ExCon: Explanation-driven Supervised Contrastive Learning

Link to the paper: https://arxiv.org/pdf/2111.14271.pdf

Contributors of this repo:

• Zhibo Zhang ([email protected])
• Jongseong Jang ([email protected])
• Ruiwen Li ([email protected])

Copyright (c) 2021 LG AI Research and University of Toronto, all rights reserved.

If you use our code, please cite our paper:

@misc{zhang2021excon,
      title={ExCon: Explanation-driven Supervised Contrastive Learning for Image Classification},
      author={Zhibo Zhang and Jongseong Jang and Chiheb Trabelsi and Ruiwen Li and Scott Sanner and Yeonjeong Jeong and Dongsub Shim},
      year={2021},
      eprint={2111.14271},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Prepare the Tiny ImageNet dataset (in the path where you want to save the dataset):

wget -nc https://image-net.org/data/tiny-imagenet-200.zip
unzip tiny-imagenet-200.zip
python3 ExCon/utils/val_format.py

Run ExCon:

CIFAR-100 dataset

python3 ExCon/main_supcon.py --epochs=200 --explainer="GradCAM" --dataset="cifar100" --batch_size=256 --method="Ex_SupCon" --learning_rate=0.5 --temp=0.1 --cosine --negative_pair=1 --validation=0 --background_anchor=0 --exp_epochs=50

Tiny ImageNet dataset

python3 ExCon/main_supcon.py --epochs=200 --explainer="GradCAM" --dataset="ImageNet" --batch_size=128 --method="Ex_SupCon" --learning_rate=0.5 --temp=0.1 --cosine --negative_pair=1 --validation=0 --background_anchor=0 --exp_epochs=0 --data_folder=$PATH_TO_DATASET

Reference Repos:

[1] https://github.com/HobbitLong/SupContrast