Updates

(2022/05/10) Code of ResTV2 is released! ResTv2 simplifies the EMSA structure in ResTv1 (i.e., eliminating the multi-head interaction part) and employs an upsample operation to reconstruct the lost medium- and high-frequency information caused by the downsampling operation.

ResT: An Efficient Transformer for Visual Recognition

Official PyTorch implementation of ResTv1 and ResTv2, from the following paper:

ResT: An Efficient Transformer for Visual Recognition. NeurIPS 2021.
ResT V2: Simpler, Faster and Stronger. NeurIPS 2022.
By Qing-Long Zhang and Yu-Bin Yang
State Key Laboratory for Novel Software Technology at Nanjing University

ResTv1 is initially described in arxiv, which capably serves as a general-purpose backbone for computer vision. It can tackle input images with arbitrary size. Besides, ResT compressed the memory of standard MSA and model the interaction between multi-heads while keeping the diversity ability.

Catalog

ImageNet-1K Training Code
ImageNet-1K Fine-tuning Code
Downstream Transfer (Detection, Segmentation) Code

Results and Pre-trained Models

ImageNet-1K trained models

name	resolution	acc@1	#params	FLOPs	Throughput	model
ResTv1-Lite	224x224	77.2	11M	1.4G	1246	baidu
ResTv1-S	224x224	79.6	14M	1.9G	1043	baidu
ResTv1-B	224x224	81.6	30M	4.3G	673	baidu
ResTv1-L	224x224	83.6	52M	7.9G	429	baidu
ResTv2-T	224x224	82.3	30M	4.1G	826	baidu
ResTv2-T	384x384	83.7	30M	12.7G	319	baidu
ResTv2-S	224x224	83.2	41M	6.0G	687	baidu
ResTv2-S	384x384	84.5	41M	18.4G	256	baidu
ResTv2-B	224x224	83.7	56M	7.9G	582	baidu
ResTv2-B	384x384	85.1	56M	24.3G	210	baidu
ResTv2-L	224x224	84.2	87M	13.8G	415	baidu
ResTv2-L	384x384	85.4	87M	42.4G	141	baidu

Note: Access code for baidu is rest. Pretrained models of ResTv1 is now available in google drive.

Installation

Please check INSTALL.md for installation instructions.

Evaluation

We give an example evaluation command for a ImageNet-1K pre-trained, then ImageNet-1K fine-tuned ResTv2-T:

Single-GPU

python main.py --model restv2_tiny --eval true \
--resume restv2_tiny_384.pth \
--input_size 384 --drop_path 0.1 \
--data_path /path/to/imagenet-1k

This should give

* Acc@1 83.708 Acc@5 96.524 loss 0.777

For evaluating other model variants, change --model, --resume, --input_size accordingly. You can get the url to pre-trained models from the tables above.
Setting model-specific --drop_path is not strictly required in evaluation, as the DropPath module in timm behaves the same during evaluation; but it is required in training. See TRAINING.md or our paper for the values used for different models.

Training

See TRAINING.md for training and fine-tuning instructions.

Acknowledgement

This repository is built using the timm library.

License

This project is released under the Apache License 2.0. Please see the LICENSE file for more information.

Citation

If you find this repository helpful, please consider citing:

ResTv1

@inproceedings{zhang2021rest,
  title={ResT: An Efficient Transformer for Visual Recognition},
  author={Qinglong Zhang and Yu-bin Yang},
  booktitle={Advances in Neural Information Processing Systems},
  year={2021},
  url={https://openreview.net/forum?id=6Ab68Ip4Mu}
}

ResTv2

@article{zhang2022rest,
  title={ResT V2: Simpler, Faster and Stronger},
  author={Zhang, Qing-Long and Yang, Yu-Bin},
  journal={arXiv preprint arXiv:2204.07366},
  year={2022}

Third-party Implementation

[2022/05/26] ResT and ResT v2 have been integrated into PaddleViT, checkout here for the 3rd party implementation on Paddle framework!

wofmanaf / ResT