Skip to content

Ym-Shan/Spiking_Multiscale_Attention_Arxiv

Repository files navigation

open source for Spiking-Multiscale-Attention

We have provided training programs and pre trained weights for the DVS128 Gesture, CIFAR10-DVS, N-Caltech101, and Imagenet-1K datasets.

Pre training weights(one GPU can load and run.)

model datasets Models
SMA-VGG Dvs128 Gesture link
SMA-AZO-VGG Dvs128 Gesture link
SMA-VGG CIFAR10-DVS link
SMA-AZO-VGG CIFAR10-DVS link
SMA-VGG N-Caltech101 link
SMA-AZO-VGG N-Caltech101 link
SMA-ResNet18 Imagenet-1K link
SMA-ResNet34 Imagenet-1K link
SMA-AZO-ResNet104 Imagenet-1K link

Operating environment

As described in the appendix of the paper, we utilized three devices in our experiments. Device one was dedicated to conducting experiments on the DVS128 Gesture, CIFAR10-DVS, and N-Caltech101 datasets. Device two was allocated for experiments involving the Imagenet-1K dataset using the ResNet18/34 architecture. Lastly, Device three was employed for experiments on the Imagenet-1K dataset using the ResNet104 network.

The specific configurations of these three devices are shown in the table below: image

Regardless of the configuration, the only core libraries used are spikingjelly==0.0.0.0.14, einops, timm and cupy.

For other unimportant configurations, please refer to requirements.txt

Run the DVS128 Gesture

CUDA_VISIBLE_DEVICE="0,1,2,3" python -m torch.distributed.launch --nproc_per_node 4 vgg8_dvs128_SMA.py

Run the CIFAR10-DVS

CUDA_VISIBLE_DEVICE="0,1,2,3" python -m torch.distributed.launch --nproc_per_node 4 vgg8_cifar10dvs_SMA.py

Run the N-Caltech101

CUDA_VISIBLE_DEVICE="0,1,2,3" python -m torch.distributed.launch --nproc_per_node 4 vgg8_NCaltech101_SMA.py

Run the Imagenet-1K

If an error occurs, you need to switch to the folder where the training and testing sets are located to execute:

rm -rf .ipynb_checkpoints

Example of using MS-ResNet for training:

CUDA_VISIBLE_DEVICES="0,1,2,3,4,5" python -m torch.distributed.launch --master_port=1234 --nproc_per_node=6 train_amp.py -net resnet18 -b 384 -lr 0.1

Our training backbone on Imagenet mainly refers to Attention SNN.

The dataset visualization methods used in this paper have been integrated into the SpikingJelly framework:

Update

A method save_as_pic has been added to save each frame of an individual event as a .png file. Prior to this, spikingjelly only had the method play_frame to save event data as .gif format.

A method save_every_frame_of_an_entire_DVS_dataset has been added that requires only one line of code to save each frame of every sample in an entire DVS dataset as a .png file.

Please Citing

@article{shan2024advancing,
  title={Advancing Spiking Neural Networks towards Multiscale Spatiotemporal Interaction Learning},
  author={Shan, Yimeng and Zhang, Malu and Zhu, Rui-jie and Qiu, Xuerui and Eshraghian, Jason K and Qu, Haicheng},
  journal={arXiv preprint arXiv:2405.13672},
  year={2024}
}

About

Formal implementation of SMA and AZO.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages