Regression-based-Analytic-Incremental-Learning

Official implementation of Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models

The paper has accepted by NeurIPS 2024.

Installation

Create a conda environment and install dependencies:

git clone https://github.com/linghan1997/Regression-based-Analytic-Incremental-Learning.git
cd RAIL

conda create -n rail python=3.8
conda activate tip_adapter

pip install -r requirements.txt

Data Preparation

We suggest putting all required datasets under the folder

RAIL/
|-- datasets/

Please refer to the following guides for setting up datasets: CoOp

Running

You may set the dataset sequence and other hyper-parameters in the config file analytic_clip.yaml

We have developed two forms of the RAIL method. For the primal RAIL:

python primal_RAIL.py

For the dual RAIL:

python dual_RAIL.py

Citation

@article{xu2024advancing,
  title={Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models},
  author={Xu, Yicheng and Chen, Yuxin and Nie, Jiahao and Wang, Yusong and Zhuang, Huiping and Okumura, Manabu},
  journal={arXiv preprint arXiv:2406.18868},
  year={2024}
}

---

Acknowledgement

Our repo benefits from CLIP and CoOp. We thank them for their wonderful works.

---