Understanding Temporal Information Dynamics in Spiking Neural Networks

Pytorch code for [Understanding Temporal Information Dynamics in Spiking Neural Networks] - AAAI23

Dependencies

• Python 3.9
• PyTorch 1.10.0
• Spikingjelly

git clone https://github.com/fangwei123456/spikingjelly.git
cd spikingjelly
python setup.py install

Training and Computing fisher information

In this anonymous code, we provide a code for

(a) train_snn.py: train SNN from scratch

python train_snn.py  --dataset 'cifar10' --arch 'resnet19' --optimizer 'sgd' --batch_size 128 --learning_rate 3e-1 --timestep 10

(b) train_snn_fisherinfo.py: computing fisher information from pretrained model

python train_snn_fisherinfo.py --dataset 'cifar10' --arch 'resnet19'  --batch_size 16 --timestep 10

Also, for skipping (a) Train SNN from scratch, we provide pretrained parameters (link) for ResNet19_CIFAR10 from epoch 20, 120, 300.

Download three check point under snapshots/ folder