yueyang1996/labo

CVPR 2023: Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification

LaBo

Code for the CVPR 2023 paper "Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification"

Set up environments

We run our experiments using Python 3.9.13. You can install the required packages using:

conda create --name labo python=3.9.13
conda activate labo
pip install -r requirements.txt

You need to modify the source code of Apricot to run the submodular optimization. See details here.

Directories

• cfg/ saves the config files for all experiments, including linear probe (cfg/linear_probe) and LaBo (cfg/asso_opt). You can modify the config files to change the system arguments.

• datasets/ stores the dataset-specific data, including images, splits, and concepts. Please check datasets/DATASET.md for details.

  Note: the images of each dataset are not provided in this repo; you need to download them and store in the corresponding folder: datasets/{dataset name}/images/. Check datasets/DATASET.md for instructions on downloading all datasets.

• exp/ is the work directories of the experiments. The config files and model checkpoints will be saved in this folder.

• models/ saves the models:

  • Linear Probe: models/linear_prob/linear_prob.py
  • LaBo: models/asso_opt/asso_opt.py
  • concept selection functions: models/select_concept/select_algo.py

• output/: the performance will be saved into .txt files stored in output/.

• Other files:

  • data.py and data_lp.py are the dataloaders for LaBo and Linear Probe, respectively.
  • main.py is the interface to run all experiments, and utils.py contains the preprocess and feature extraction functions.
  • linear probe.sh is the bash file to run the linear probe. labo_train.sh and alabo_test.sh are the bash file to train and test LaBo.

Linear Probe

To get the linear probe performance, just run:

sh linear_probe.sh {DATASET} {SHOTS} {CLIP SIZE}

For example, for flower dataset 1-shot with ViT-L/14 image encoder, the command is:

sh linear_probe.sh flower 1 ViT-L/14

The code will automatically encode the images and run a hyperparameter search on the L2 regularization using the dev set. The best validation and test performance will be saved in the output/linear_probe/{DATASET}.txt.

LaBo Training

To train the LaBo, run the following command:

sh labo_train.sh {SHOTS} {DATASET}

The training logs will be uploaded to the wandb. You may need to set up your wandb account locally. After reaching the maximum epochs, the checkpoint with the highest validation accuracy and the corresponding config file will be saved to exp/asso_opt/{DATASET}/{DATASET}_{SHOT}shot_fac/.

LaBo Testing

To get the test performance, use the model checkpoint and corresponding configs saved in exp/asso_opt/{DATASET}/{DATASET}_{SHOT}shot_fac/ and run:

sh labo_test.sh {CONFIG_PATH} {CHECKPOINT_PATH}

The test accuracy will be printed to output/asso_opt/{DATASET}.txt.

Please cite our paper if you find it useful!

@inproceedings{yang2023language,
  title={Language in a bottle: Language model guided concept bottlenecks for interpretable image classification},
  author={Yang, Yue and Panagopoulou, Artemis and Zhou, Shenghao and Jin, Daniel and Callison-Burch, Chris and Yatskar, Mark},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19187--19197},
  year={2023}
}