rachmadvwp / hybrid-snn-conversion

Training spiking networks with hybrid ann-snn conversion and spike-based backpropagation

Enabling Deep Spiking Neural Networks with Hybrid Conversion and Spike Timing Dependent Backpropagation

This is the code related to the paper titled "Enabling Deep Spiking Neural Networks with Hybrid Conversion and Spike Timing Dependent Backpropagation" published in ICLR, 2020

Training Methodology

The training is performed in the following two steps:

• Train an ANN ('ann.py')
• Convert the ANN to SNN and perform spike-based backpropagation ('snn.py')

Files

• 'ann.py' : Trains an ANN, the architecutre, dataset, training settings can be provided an input argument
• 'snn.py' : Trains an SNN from scratch or performs ANN-SNN conversion if pretrained ANN is available.
• /self_models : Contains the model files for both ANN and SNN
• 'ann_script.py' and 'snn_script.py': These scripts can be used to design various experiments, it creates 'script.sh' which can be used to run multiple models

Trained ANN models

• VGG5 CIFAR10
• VGG16 CIFAR10
• VGG11 CIFAR100

Trained SNN models

• VGG5 CIFAR10
• VGG16 CIFAR10
• VGG11 CIFAR100

Issues

• Sometimes the 'STDB' activation becomes unstable during training, leading to accuracy drop. The solution is to modulate the alpha and beta parameter or change the activation to 'Linear' in 'main.py'
• Another reason for drop in accuracy could be the leak parameter. Please change 'leak_mem=1.0' in 'main.py'. This changes the leaky-integrate-and-fire (LIF) neuron to integrate-and-fire (IF) neuron.

Citation

If you use this code in your work, please cite the following paper

@inproceedings{
Rathi2020Enabling,
title={Enabling Deep Spiking Neural Networks with Hybrid Conversion and Spike Timing Dependent Backpropagation},
author={Nitin Rathi and Gopalakrishnan Srinivasan and Priyadarshini Panda and Kaushik Roy},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=B1xSperKvH}
}