Awesome
ICCV2023: "GET: Group Event Transformer"
<img src="figures/teaser.png" alt="teaser" width="600"/>
Installations
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
pip install timm yacs spikingjelly termcolor opencv-python
Required Data
"CIFAR10DVS": spikingjelly will download this dataset automatically.<br /> "DVS128Gesture": Follow the spikingjelly instructions to download files manually.<br /> "Other datasets": See spikingjelly.
Evaluation (switch to "master_with_checkpoints" branch)
⭐CIFAR10DVS (top-1 acc. 84.8%)
python -m torch.distributed.launch --nproc_per_node 1 --master_port 23333 main.py \
--use_env --batch-size 64 --embed_split 24 --group_num 12 --patch_size 4 \
--data-path datasets/cifar10dvs --cfg configs/GET/GET_48.yaml --eval --resume checkpoints/GET_group12_cifar.pth
⭐DVS128Gesture (top-1 acc. 97.9%)
python -m torch.distributed.launch --nproc_per_node 1 --master_port 23333 main.py \
--use_env --batch-size 64 --embed_split 48 --group_num 24 --patch_size 4 \
--data-path datasets/dvs128gesture --cfg configs/GET/GET_48.yaml --eval --resume checkpoints/GET_group24_gesture.pth
Training
CIFAR10DVS
python -m torch.distributed.launch --nproc_per_node 8 --master_port 23333 main.py --use_env \
--use-checkpoint --batch-size 64 --epoch 1000 --embed_split 24 --group_num 12 \
--patch_size 4 --lr 0.001 --data-path datasets/cifar10dvs --cfg configs/GET/GET_48.yaml
DVS128Gesture
python -m torch.distributed.launch --nproc_per_node 8 --master_port 23333 main.py --use_env \
--use-checkpoint --batch-size 64 --epoch 1000 --embed_split 48 --group_num 24 \
--patch_size 4 --lr 0.0008 --data-path datasets/dvs128gesture --cfg configs/GET/GET_48.yaml
Citing GET👏
@inproceedings{peng2023get,
title={GET: Group Event Transformer for Event-Based Vision},
author={Yansong Peng and Yueyi Zhang and Zhiwei Xiong and Xiaoyan Sun and Feng Wu},
booktitle = {International Conference on Computer Vision (ICCV)},
year={2023},
}