This repo is the official PyTorch implementation of CVPR2025 paper "Language Guided Concept Bottleneck Models for Interpretable Continual Learning".

We follow the framework of Pilot to implement this project.

conda create -n <env name> python=3.8.10 -y
conda activate <env name>
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.7 -c pytorch -c nvidia --no-deps
# other dependencies
pip install -r requirements.txt
pip install git+https://github.com/openai/CLIP.git

We evaluated our approch on 3 coarse-grained datasets: CIFAR-100, TinyImageNet, ImageNet100; and 4 fine-grained datasets: CUB-200, Flower102, Food101, Stanford-cars.

You should specify your data folder in utils/data.py:

def download_data(self):
    assert 0,"You should specify the folder of your dataset"
    train_dir = '[DATA-PATH]/train/'
    test_dir = '[DATA-PATH]/val/'

or

def download_data(self):
    assert 0,"You should specify the folder of your dataset"
    root="/[DATA_PATH]/"
    ....

We provide the concepts that are utilized in our experiments in data folder.

The configuration of hyper-parameters of our method is provided in exps folder. You can run an experiment follows:

python main.py --config=<config json path>

You can modify the configuration file directly to specialize the training process.