Unified Vision Transformer Compression

Codes for the paper: [ICLR 2022] Unified Vision Transformer Compression.

Shixing Yu*, Tianlong Chen*, Jiayi Shen, Huan Yuan, Jianchao Tan, Sen Yang, Ji Liu, Zhangyang Wang

Overall Results

Extensive experiments are conducted with several DeiT backbones on ImageNet, which consistently verify the effectiveness of our proposal. For example, UVC on DeiT-Tiny (with/without distillation tokens) yields around 50% FLOPs reduction, with little performance degradation (only 0.3%/0.9% loss compared to the baseline).

Method	Acc	FLOPs(G)	Compression Ratio (%)
DeiT-Small	79.8	4.6	100
SCOP	77.5 (-2.3)	2.6	56.4
PoWER	78.3 (-1.5)	2.7	58.7
HVT	78.0 (-1.8)	2.4	52.2
Patch Slimming	79.4 (-0.4)	2.6	56.5
UVC (Ours)	79.44 (-0.36)	2.65	57.61
UVC (Ours)	78.82 (-0.98)	2.32	50.41

Overview of Proposed UVC

We formulate and solve UVC as a unified constrained optimization problem. It simultaneously learns model weights, layer-wise pruning ratios/masks, and skip configurations, under a distillation loss and an overall budget constraint.

Implementations of UVC

Set the 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
pip install ml_collections
pip install tensorboard

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 ./

Running command

The training contains two parts.

• The first part is UVC Training. In this stage, it optimizes the architecture with primal-dual algorithm to find the optimal block-wise layout and skip configuration.
• The second part is Post Training. In this stage, the architecture is fixed while only updating the weights to help the network to regain accuracy.

Stage1 UVC Training

python -W ignore -m torch.distributed.launch \
--nproc_per_node=2 \
--master_port 6019 joint_train.py \
--gpu_num '0,1' \
--uvc_train \
--model_type deit_tiny_patch16_224 \
--model_path https://dl.fbaipublicfiles.com/deit/deit_tiny_patch16_224-a1311bcf.pth \
--distillation-type soft \
--distillation-alpha 0.1 \
--train_batch_size 512 \
--num_epochs 30 \
--eval_every 1000 \
--flops_with_mhsa 1 \
--zlr_schedule_list "1,5,9,13,17" \
--learning_rate 1e-4 \
--enable_deit 0 \
--budget 0.5 \
--enable_pruning 1 \
--enable_block_gating 1 \
--enable_patch_gating 1 \
--gating_weight 5e-4 \
--patch_weight 5 \
--patch_l1_weight 0.01 \
--patchloss "l1" \
--use_gumbel 1 \
--glr 0.1 \
--patchlr 0.01 \
--num_workers 64 \
--seed 730 \
--output_dir mc_deit_tiny_patch16_224_with_patch \
--log_interval 1000 \
--eps 0.1 \
--eps_decay 0.92 \
--enable_warmup 1 \
--warmup_epochs 5 \
--warmup_lr 1e-4 \
--z_grad_clip 0.5 \
--gating_interval 50

Stage2 Post Training

python -m torch.distributed.launch \
--nproc_per_node=2 --master_port 6382 post_train.py \
--pretrained 0 \
--model_type "deit_small_patch16_224" \
--model_path https://dl.fbaipublicfiles.com/deit/deit_small_patch16_224-cd65a155.pth \
--checkpoint_dir /home/shixing/deit_small_patch16_224_11.pth.tar \
--distillation-type soft \
--distillation-alpha 0.1 \
--train_batch_size 256 \
--gpu_num '2,3' \
--epochs 120 \
--eval_every 1000 \
--output_dir exp/deit_small_nasprune_0.58 \
--num_workers 64

Citation

@inproceedings{yu2022unified,
  author = {Yu, Shixing and Chen, Tianlong and Shen, Jiayi and Yuan, Huan and Tan, Jianchao and Yang, Sen and Liu, Ji and Wang, Zhangyang},
  title = {Unified Visual Transformer Compression},
  booktitle = {ICLR},
  year = {2022},
}

Results

• Our full compression log for reported results are under file log/.

• Checkpoint of deit-tiny-distilled-patch16-224: https://drive.google.com/drive/folders/1kjhNsppWCLuaGm-fAf4tVbVRENL14PeD?usp=sharing (with distillation token)

Acknowledgement

ViT : https://github.com/jeonsworld/ViT-pytorch

ViT : https://github.com/google-research/vision_transformer

DeiT: https://github.com/facebookresearch/deit

T2T-ViT: https://github.com/yitu-opensource/T2T-ViT