SNN-simulator-on-PYNQcluster

Openhec lab，Jiangnan university

Introduction

A Spiking neural network simulator NEST base on FPGA‘s cluster（LIF NEURON）

Spiking neuron network simulator NEST
SNN image classification
Neuron computing and STDP accelerator base on FPGA
MPI communication between PYNQ

This system consists of PyNN brain-like framework, NEST simulator, PYNQ framework, FPGA neurons and STDP hardware modules. As shown in picture, the top-level application design language is Python. With the assistance of the PYNN architecture, the NEST simulator is called. Various commands are interpreted by the python interpreter and the SLI interpreter, and then enter the NEST kernel. The underlying network creation according to various commands includes neuron creation, synapse connection creation, simulation time setting, etc.

On this basis, we designed FPGA neuron acceleration module and FPGA STDP synapse acceleration module, and provided acceleration modules for different computation-intensive points according to the network topology and computing requirements.

The general platform of this project integrates 16 PYNQ boards, and the board-level connection follows the TCP/IP protocol. The PYNQ-Z2 development board is based on the ZYNQ XC7Z020 FPGA and is equipped with Ethernet, HDMI input/output, MIC input, audio output, Arduino interface, Raspberry Pi interface, 2 Pmods, user LEDs, buttons and switches. Each board in the cluster contains a complete NEST emulator. Connect the PYNQ development board to the same network to build a cluster. You can customize one of the boards as the master node. When using it, first send all the data to the master node, and then evenly distribute the data to each board through the NEST's own distribution mechanism for processing. After the processing is completed, each board will send the results back to the master node.

The first advantage of using the cluster form is that it can achieve performance scalability. In addition, using a multi-node form of clusters in memory access will optimize memory access time. The cache size of a single board is limited. When the hit rate is similar, The form of cluster can also increase the calculation speed.

Quick Start

If you would like to use the Jupyter notebook on PYNQ to perform the simulation on NEST, please click here.
Petalinux
https://github.com/OpenHEC/SNN-simulator-on-PYNQcluster/tree/master/petalinux

Install NEST-14.0-FPGA

In this project, we use PYNQ-Z2 v2.5 PYNQ image. (If you use the old version of the image file, it should also work.)
Installing PyNN requires:

 sudo apt-get install -y \  
 cython \  
 libgsl-dev \  
 libltdl-dev \  
 libncurses-dev \  
 libreadline-dev \  
 python-all-dev \  
 python-numpy \  
 python-scipy \  
 python-matplotlib \  
 python-nose \  
 openmpi-bin \  
 libopenmpi-dev

Configure NEST:

cd NEST-14.0-FPGA  
cmake -Dwith-python=3 -DCMAKE_INSTALL_PREFIX:PATH=./ </path/to/NEST/src>

Compile and install NEST:

make  
make install  
make installcheck  
source </path/to/nest_install_dir>/bin/nest_vars.sh

References: https://nest-simulator.readthedocs.io/en/stable/installation/linux_install.html

Install PYNN

Installing PyNN requires:

Python (version 2.7, 3.3-3.7)  
a recent version of the NumPy package  
the lazyarray package  
the Neo package (>= 0.5.0)  
at least one of the supported simulators: e.g. NEURON, NEST, or Brian.

Optional dependencies are:

mpi4py (if you wish to run distributed simulations using MPI)  
either Jinja2 or Cheetah (templating engines)  
the CSA library

Install PyNN:

pip install pyNN

References:http://neuralensemble.org/docs/PyNN/installation.html

Run snn_object:

1 PYNQ node:

cd snn_object
python3 dump-c1-spikes.py --training-dir airplanes_10_6 --dataset-label train

8 PYNQ node:

mpirun -n 8 -machinefile ./machinefile python3 dump-c1-spikes.py --training-dir airplanes_10_6 --dataset-label train

1 PYNQ node run with jupyter notebook: dump_c1_spikes.ipynb
others commond: https://github.com/roberthangu/snn_object_recognition

Repo organization

The repo is organized as follows:

snn_object：Describes a new biologically plausible mechanism for generating intermediate-level visual representations using an unsupervised learning scheme.
- iaf_psc_exp.bin：LIF NEURON bitstream
NEST-14.0-FPGA：implemention Spiking neural network simulator NEST on FPGA cluster
hls: NEST LIF Neuron accelerator implemented in vivado HLS 2018.2.
vivado: creating vivado project to get block_design.tcl and bitstream.
NEST_PYNQ_Jupyter: An example of image classification based on Jupyter Notebook on PYNQ.

References

NEST

NEST is a simulator for spiking neural network models, ideal for networks of any size.
NEST Simulation:http://www.nest-simulator.org
NEST-14.0 github：https://github.com/nest/nest-simulator/releases/tag/v2.14.0

pyNN

A Python package for simulator-independent specification of neuronal network models.
PyNN :http://neuralensemble.org/PyNN/

NEST base on GPU

Implemention Spiking neural network simulator NEST on multi-GPU and distributed GPU
https://github.com/pnquanganh/opencl-nest
Nguyen Q A P, Andelfinger P, Cai W, et al. Transitioning Spiking Neural Network Simulators to Heterogeneous Hardware[C]//Proceedings of the 2019 ACM SIGSIM Conference on Principles of Advanced Discrete Simulation. ACM, 2019: 115-126.

SNN simulator base on FPGA

A.Podobas,S.Matsuoka,Luk Wayne.Designing and accelerating spiking neural networks using OpenCL for FPGAs[C].//International Conference on Field Programmable Technology (ICFPT).Melbourne,VIC, Australia: IEEE,2017.

snn_object_recognition on PYNN-NEST

Describes a new biologically plausible mechanism for generating intermediate-level visual representations using an unsupervised learning scheme.
https://github.com/roberthangu/snn_object_recognition
Masquelier, Timothée, Thorpe S J. Unsupervised Learning of Visual Features through Spike Timing Dependent Plasticity[J].PLoS Computational Biology, 2007, 3(2):e31.

others

Schuman C D , Potok T E , Patton R M , et al. A Survey of Neuromorphic Computing and Neural Networks in Hardware

OpenHEC / SNN-simulator-on-PYNQcluster