Skip to content

OpenCompatible provides a standard compatible training benchmark, covering practical training scenarios.

License

Notifications You must be signed in to change notification settings

TencentARC/OpenCompatible

Repository files navigation

Open Compatible Training Benchmark

OpenCompatible provides a standard compatible training benchmark, covering practical training scenarios.

Mainstreaming paradigms are easily achieved using our benchmark, which includes:

  • backfilling-free backward compatibility
  • hot-refresh backward compatibility
  • feature-level forward compatibility (onging)

Various downstream tasks are supported:

  • Content-based Image Retrieval (Google Landmark, Revisited Oxford, and Revisited Paris)
  • Face Recognition (MS1M-v3 and IJB-C)
  • Person Re-ID (onging)

Published papers

Requirements

  • Python >= 3.6
  • PyTorch >= 1.6
  • tensorflow >= 2.1
  • termcolor
  • sklearn
  • faiss-gpu
  • numpy
  • tqdm

Dataset Preparation

Train and Test

Model Zoo

Pre-trained models are provided (see model_zoo.md )

License

This project is licensed under the Apache v2 License.

More details are in LICENSE.

Acknowledgements

This project is inspired by the project Pytorch-Template, and OpenBCT.

Contributors: Binjie Zhang, Yixiao Ge, and Shupeng Su.

References:

[1] Towards backward-compatible representation learning. CVPR 2020.
[2] Learning without Forgetting. T-PAMI 2017.
[3] Positive-congruent training: Towards regression-free model updates. CVPR 2021.

Citatations

@inproceedings{zhang2021hot,
  title={Hot-Refresh Model Upgrades with Regression-Free Compatible Training in Image Retrieval},
  author={Zhang, Binjie and Ge, Yixiao and Shen, Yantao and Li, Yu and Yuan, Chun and Xu, Xuyuan and Wang, Yexin and Shan, Ying},
  booktitle={International Conference on Learning Representations},
  year={2021}
}
@article{zhang2022towards,
  title={Towards Universal Backward-Compatible Representation Learning},
  author={Zhang, Binjie and Ge, Yixiao and Shen, Yantao and Su, Shupeng and Yuan, Chun and Xu, Xuyuan and Wang, Yexin and Shan, Ying},
  journal={arXiv preprint arXiv:2203.01583},
  year={2022}
}

Contact

Binjie Zhang (homepage): zbj19@tsinghua.org.cn, Yixiao Ge (homepage): yixiaoge@tencent.com.

About

OpenCompatible provides a standard compatible training benchmark, covering practical training scenarios.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages