LzVv123456/Contrastive-Prototypical-Prompt

Steering Prototypes with Prompt-tuning for Rehearsal-free Continual Learning

1. Introduction

This repository contains the official source code for the research paper titled "Steering Prototypes with Prompt-tuning for Rehearsal-free Continual Learning". Please refer to the paper https://arxiv.org/abs/2303.09447 for detailed methods

2. Setting Up the Environment

To ensure reproducibility and smooth execution of the code, we recommend setting up a dedicated environment using conda.

Steps:

1. First, make sure you have Anaconda or Miniconda installed.

2. Create a new conda environment:

   conda create --name your_env_name python=3.9

3. Activate the environment:

   conda activate your_env_name

4. Install the required packages:

   pip install -r requirements.txt

3. Preparing and Downloading Datasets

1. For CIFAR-100 and 5datasets, they should download automatically when running the code.
2. For ImageNet-R, download and unzip the original data from https://people.eecs.berkeley.edu/~hendrycks/imagenet-r.tar. Please also download the data split from https://drive.google.com/drive/folders/1D5ADrPs9OweevMNA-ZjuAGdbilYfA1io?usp=sharing.
3. For ImageNet-Sub, download and unzip from https://drive.google.com/file/d/1n5Xg7Iye_wkzVKc0MTBao5adhYSUlMCL/view?usp=sharing
4. Save all dowmloaded contents in the data/ folder.

4. How to Run the Code

1. Navigate to the script directory:

   cd script/

2. Run the bash scripts:

   bash cifar100.sh   # for CIFAR-100

   bash imagenet.sh   # for ImageNet-R

   bash imagenet_sub.sh   # for ImageNet-Sub

   bash 5datasets.sh   # for 5datasets

3. How to run other settings:

   Please refer to the "main.py" for the detailed arguments.

5. How to Cite the Paper

If you find our research or this repository useful, please consider citing our work:

@misc{li2023steering,
      title={Steering Prototypes with Prompt-tuning for Rehearsal-free Continual Learning}, 
      author={Zhuowei Li and Long Zhao and Zizhao Zhang and Han Zhang and Di Liu and Ting Liu and Dimitris N. Metaxas},
      year={2023},
      eprint={2303.09447},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

6. Acknowledgments

We would like to extend our gratitude to dino, mae and SupContrast which we use partial of their code in our project.