continental/guided-concept-projection-vectors

The implementation of Guided Concept Projection Vectors (GCPV) framework for concept-based explainability of computer vision models.

Installation

1. Download repo

git clone https://github.com/continental/guided-concept-projection-vectors.git

2. Create & activate venv (optionally)

We used Python 3.9.17

python -m venv gcpv_venv
source ./gcpv_venv/bin/activate

3. Install requirements

pip install -r requirements.txt

Demo

Download MS COCO 2017 annotations + validation subset (240 MB + 780 MB)

Execute: ./data/download_ms_coco_2017val_dataset.sh

Try demo Jupyter Notebooks files

1. Optimize GCPVs for a single sample: ./demo/gcpv_optimization.ipynb
2. Optimize & cluster several GCPVs + weak concept localization: ./demo/gcpv_clustering.ipynb
3. Find subconcepts with GCPVs + weak sub-concept localization: ./demo/gcpv_clustering_subconcepts.ipynb

Reference

ArXiv.org:

@article{mikriukov2023gcpv,
  title={GCPV: Guided Concept Projection Vectors for the Explainable Inspection of CNN Feature Spaces},
  author={Mikriukov, Georgii and Schwalbe, Gesina and Hellert, Christian and Bade, Korinna},
  journal={arXiv preprint arXiv:2311.14435},
  year={2023}
}

Documentation

For further help, see the API-documentation or contact the maintainers.

License

Copyright (C) 2024 co-pace GmbH (a subsidiary of Continental AG). All rights reserved.