Tangent Model Composition (ICCV 2023, ICLR 2024)

Official code repository for

• Tangent Model Composition for Ensembling and Continual Fine-tuning (ICCV 2023)
• Tangent Transformers for Composition, Privacy and Removal (ICLR 2024)

Requirements

Our repository is based on PyTorch. We use Torch 1.12 and Python 3.9, other versions have not been tested.

In addition, the following packages are also needed:

pip install hydra-core==1.2.0

Datasets

Create a folder for storing datasets in the main directory

mkdir data

We provide example scripts for setting up MIT-67 and Oxford Pets in the setup directory

bash setup/setup_mit.sh
bash setup/setup_oxfordpets.sh

Reproducing results

Our results for the Class Incremental (Class-IL) setting and Data Incremental (Data-IL) can be reproduced using

bash scripts/compose.sh

and changing the variables appropriately.

For composition tasks on Tangent Transformers, an example script can be found in

bash scripts/compose_vit.sh

which can be adapted to one's needs. To obtain the best hyperparameters for each dataset, please refer to Appendix A of the original paper

If you find this useful for your work, please consider citing

@inproceedings{liu2023tangent,
  title={Tangent Model Composition for Ensembling and Continual Fine-tuning},
  author={Liu, Tian Yu and Soatto, Stefano},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={18676--18686},
  year={2023}
}

@inproceedings{liu2024tangent,
  title={Tangent Transformers for Composition, Privacy and Removal},
  author={Liu, Tian Yu and Golatkar, Aditya and Soatto, Stefano},
  journal={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://arxiv.org/abs/2307.08122}
}