TencentARC / OpenCompatible

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

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

• [ICLR 2022] B. Zhang, Y. Ge, Y. Shen, Y. Li, C. Yuan, X. Xu, Y. Wang, and Y. Shan. Hot-Refresh Model Upgrades with Regression-Free Compatible Training in Image Retrieval[C]. [video] [slides]
• [IJCAI 2022 (oral)] B. Zhang, Y. Ge, Y. Shen, S. Su, F. Wu, C. Yuan, X. Xu, Y. Wang, and Y. Shan. Towards Universal Backward-Compatible Representation Learning[J]. video slides
• [arXiv] Su S^*, Zhang B^*, Ge Y, et al. Privacy-Preserving Model Upgrades with Bidirectional Compatible Training in Image Retrieval[J].

Requirements

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

Dataset Preparation

• Please refer to datasets.md for more details.

Train and Test

• Please refer to train_test.md for more details.

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): [e-mail], Yixiao Ge (homepage): [e-mail].