Chasing Sparsity in Vision Transformers: An End-to-End Exploration

Codes for [NeurIPS'21] Chasing Sparsity in Vision Transformers: An End-to-End Exploration.

Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, Zhangyang Wang

Overall Results

Extensive results on ImageNet with diverse ViT backbones validate the effectiveness of our proposals which obtain significantly reduced computational cost and almost unimpaired generalization. Perhaps most surprisingly, we find that the proposed sparse (co-)training can even improve the ViT accuracy rather than compromising it, making sparsity a tantalizing “free lunch”. For example, our sparsified DeiT-Small at (5%, 50%) sparsity for (data, architecture), improves 0.28% top-1 accuracy, and meanwhile enjoys 49.32% FLOPs and 4.40% running time savings.

Proposed Framework of SViTE

Implementations of SViTE

Set Environment

conda create -n vit python=3.6

pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

pip install tqdm scipy timm

git clone https://github.com/NVIDIA/apex

cd apex

pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

pip install -v --disable-pip-version-check --no-cache-dir ./

Cmd

Command for unstructured sparsity, i.e., SViTE.

SViTE-Small

bash cmd/ vm/0426/vm1.sh 0,1,2,3,4,5,6,7

Details

CUDA_VISIBLE_DEVICES=$1 \
python -m torch.distributed.launch \
    --nproc_per_node=8 \
    --use_env main.py \
    --model deit_small_patch16_224 \
    --epochs 600 \
    --batch-size 64 \
    --data-path ../../imagenet \
    --output_dir ./small_dst_uns_0426_vm1 \
    --dist_url tcp://127.0.0.1:23305 \
    --sparse_init fixed_ERK \
    --density 0.4 \
    --update_frequency 15000 \
    --growth gradient \
    --death magnitude \
    --redistribution none

SViTE-Base

bash cmd/ vm/0426/vm3.sh 0,1,2,3,4,5,6,7

Details

CUDA_VISIBLE_DEVICES=$1 \
python -m torch.distributed.launch \
    --nproc_per_node=8 \
    --use_env main.py \
    --model deit_base_patch16_224 \
    --epochs 600 \
    --batch-size 128 \
    --data-path ../../imagenet \
    --output_dir ./base_dst_uns_0426_vm3 \
    --dist_url tcp://127.0.0.1:23305 \
    --sparse_init fixed_ERK \
    --density 0.4 \
    --update_frequency 7000 \
    --growth gradient \
    --death magnitude \
    --redistribution none

Remark. More commands can be found under the "cmd" folder.

Citation

@misc{chen2021chasing,
      title={Chasing Sparsity in Vision Transformers:An End-to-End Exploration}, 
      author={Tianlong Chen and Yu Cheng and Zhe Gan and Lu Yuan and Lei Zhang and Zhangyang Wang},
      year={2021},
      eprint={2106.04533},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}