Official code for the paper AdaFocal: Calibration-aware Adaptive Focal Loss
Authors: Arindam Ghosh, Thomas Schaaf, and Matt Gormley
URL: https://proceedings.neurips.cc/paper_files/paper/2022/hash/0a692a24dbc744fca340b9ba33bc6522-Abstract-Conference.html
Arxiv: https://arxiv.org/abs/2211.11838
The code provides the bare minimum to reproduce the calibration related results obtained from training a ResNet-50 model on CIFAR-10 dataset using Adafocal loss.
Most of the starter code for training, evaluation and calculating calibration related metrics is borrowed from https://github.com/torrvision/focal_calibration.
To train Resnet-50 on CIFAR-10 with default settings for Adafocal, run:
python main.py --dataset cifar10 --model resnet50 --loss adafocal -e 350 --save-path exp/cifar10_resnet50_adafocal
If the code or the paper has been useful in your research, please add the citation:
@inproceedings{NEURIPS2022_0a692a24,
author = {Ghosh, Arindam and Schaaf, Thomas and Gormley, Matthew},
booktitle = {Advances in Neural Information Processing Systems},
pages = {1583--1595},
title = {AdaFocal: Calibration-aware Adaptive Focal Loss},
url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/0a692a24dbc744fca340b9ba33bc6522-Paper-Conference.pdf},
volume = {35},
year = {2022}
}