cuSNN is a C++ library that enables GPU-accelerated simulations of large-scale Spiking Neural Networks (SNNs).

• More about cuSNN
  • Layer Types
  • Neuron Models
  • Synapse Models
  • Learning Rules
• Installation
• cuSNN samples
• The Team

This project was created for the work "Unsupervised Learning of a Hierarchical Spiking Neural Network for Optical Flow Estimation: From Events to Global Motion Perception" (Paredes-Vallés, F., Scheper, K.Y., and de Croon, G.C.H.E., 2018).

If you use this library in an academic publication, please cite our work:

@article{paredes2020unsupervised,
  title={Unsupervised Learning of a Hierarchical Spiking Neural Network for Optical Flow Estimation: From Events to Global Motion Perception},
  author={Paredes-Valles, Federico and Scheper, Kirk Yannick Willehm and De Croon, Guido Cornelis Henricus Eugene},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2020},
  volume={42},
  number={8},
  pages={2051-2064},
  publisher={IEEE}
}

cuSNN is a library that consists of the following files:

In cuSNN, a SNN is defined using the following host/device classes:

The cuSNN library has been tested on Linux systems only.

• Conv2d: 2D convolution over an input neural map composed of several input channels.
• Conv2dSep: 2D separable convolution over an input neural map composed of several input channels.
• Pooling: 2D pooling over an input neural map composed of several input channels.
• Merge: 1x1 convolution with unitary weights over an input neural map composed of several input channels.
• Dense: Full connectivity to an input neural map composed of several input channels.

At the moment, only the models proposed in our work are implemented. These are:

• Leaky Integrate-and-Fire (LIF) neuron model
• Trace-based adaptive LIF neuron model
• Static synapse model

At the moment, only the following learning rules are implemented:

• Unsupervised Learning:
  • Our trace-based STDP formulation
  • Gerstner's STDP formulation
  • Kheradpisheh et al. STDP formulation
  • Shrestha et al. STDP formulation

Requirements:

• CMake 3.8 (or higher)
• NVIDIA CUDA Toolkit 8.0 (or higher)

Get the cuSNN source:

git clone https://github.com/fedepare/cuSNN.git
cd cuSNN

Install cuSNN:

Run the following Python script to install the cuSNN library:

python setup.py <install_dir>   # Python 2.*
python3 setup.py <install_dir>  # Python 3.*

By default, cuSNN installation directory is /usr/local. Hence, administrative privileges may be required. A different installation directory can be specified using the <install_dir> argument, as indicated above.

To uninstall all files, run:

xargs rm < build/install_manifest.txt

Again, administrative privileges may be required depending on your instllation directory.

Several samples are available to demonstrate the main features provided by the cuSNN library. These are stored in the submodule cuSNN-samples. To incorporate them to your current directory, run:

git submodule update --init --recursive

P.S.: This process can be a bit slow since a small dataset is included.

In Linux, the samples are built using makefiles. For this, go to the sample directory you wish and run:

make clean
make

In the makefile of each sample, there are compilation flags to adapt the simulation to your needs. Depending on their value, some of the following libraries may be required:

• PLOTTER: OpenGL and FreeGLUT for visualization
• SNAPSHOT: cnpy library for converting C++ arrays to Numpy

In case your installation directory differs from /usr/local, -lcuSNN needs to be removed from the definition of the LIBSUSR* compilation flags, and LIBSCUSTOM needs to be defined as follows:

LIBSCUSTOM = -I<install_dir>/include -L<install_dir>/lib -lcuSNN

Further, if your installation directory differs from /usr/local, you may need to include the installation directory to your LD_LIBRARY_PATH environment variable so the executable can find the library at runtime. You can do this by adding the following line to your ~/.bashrc file:

LD_LIBRARY_PATH="$LD_LIBRARY_PATH:<install_dir>/lib"
export LD_LIBRARY_PATH

or by running the following command in every new terminal window in which a cuSNN sample wants to be run:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<install_dir>/lib

Once the sample of your choice is built as explained above, simply run the executable as follows:

./build/main

cuSNN is currently maintained by Fede Paredes-Vallés, Kirk Scheper, and Guido de Croon.