[TPAMI 2023] Vision Transformer with Quadrangle Attention

This is the official repository of the paper Vision Transformer with Quadrangle Attention.

Qiming Zhang, Jing Zhang, Yufei Xu, and Dacheng Tao

News | Abstract | Method | Usage | Results | Statement

Current applications

Classification: Hierarchical models has been released; Plain ones will be released soon.

Object Detection: Will be released soon;

Semantic Segmentation: Will be released soon;

Human Pose: Will be released soon

News

24/01/2024

• The code of hierarchical models on classification has been released.

30/12/2023

• The paper is accepted by IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) with IF 24.314.

27/03/2023

• The paper is post on arxiv! The code will be made public available once cleaned up.

Abstract

This repository contains the code, models, test results for the paper Vision Transformer with Quadrangle Attention, which is an substantial extention of our ECCV 2022 paper VSA. We extends the window-based attention to a general quadrangle formulation and propose a novel quadrangle attention. We employs an end-to-end learnable quadrangle regression module that predicts a transformation matrix to transform default windows into target quadrangles for token sampling and attention calculation, enabling the network to model various targets with different shapes and orientations and capture rich context information. With minor code modifications and negligible extra computational cost, our QFormer outperforms existing representative (hierarchical and plain) vision transformers on various vision tasks, including classification, object detection, semantic segmentation, and pose estimation.

Method

Fig.1 - The comparison of the current design (hand-crafted windows) and Quadrange attention.

Fig.2 - The pipeline of our proposed quadrangle attention (QA).

Fig.3 - The transformation process in quadrangle attention.

Fig.4 - The architecture of our plain QFormer_p (a) and hierarchical QFormer_h (b).

Usage

Requirements

• PyTorch==1.7.1
• torchvision==0.8.2
• timm==0.3.2

The Apex is optional for faster training speed.

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" ./

Other Requirements

pip install opencv-python==4.4.0.46 termcolor==1.1.0 yacs==0.1.8 timm==0.4.9
pip install einops

Train & Eval

For classification on ImageNet-1K, to train from scratch, run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch \
  --nnodes ${NNODES} \
  --node_rank ${SLURM_NODEID} \
  --master_addr ${MHOST} \
  --master_port 25901 \
  --nproc_per_node 8 \
  ./main.py \
  --cfg configs/swin/qformer_tiny_patch4_window7_224.yaml \
  --data-path ${IMAGE_PATH} \
  --batch-size 128 \
  --tag 1024-dpr20-coords_lambda1e-1 \
  --distributed \
  --coords_lambda 1e-1 \
  --drop_path_rate 0.2 \

For single GPU training, run

python ./main.py \
  --cfg configs/swin/qformer_tiny_patch4_window7_224.yaml \
  --data-path ${IMAGE_PATH} \
  --batch-size 128 \
  --tag 1024-dpr20-coords_lambda1e-1 \
  --coords_lambda 1e-1 \
  --drop_path_rate 0.2 \

To evaluate, run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch \
  --nnodes ${NNODES} \
  --node_rank ${SLURM_NODEID} \
  --master_addr ${MHOST} \
  --master_port 25901 \
  --nproc_per_node 8 \
  ./main.py \
  --cfg configs/swin/qformer_tiny_patch4_window7_224.yaml \
  --data-path ${IMAGE_PATH} \
  --batch-size 128 \
  --tag eval \
  --distributed \
  --resume ${MODEL PATH} \
  --eval

For single GPU evaluation, run

python ./main.py \
  --cfg configs/swin/qformer_tiny_patch4_window7_224.yaml \
  --data-path ${IMAGE_PATH} \
  --batch-size 128 \
  --tag eval \
  --resume ${MODEL PATH} \
  --eval

Results

Results on plain models

Classification results on ImageNet-1K with MAE pretrained models

model	resolution	acc@1	Weights & Logs
ViT-B + Window attn	224x224	81.2	\
ViT-B + Shifted window	224x224	82.0	\
QFormerp-B	224x224	82.9	Coming soon

Detection results on COCO with MAE pretrained models and the Mask RCNN detector, following ViTDet

model	box mAP	mask mAP	Params	Weights & Logs
ViTDet-B	51.6	45.9	111M	\
QFormerp-B	52.3	46.6	111M	Coming soon

Semantic segmentation results on ADE20k with MAE pretrained models and the UPerNet segmentor

model	image size	mIoU	mIoU*	Weights & Logs
ViT-B + window attn	512x512	39.7	41.8	\
ViT-B + shifted window attn	512x512	41.6	43.6	\
QFormerp-B	512x512	43.6	45.0	Coming soon
ViT-B + window attn	640x640	40.2	41.5	\
ViT-B + shifted window attn	640x640	42.3	43.5	\
QFormerp-B	640x640	44.9	46.0	Coming soon

Human pose estimation results on COCO with MAE pretrained models, following ViTPose

attention	model	AP	AP50	AR	AR50	Weights & Logs
Window	ViT-B	66.4	87.7	72.9	91.9	\
Shifted window	ViT-B	76.4	90.9	81.6	94.5	\
Quadrangle	ViT-B	77.0	90.9	82.0	94.7	Coming soon
Window + Full	ViT-B	76.9	90.8	82.1	94.7	\
Shifted window + Full	ViT-B	77.2	90.9	82.2	94.7	\
Quadrangle + Full	ViT-B	77.4	91.0	82.4	94.9	Coming soon

Results on hierarchical models

Main Results on ImageNet-1K

name	resolution	acc@1	acc@5	acc@RealTop-1	Weights & Logs
Swin-T	224x224	81.2	\	\	\
DW-T	224x224	82.0	\	\	\
Focal-T	224x224	82.2	95.9
QFormerh-T	224x224	82.5	96.2	87.5	model & logs
Swin-S	224x224	83.2	96.2	\	\
Focal-S	224x224	83.5	96.2	\	\
QFormerh-S	224x224	84.0	96.8	88.6	model & logs
Swin-B	224x224	83.4	96.5	\	\
DW-B	224x224	83.4	\	\	\
Focal-B	224x224	83.8	96.5	\	\
QFormerh-B	224x224	84.1	96.8	88.7	model & logs

Object Detection Results

Mask R-CNN

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	config	log	model
Swin-T	ImageNet-1K	1x	43.7	39.8	48M	\	\	\
DAT-T	ImageNet-1K	1x	44.4	40.4	48M	\	\	\
Focal-T	ImageNet-1K	1x	44.8	41.0	49M	\	\	\
QFormerh-T	ImageNet-1K	1x	45.9	41.5	49M	config	log	onedrive
Swin-T	ImageNet-1K	3x	46.0	41.6	48M	\	\	\
DW-T	ImageNet-1K	3x	46.7	42.4	49M	\	\	\
DAT-T	ImageNet-1K	3x	47.1	42.4	48M	\	\	\
DAT-T	ImageNet-1K	3x	47.1	42.4	48M	\	\	\
QFormerh-T	ImageNet-1K	3x	47.5	42.7	49M	config	log	onedrive
Swin-S	ImageNet-1K	3x	48.5	43.3	69M	\	\	\
Focal-S	ImageNet-1K	3x	48.8	43.8	71M	\	\	\
DAT-S	ImageNet-1K	3x	49.0	44.0	69M	\	\	\
QFormerh-S	ImageNet-1K	3x	49.5	44.2	70M	config	log	onedrive

Cascade Mask R-CNN

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	config	log	model
Swin-T	ImageNet-1K	1x	48.1	41.7	86M	\	\	\
DAT-T	ImageNet-1K	1x	49.1	42.5	86M	\	\	\
QFormerh-T	ImageNet-1K	1x	49.8	43.0	87M	config	log	onedrive
Swin-T	ImageNet-1K	3x	50.2	43.7	86M	\	\	\
QFormerh-T	ImageNet-1K	3x	51.4	44.7	87M	config	log	onedrive
Swin-S	ImageNet-1K	3x	51.9	45.0	107M	\	\	\
QFormerh-S	ImageNet-1K	3x	52.8	45.7	108M	config	log	onedrive

Semantic Segmentation Results for ADE20k

UperNet

Backbone	Pretrain	Lr Schd	mIoU	mIoU*	#params	config	log	model
Swin-T	ImageNet-1k	160k	44.5	45.8	60M	\	\	\
DAT-T	ImageNet-1k	160k	45.5	46.4	60M	\	\	\
DW-T	ImageNet-1k	160k	45.7	46.9	61M	\	\	\
Focal-T	ImageNet-1k	160k	45.8	47.0	62M	\	\	\
QFormerh-T	ImageNet-1k	160k	46.9	48.1	61M	Coming soon	Coming soon	Coming soon
Swin-S	ImageNet-1k	160k	47.6	49.5	81M	\	\	\
DAT-S	ImageNet-1k	160k	48.3	49.8	81M	\	\	\
Focal-S	ImageNet-1k	160k	48.0	50.0	61M	\	\	\
QFormerh-S	ImageNet-1k	160k	48.9	50.3	82M	Coming soon	Coming soon	Coming soon
Swin-B	ImageNet-1k	160k	48.1	49.7	121M	\	\	\
DW-B	ImageNet-1k	160k	48.7	50.3	125M	\	\	\
Focal-B	ImageNet-1k	160k	49.0	50.5	126M	\	\	\
QFormerh-B	ImageNet-1k	160k	49.5	50.6	123M	Coming soon	Coming soon	Coming soon

Statement

This project is for research purpose only. For any other questions please contact qmzhangzz at hotmail.com.

The code base is borrowed from Swin.

Citing QFormer, VSA and ViTAE

@article{zhang2023vision,
  title={Vision Transformer with Quadrangle Attention},
  author={Zhang, Qiming and Zhang, Jing and Xu, Yufei and Tao, Dacheng},
  journal={arXiv preprint arXiv:2303.15105},
  year={2023}
}
@inproceedings{zhang2022vsa,
  title={VSA: learning varied-size window attention in vision transformers},
  author={Zhang, Qiming and Xu, Yufei and Zhang, Jing and Tao, Dacheng},
  booktitle={Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXV},
  pages={466--483},
  year={2022},
  organization={Springer}
}
@article{zhang2023vitaev2,
  title={Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond},
  author={Zhang, Qiming and Xu, Yufei and Zhang, Jing and Tao, Dacheng},
  journal={International Journal of Computer Vision},
  pages={1--22},
  year={2023},
  publisher={Springer}
}
@article{xu2021vitae,
  title={Vitae: Vision transformer advanced by exploring intrinsic inductive bias},
  author={Xu, Yufei and Zhang, Qiming and Zhang, Jing and Tao, Dacheng},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

Our other Transformer works

ViTPose: Please see Baseline model ViTPose for human pose estimation;

VSA: Please see ViTAE-Transformer for Image Classification and Object Detection;

ViTAE & ViTAEv2: Please see ViTAE-Transformer for Image Classification, Object Detection, and Sementic Segmentation;

Matting: Please see ViTAE-Transformer for matting;

Remote Sensing: Please see ViTAE-Transformer for Remote Sensing; Advancing Plain Vision Transformer Towards Remote Sensing Foundation Model ;