[ML Reproducibility Challenge:] Temporal Spike Sequence Learning via Backpropagation for Deep Spiking Neural Networks (TSSL-BP)

This repository is for the reproduction of Temporal Spike Sequence Learning via Backpropagation for Deep Spiking Neural Networks.

Requirements

Instructions from https://github.com/stonezwr/TSSL-BP/

Dependencies and Libraries

• python 3.7
• pytorch
• torchvision

Installation

To install requirements:

pip install -r requirements.txt

Training

Before running

Modify the data path and network settings in the config files.

Select the index of GPU in the main.py (0 by default)

Run the code

$ python main.py -config Networks/config_file.yaml
$ python main.py -config Networks/config_file.yaml -checkpoint checkpoint/ckpt.pth // load the checkpoint

Results

Performance comparison between original paper and this reproduction:

MNIST

Paper	Network Size	Time Steps	Epochs	Mean	Stddev	Best
Original paper	15C5-P2-40C5-P2-300	5	100	99.50%	0.02%	99.53%
Reproduction	15C5-P2-40C5-P2-300	5	100	99.40%	0.04%	99.47%

CIFAR 10- CNN1

Paper	Network Size	Time Steps	Epochs	Mean	Stddev	Best
Original paper	96C3-256C3-P2-384C3-P2-384C3-256C3-1024-1024	5	150	88.98%	0.27%	89.22%
Reproduction	96C3-256C3-P2-384C3-P2-384C3-256C3-1024-1024	5	150	88.96%	0.10%	89.07%

CIFAR 10- CNN2

Paper	Network Size	Time Steps	Epochs	Mean	Stddev	Best
Original paper	128C3-256C3-P2-512C3-P2-1024C3-512C3-1024-512	5	150	N/A	N/A	91.41%
Reproduction	128C3-256C3-P2-512C3-P2-1024C3-512C3-1024-512	5	150	N/A	N/A	89.61%