EC-SNN

This is the repository of our article published in IJCAI 2024 "EC-SNN: Splitting Deep Spiking Neural Networks on Edge Devices".

Overview

Requirements

torch==2.0.1
torchvision==0.15.2
librosa==0.10.1
spikingjelly==0.0.0.0.14
numpy==1.23.5
pandas==1.5.3
scikit-learn==1.2.1
opencv-python==4.8.1.78

How to run

Examples of running commands for different purposes are listed below, please modify the corresponding parts to implement your expected task. (All running commands in shell scripts will be attached to Github Pages later.)

This repository is a simulation toolkit for researchers to learn the logistics of EC-SNN. To get the results listed in our paper, please deploy the corresponding models to edge devices like Raspberry PI!

model training

python ecsnn.py -train -arch=vgg9 -act=snn -device=cuda -data_dir=. -dataset=cifar10 -b=128

model pruning for one edge device with all classes selected

make sure the class tokens are integers starting from 0.

python ecsnn.py -prune -arch=vgg9 -act=snn -data_dir=. -dataset=cifar10 -b=128 -split_dir=./splitted/ -device=cuda -apoz=95 -c 0 1 2 3 4 5 6 7 8 9

energy consumption

python ecsnn.py -split -energy -arch=vgg9 -act=snn -device=cuda -split_dir=./splitted/ -data_dir=. -dataset=cifar10 -b=128

latency

make sure ./infer_data/ contains the specific frame you want before making an inference

python ecsnn.py -split -infer -arch=vgg9 -act=snn -device=cpu -split_dir=./splitted/ -dataset=cifar10

For more details about each argument, try reaching our Github pages later.

cifarnet quick start

Implement the following commands step by step to get quick results. We provide this part for quick access to the whole workflow of EC-SNN.

# training
python ecsnn.py -arch=cifarnet -act=snn -device=cuda -train 
python ecsnn.py -arch=cifarnet -act=ann -device=cuda -train 
python ecsnn.py -arch=cifarnet -act=snn -prune -b=128 -split_dir=./splitted/ -device=cuda -apoz=95 -c 0 1 2 3 4 5 6 7 8 9
python ecsnn.py -arch=cifarnet -act=ann -prune -b=128 -split_dir=./splitted/ -device=cuda -apoz=56 -c 0 1 2 3 4 5 6 7 8 9
python ecsnn.py -arch=cifarnet -act=snn -fusion -split_dir=./splitted/ -device=cuda -b=128
python ecsnn.py -arch=cifarnet -act=ann -fusion -split_dir=./splitted/ -device=cuda -b=128

# latency
python ecsnn.py -arch=cifarnet -act=snn -device=cuda -infer 
python ecsnn.py -arch=cifarnet -act=ann -device=cuda -infer 
python ecsnn.py -arch=cifarnet -act=snn -device=cuda -infer -split -split_dir=./splitted/
python ecsnn.py -arch=cifarnet -act=ann -device=cuda -infer -split -split_dir=./splitted/

# energy consumption
python ecsnn.py -arch=cifarnet -act=snn -device=cuda -energy -b=128
python ecsnn.py -arch=cifarnet -act=ann -device=cuda -energy -b=128
python ecsnn.py -arch=cifarnet -act=snn -device=cuda -energy -split -split_dir=./splitted/ -b=128
python ecsnn.py -arch=cifarnet -act=ann -device=cuda -energy -split -split_dir=./splitted/ -b=128

Datasets

You can download experiment data and put them into the data folder. All data are available in the links below:

• CIFAR10
• Caltech 101
• CIFAR10-DVS
• N-Caltech101 (Recommend One Drive)
• GTZAN
• UrbanSound8K

Cite

Please cite the following paper if you find our work contributes to yours in any way:

@inproceedings{ijcai2024p5149,
  title     = {EC-SNN: Splitting Deep Spiking Neural Networks on Edge Devices},
  author    = {Di, Yu and Xin, Du and Linshan, Jiang and Wentao, Tong and Shuiguang, Deng},
  booktitle = {Proceedings of the Thirty-Third International Joint Conference on
               Artificial Intelligence, {IJCAI-24}},
  year      = {2024},
}