🌄 Graia

An experimental neural network library.

Goals

• Not using a retropropagation algorithm for training. Retropropagation works pretty well but the main goal of this project is to find a biologically plausible training algorithm that would work using only the information available locally to the nodes. Current status: Work In Progress 👩‍💻 (70% accuracy for now).
• Python API similar to TensorFlow.

Graia’s core is written in Futhark to leverage OpenCL for GPU acceleration.

Online Demo

For Graia to be usable, you’ll have to use it locally with a GPA as described below.

Prerequisites

To build Graia on a Debian/Ubuntu system, you need:

• Futhark
• Futhark FFI pip install futhark-ffi
• OpenCL
  • Native GPU drivers are prefered but if no OpenCL device is listed with clinfo -l you can install pocl-opencl-icd (slower and sometimes buggy).
  • If missing CL/cl.h error, install opencl-headers.
  • If missing -lOpenCL error, create an OpenCL link: sudo ln -s /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 /usr/lib/libOpenCL.so.

Then simply run make to test and compile the Futhark files to the OpenCL library used by the Graia Python class.

Jupyter Notebooks

To be sure to have all the Python packages needed, Anaconda is highly recommended.

• Install Futhark FFI: pip install futhark-ffi.
• As mentioned above, Native GPU drivers are prefered but if they don’t work you can Install PoCL:conda install conda-forge::pocl.

TODO

• implement real boolean values instead of infinite relu slope
• Save/load model https://documen.tician.de/pyopencl/array.html#pyopencl.array.Array.get then https://numpy.org/devdocs/reference/generated/numpy.save.html
• More Futhark tests