Implementation of the paper Keys to Accurate Feature Extraction Using Residual Spiking Neural Networks

Published in IOP Neuromorphic Computing and Engineering

Cite as:

@article{vicentesola2022keys,
author={Alex Vicente-Sola and Davide L Manna and Paul Kirkland and Gaetano Di Caterina and Trevor Bihl},
title={Keys to accurate feature extraction using residual spiking neural networks},
journal={Neuromorphic Computing and Engineering },
doi = {10.1088/2634-4386/ac8bef},
url = {https://doi.org/10.1088/2634-4386/ac8bef},
year={2022},
month = {sep},
publisher = {{IOP} Publishing},
volume = {2},
number = {4},
pages = {044001},
}

Code tested in Python 3.8.10 and PyTorch 1.8.1

Linux:
sh requirements

Windows:
requirements.bat

• CIFAR-10:

python train.py --num_steps 50 --lr 0.026844 --leak_mem 0.8744 --arch 'sresnet' --dataset  'cifar10' --num_workers 4 --num_epochs 70

• CIFAR-100:

python train.py --num_steps 50 --lr 0.026844 --leak_mem 0.8744 --arch 'sresnet' --dataset  'cifar100' --num_workers 4 --num_epochs 70

• DVS-CIFAR10:

    python train.py --num_steps 50 --lr 0.026844 --leak_mem 0.8744 --arch 'sresnet_nm' --dataset  'cifar10dvs' --num_workers 4 --num_epochs 70

Set the batch size with the flag: --batch_size (use the biggest batch size your GPU can support)

You can set the GPU device to use with the flag --device

You can set S-ResNet's depth using the flag --n and its width using the flag --nFilters

To resume training from a saved checkpoint, indicate the checkpoint location using the flag --reload

If you want to fine tune a checkpoint trained with a different dataset, use the flag --fine_tune. This will start the epoch count from 0, reset the accuracy history and skip the loading of the Fully Connected layer and Conv1.

To use the boosting strategy in the final layer use the flag --boosting. Note that if a checkpoint was trained using this strategy, the flag will need to be present when we load it for further training or testing. The pre-trained network in CIFAR-10 "cifar10_sresnet-boost.pth.tar" needs this flag.

For testing, the location of the saved model is passed with --model_path alongside the architecture flags such as --arch --num_steps. Remember to keep the same architecture and parameters that you used for training.

Also notice that the batch size can be bigger for inference than it was at training time.

• CIFAR-10:

python test.py --model_path "path to saved model" --num_steps 50 --arch 'sresnet' --dataset 'cifar10'

• CIFAR-100:

python test.py --model_path "path to saved model" --num_steps 50 --arch 'sresnet' --dataset 'cifar100'

• DVS-CIFAR10:

python test.py --model_path "path to saved model" --num_steps 50 --arch 'sresnet_nm' --dataset 'cifar10dvs'