Awesome

Generative Multi-modal Models are Good Class Incremental Learners

This is the official code for our CVPR paper: <a href='https://arxiv.org/abs/2403.18383.pdf'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a>

Getting Started

Installation

Git clone our repository, creating a python environment and activate it via the following command

git clone https://github.com/DoubleClass/GMM
cd GMM
conda env create -f env_GMM.yaml
conda activate GMM
pip install git+https://github.com/openai/CLIP.git

Vicuna

You can get the LLM Vicuna in huggingface-vicuna-7b, or you can Dowload it via baidunetdisk. (code: s3pu) or google-drive

Then set the downloaded vicuna folder path here and the initial checkpoint here

EVA_VIT_G

The code will automatically downloading the eva_vit_g.pth, we alse put it here or huggingface, you can manually download it and put it in the cache dir: .cache/torch/hub/checkpoints

bert-base-uncased

The code will automatically downloading this, but in case you don't have access to huggingface, we also put it here , you can manually download it and alse put it in cache dir: .cache/huggingface/hub/models--bert-base-uncased

datasets

ImageNet-R

You can download it here

Then set the dataset folder path here

Besides, you need to customize the dataset for the GPT fine-tuning process. We prepare a example here you can follow: download or google-dirve

After downloaded the customized dataset, you can set the data root path here.

Training

After setting all model and dataset config, you can run the following command to start fine-tuning.

python train.py --cfg-path train_configs/minigpt4_stage2_finetune.yaml

Testing

After training, you will get a model checkpoint of the last continual learning stage. put the path to scipts in eval_all.sh and specify a results directory.

Then set the results path in the get_score_all.py

Run the script:

bash eval_all.sh

Citation

If you find our repo useful for your research, please consider citing our paper:

@article{cao2024GMM,
  title={Generative Multi-modal Models are Good Class Incremental Learners},
  author={Cao, Xusheng and Lu, Haori and Huang, Linlan and Liu, Xialei and Cheng, Ming-Ming},
  journal={IEEE Computer Vision and Pattern Recognition (CVPR)},
  year={2024}
}

License

This code is licensed under the Creative Commons Attribution-NonCommercial 4.0 International for non-commercial use only. Please note that any commercial use of this code requires formal permission prior to use.

Contact

For technical questions, please contact <a href="caoxusheng@mail.nankai.edu.cn">caoxusheng@mail.nankai.edu.cn</a>