This is the official code for our paper :

The environment is the same as that of our RAPF.

create enviroment using Miniconda (or Anaconda)

conda create -n continual_clip python=3.8
conda activate continual_clip

install dependencies:

bash setup_environment.sh

We provide the scripts for imagenet100. Please run:

python main.py \
    --config-path configs/class \
    --config-name imagenet100_10-10.yaml \
    dataset_root="[imagenet1k_path]" \
    class_order="class_orders/imagenet100.yaml"

Note: To obtain the epoch parameter from the first task described in Eq. (3), please run the epoch.py file.

The dataset_root folder should contain the train and val folders.

imagenet1k_path
├── train
│   ├── n01440764 
│   └── ···
├── val
│   ├── n01440764 
│   └── ···

imagenet-r_path
├── train
│   ├── n01443537 
│   └── ···
├── val
│   ├── n01443537 
│   └── ···

The command to run the other two datasets is similar, in run_experiment.sh

Cifar100 will download automatically. Imagenet-R is randomly splited. You can also use our splited list in RAPF/imgr_split/imgr_train_test_split.txt.

The format of imgr_train_test_split.txt:

train
n02051845/art_0.jpg
...
test
n02051845/tattoo_4.jpg
...

Our method implementation is based on the Continual-CLIP.

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

@inproceedings{huang2025mind,
  title={Mind the gap: Preserving and compensating for the modality gap in clip-based continual learning},
  author={Huang, Linlan and Cao, Xusheng and Lu, Haori and Meng, Yifan and Yang, Fei and Liu, Xialei},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={3777--3786},
  year={2025}
}

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.

For technical questions, please contact [email protected]