means that the API is not stable and that the software is likely to contain bugs. Although it has been used on various topologies, has thousands of lines of unit tests it is still likely to fail at one point or the other.

Also, certain decisions are not thought out to the very end and may be considered experimental.