Augmentation by Counterfactual Explanation - Fixing an Overconfident Classifier

This is the GitHub repo for the WACV'23 paper: "Augmentation by Counterfactual Explanation - Fixing an Overconfident Classifier" (https://arxiv.org/abs/2210.12196)

-- by Sumedha Singla, Nihal Murali, Forough Arabshahi, Sofia Triantafyllou, Kayhan Batmanghelich

To cite our work, please use the following bibtex entry:

@inproceedings{singla2023augmentation,
  title={Augmentation by Counterfactual Explanation-Fixing an Overconfident Classifier},
  author={Singla, Sumedha and Murali, Nihal and Arabshahi, Forough and Triantafyllou, Sofia and Batmanghelich, Kayhan},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={4720--4730},
  year={2023}
}

Example Use-Cases:

general syntax: python <code_path> --config <config_path>

(1) Train a DenseNet classifier on the AFHQ dataset:

python ./Train_Classifier_DenseNet.py --config ./Configs/Classifier/DenseNet_CelebA.yaml

(2) Train a StyleGANv2 explainer on the CelebA dataset:

python ./Explainer_StyleGANv2/Train_Explainer_StyleGANv2.py --config ./Configs/Explainer/styleGAN_CelebA.yaml

Once the classifier and explainer are trained, refer to ./Misc/retrain_clf_via_ace.ipynb jupyter notebook for re-training the classifier using the proposed ACE method.