This is a project for the proposed Homeostasis-based ANN-to-SNN conversion for Inception and Residual architecture. This work has been summarized in paper Homeostasis-based ANN-to-SNN Conversion of Inception and Residual Architecture for Object Classification submitted to IEEE ICASSP-2019. The CNN architectures that has been converted in this project includes:
- VGG-16
- ResNet-20,32,44,56
- Inception-v4
- Inception-ResNet-v2
Take the conversion of Inception as an example, three parts are included:
- pre-train CNN models:
- network definition: inception_init.m
- training CNNs: train_cifar.m
- parse Inception(dagnn format): parse_dagnn.m
- normalize weights
- construct SNN
- simulate SNN: dagnn2snn.m
- run script: conversion_dagnn.m
The training of CNNs and the simulation of SNN all are implemented on Matlab with Deep Learning Library: Matconvnet.