This repository contains the code for the paper "Revisiting Reweighted Risk for Calibration: AURC, Focal, and Inverse Focal Loss" [arxiv].

To run the code, you will need the following dependencies (excluding common packages like scipy, numpy, and torch):

• Python ≥ 3.8

• Tiny-ImageNet [link]

To download and preprocess the dataset, use the following commands:

cd data
python tiny_imagenet_utils.py

• CIFAR-10/100

To train the model with select AU loss, you can copy the file loss/select_au.py into your repository.

To train the model with select AU loss, use the following commands:

python src/train.py --arch vit_small --dataset tiny-imagenet --loss_type select_au --seed 40 --workers 1

If you found this work or code useful, please cite:

@misc{zhou2025revisitingreweightedriskcalibration,
      title={Revisiting Reweighted Risk for Calibration: AURC, Focal, 
and Inverse Focal Loss}, 
      author={Han Zhou and Sebastian G. Gruber and Teodora Popordanoska 
and Matthew B. Blaschko},
      year={2025},
      eprint={2505.23463},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.23463}, 
}

This project is licensed under the MIT License.