Recalibrating Conformal Prediction

Test-time recalibration of conformal predictors based on unlabeled data for improved performance under test distribution shift.

Requirements

The following Python libraries are required to run the code in this repository:

torch
torchvision

Other requirements can be installed with pip install -r requirements.txt. The above two libraries are not included in requirements.txt for safe measure.

Usage

All the figures in the paper can be reproduced by running this notebook

Citation

@article{yilmazheckel2022RecalibratingConformal,
    author    = {Fatih Furkan Yilmaz and Reinhard Heckel},
    title     = {Test-time Recalibration of Conformal Predictors Under Distribution Shift Based on Unlabeled Examples},
    journal   = {arXiv:2210.04166},
    year      = {2022}
}

Licence

All files are provided under the terms of the Apache License, Version 2.0.