This is the repo for the paper Beyond Pinball Loss: Quantile Methods for Calibrated Uncertainty Quantification.
The core algorithm aims to learn the conditional quantiles of the data distribution by optimizing directly for calibration, sharpness, and adversarial group calibration.
python main.py --loss scaled_batch_cal
There are currently 4 losses implemented: "calibration loss",
"scaled calibration loss",
"interval score", and "qr", which is simply the pinball loss.
The versions of the losses with the prefix "batch" means the code for calculating the loss is
vectorized.
If you find this work useful, please consider citing the corresponding paper:
@article{chung2020beyond,
title={Beyond Pinball Loss: Quantile Methods for Calibrated Uncertainty Quantification},
author={Chung, Youngseog and Neiswanger, Willie and Char, Ian and Schneider, Jeff},
journal={arXiv preprint arXiv:2011.09588},
year={2020}
}
This work led to the development of a general uncertainty quantification repo, called Uncertainty-Toolbox. This toolbox evaluates a quantification of uncertainty based on a suite of metrics, including calibration, sharpness, adversarial group calibration, proper scoring rules, and accuracy.