Skip to content

BorealisAI/towards-better-sel-cls

Repository files navigation

Towards Better Selective Classification

This is the official implementation of the paper Towards Better Selective Classification.

In this work, we confirm that the superior performance of state-of-the-art methods such as SelectiveNet, Deep Gamblers, and Self-Adaptive Training is owed to training a more generalizable classifier rather than their proposed selection mechanisms. We propose an entropy-based regularizer that improves the performance and achieves new state-of-the-art results.

Install

Create and activate a conda environment. Install the dependencies as listed in requirements.txt:

conda create --name sel_cls python=3.7
conda activate sel_cls
pip install -r requirements.txt

Training and Evaluation

Self-Adaptive Training (SAT):

bash run_${dataset}.sh

Self-Adaptive Training (SAT) + Entropy Minimization (EM):

bash run_${dataset}_entropy.sh

Reference

For technical details, please check the conference version of our paper.

@inproceedings{
    feng2023towards,
    title={Towards Better Selective Classification},
    author={Leo Feng and Mohamed Osama Ahmed and Hossein Hajimirsadeghi and Amir H. Abdi},
    booktitle={International Conference on Learning Representations},
    year={2023},
    url={https://openreview.net/forum?id=5gDz_yTcst}
}

Acknowledgement

This code is based on the official code base of Self-Adaptive Training (which is based on the official code base of Deep Gambler).

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published