SpKeras: CNN-to-SNN Conversion Toolbox

Published: March 01, 2021

Convert CNN to SNN within one-line code

Spkeras_Framework

SpKeras can easily get and evaluate rate-based spiking neural networks (SNNs), by following steps:

Pre-train Convolutional neural networks (CNNs) using Tensorflow-keras
Convert CNNs into SNNs using SpKeras
Evaluate SNNs and get parameters, e.g. weights, bias and thresholds

Built With

SpKeras has extension layers, e.g. CurrentBias and SpikeForward, following Tensorflow-keras design syntax.

Tensorflow-keras

New Features

Works with Keras Functional API, for example:
- convert Add Layer

Getting Started

The package is tested in Python 3.7.6 and Tensorflow 2.3.1.

Prerequisites

Install tensorflow
```
pip install tensorflow
```

Installation

Clone the repo

git clone https://github.com/(Anonymous)/spkeras.git

Coding for SpKeras

SpKeras will detect the Activation Layer in CNN to create SpikeActivation Layer. It means all activation function should stay inside Activation Layer, including Softmax and Sigmoid.

#Sequential model
model.add(Conv2D(64, (3, 3), padding='same')
model.add(BatchNormalization())
model.add(Activation('relu'))

#Functional API
x = Conv2D( 64, (3,3), padding="same")(inputs)
x = BatchNormalization()(x)
x = Activation("relu")(x)
x = Conv2D( 64, (3,3), padding="same")(x)
x = BatchNormalization()(x)
x = Activation("relu")(x)
x = add([x, inputs])

Example

#load dataset and cnn model
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.models import load_model

(x_train, y_train), (x_test, y_test) = cifar10.load_data()
x_train = x_train/255
x_test = x_test/255
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
y_train = keras.utils.to_categorical(y_train, 10)
y_test = keras.utils.to_categorical(y_test, 10)

cnn_model = load_model('cnn_model.h5')

#Convert CNN into SNN
from spkeras.models import cnn_to_snn

#Current normalisation using cnn_to_snn
##Default: signed_bit=0, amp_factor=100, method=1, epsilon = 0.001

snn_model = cnn_to_snn(signed_bit=0)(cnn_model,x_train)

#Evaluate SNN accuracy
##Default: timesteps=256, thresholding=0.5, scaling_factor=1, noneloss=False, spike_ext=0 
_,acc = snn_model.evaluate(x_test,y_test,timesteps=256)

#Count SNN spikes
##Default: timesteps=256, thresholding=0.5, scaling_factor=1, noneloss=False, spike_ext=0, mode=0
s_max,s = snn_model.SpikeCounter(x_train,timesteps=256)

#Count neuron numbers
##Default: mode = 0
n = snn_model.NeuronNumbers(mode=0)

Attributes

'''
--------------------------
cnn_to_snn
--------------------------
sigbed_bit: bitwidth of weights, default 0 (32-bit) 
amp_factor: amplification factor, default 100
method    : default 1
epsilon   : 0.001
--------------------------
evaluate & SpikeCounter
--------------------------
timesteps   : inference time, default 256.
thresholding: default 0.5.
noneloss    : noneloss mode, default False.
spike_ext   : extra inference time, default 0. (-1 for unlimited inference time) 
--------------------------
SpikeCounter
--------------------------
mode: set 1 to count number of neurons under different spikes, default 0
--------------------------
NeuronNumber
--------------------------
mode: set 1 to exclude average pooling layer, default 0
'''

More details

For more details please follow github, paper

Dengyu Wu