SNN_HAR

Pytorch implementation of Spiking Neural Networks for Human Activity Recognition. Paper link: Wearable-based Human Activity Recognition with Spatio-Temporal Spiking Neural Networks.

Installation and usage

Our implementation is based on Pytorch 1.10 or higher and other libraries. Please install the packages in the requirements.txt:

pip install -r requirements.txt

Datasets download

Please download the dataset and decompress them under the data directory before running the code:

• UCIHAR link
• SHAR link
• HHAR link

Running ANNs

python main.py --dataset ucihar --backbone FCN --rep 5
python main.py --dataset shar --backbone FCN --rep 5
python main.py --dataset hhar --backbone FCN --rep 5

Here, for ANN baselines, we can change our backbone to FCN (CNN in the paper), DCL, LSTM, Transformer. --rep specifies the number of repeats for computing mean and standard deviation of the accuracies.

Running SNNs

python main.py --dataset ucihar --backbone SFCN --lr 1e-3 --tau 0.75 --thresh 0.5
python main.py --dataset shar --backbone SFCN --lr 1e-3 --tau 0.25 --thresh 0.5
python main.py --dataset hhar --backbone SFCN --lr 1e-3 --tau 0.75 --thresh 0.5

For spiking versions of backbone, we offer SFCN adn SDCL. Note that they need to be trained with slightly larger learning rate. The --tau specifies the decay factor in LIF neurons (Sec 3.3 in paper).

Acknowledgement

Our code framework is strongly based on Tian0426/CL_HAR. We'd like to thank the authors of this repo for their efforts.

If you find our paper interesting, please kindly consider cite the paper.

@article{li2022wearable,
  title={Wearable-based Human Activity Recognition with Spatio-Temporal Spiking Neural Networks},
  author={Li, Yuhang and Yin, Ruokai and Park, Hyoungseob and Kim, Youngeun and Panda, Priyadarshini},
  journal={arXiv preprint arXiv:2212.02233},
  year={2022}
}