ldachuan / wenn

Multidimensional Uncertainty-Aware Evidential Neural Networks ( AAAI2021)

WGAN-ENN

This repository contains the essential code for the paper Multidimensional Uncertainty-Aware Evidential Neural Networks (AAAI 2021).

The code is written by Python 3.8 and pytorch 1.5 in GPU version. It has also been tested under Python 3.6 and pytorch 1.7.

Quick Start

1. Create folders 'datasets' and 'results' to save downloaded datasets and output results.
2. Most of the datasets will be downloaded automatically. You can download notMNIST_small You can evaluate on either LSUN_classroom or LSUN_resize. The results are close.
3. Run test_demo.py

More details

4. We first pretrain the ENN classfier to reach a good accuracy and then feed it into the alogrithm to calibrate its uncertainty. All the pretrained classifiers are under the folder 'pretrain'. Note that the pretrained ENN models are different from the normal softmax pretrained model. It is trained using Eq.9, not cross-entropy. Please also refer to the pretrain function in main.py.

5. The repo below contain a lot of baselines considered in our paper. But we change all the base classfier as ResNet-20 in our experiments.

   • Outlier Explosure

   • CCU

   • BayesByHpernet

6. We also provide an example of plotting Figure 4, Figure 5 and Figure 6.

Citation

If you find this repo useful in your research, please consider citing:

@article{hu2020multidimensional,
  title={Multidimensional Uncertainty-Aware Evidential Neural Networks},
  author={Hu, Yibo and Ou, Yuzhe and Zhao, Xujiang and Cho, Jin-Hee and Chen, Feng},
  journal={arXiv preprint arXiv:2012.13676},
  year={2020}
}