NeuromorphicProcessor

Neuromorphic processor implementation for Master thesis - WIP!

The accelerator may be built by

$make all

All possible tests are run by

$make test

The folder structure may be cleaned by

$make clean

Accelerator parameters are listed in src/main/scala/neuroproc/package.scala and per default target a Xilinx Kintex-7 FPGA with BUFGCE primitives inserted instead of the typical latch-based clock-gating cells used in ASICs.

Install guide (tested with Ubuntu 20.04)

The accelerator is written in Chisel3 and tests are run using ChiselTest with either the built-in Treadle backend or the Verilator backend.

Java 8

Java 8 is preferred for Chisel3 thus far, although the code runs with Java 11 as well.

$sudo apt update
$sudo apt install openjdk-8-jdk openjdk-8-jre

Make java 8 the default java version

$sudo update-alternatives --config java
$sudo update-alternatives --config javac

Choose java 8 for both

Scala build tool (SBT)

$echo "deb https://dl.bintray.com/sbt/debian /" | sudo tee -a /etc/apt/sources.list.d/sbt.list
$curl -sL "https://keyserver.ubuntu.com/pks/lookup?op=get&search=0x2EE0EA64E40A89B84B2DF73499E82A75642AC823" | sudo apt-key add
$sudo apt-get update
$sudo apt-get install sbt

Verilator

Verilator needs to be version >= 4.028.

$sudo apt-get install verilator

If apt cannot provide a new enough version, consider installing directly from GitHub

$git clone -b v4.028 https://github.com/verilator/verilator.git
$cd verilator
$autoconf
$export VERILATOR_ROOT=`pwd`
$make

Consider adding export VERILATOR_ROOT="/path/to/verilator/binary" to your .bashrc file and symlink the binary into /usr/local/bin to get access to the verilator command directly in bash.

Python libraries

$sudo apt-get install python3-pip
$pip3 install scipy seaborn bindsnet tikzplotlib

Python models

The SNN models are developed in Python using the BindsNET library. Pretrained models can be mapped to the accelerator with the mapping functions included in src/main/scala/neuroproc/package.scala. Ubuntu 18.04 and newer versions ship with Python3.

The included notebook netwkdev/KWSonSNN.ipynb presents the Google Speech Commands dataset (GSCD) with a number of plots.

The work bases itself on the DiehlAndCook2015 model included in BindsNET. A similar PyTorch network is included in the repository for comparison. Both training scripts perform checkpointing after each epoch and allow early-stopping by pressing Ctrl+C. Trained networks are pickled and stored in netwkdev/pretrained/.

PyTorch model

The model is included in its training script. Instead of modeling excitatory and inhibitory links, the model is a simple FFNN with just one hidden layer; thus, the network has the same number of neurons as the DiehlAndCook2015 network. Training this model can be done by running the training script

$python3 ./netwkdev/trainpt.py [--params]

The available parameters are as follows

Parameter	Default	Description
--gpu	False	Enables running on CUDA-based GPUs.
--use_mnist	False	Switches to using MNIST instead of GSCD.
--seed	2	Seed for the PyTorch random number generation.
--epochs	100	Number of epochs to train the network for.
--lr	0.0002	Learning rate used in the Adam optimizer.
--n_train	-1	How many training examples to use. -1 means all.
--n_valid	-1	How many validation examples to use. -1 means all.

This model achieves higher than 95% accuracy on MNIST and roughly 70% accuracy on GSCD.

BindsNET model

The model is available in netwkdev/kwsonsnn/model.py. Training this model can be done by running the training script

$python3 ./netwkdev/train.py [--params]

The available parameters are as follows

Parameter	Default	Description
--gpu	False	Enables running on CUDA-based GPUs.
--use_mnist	False	Switches to using MNIST instead of GSCD.
--seed	2	Seed for the PyTorch random number generation.
--epochs	5	Number of epochs to train the network for.
--n_clamp	1	Number of neurons to clamp while training.
--n_train	-1	How many training examples to use. -1 means all.
--n_valid	-1	How many validation examples to use. -1 means all.
--exc	4.5	Excitatory connection maximum absolute weight.
--inh	3.0	Inhibitory connection maximum absolute weight.
--time	500	Time for a single evaluation run in ms.
--dt	1.0	Time step length in ms. Number of time steps is time/dt.
--intensity	128.0	Scaling factor for input images applied before spike encoding.
--plot	False	Enables debug plots during training.
--plot_interval	250	Number of examples between updates to debug plots.
--update_interval	250	Number of examples between updates to neuron assignments and statistics prints.

This model achieves just above 50% accuracy on MNIST and roughly 40% accuracy on GSCD.