This repo contains the code for the paper Conservative Uncertainty Estimation By Fitting Prior Networks.
The code requires Python >= 3.6 and a GPU with CUDA and CuDNN installed.
To install the reqirements, run
pip install -r requirements.txt
from the root directory.
All the scripts are in src/scripts and the different uncertainty models are in src/scripts/models.
The CIFAR-10 experiments from the paper can be reproduced by running
./reproduce_cifar_experiment.sh
from the src/scripts directory.
This should take about 16 hours on a single GPU and will produce three figures in the figures directory that should look like this:
The different models can be trained using the training script in src/scripts/train_uncertainties.py.
The command line options can be shown by running
python train_uncertainties.py --help
The neural network architectures used in this implementation are adapted from David Page and the dropout baseline is adapted from Yarin Gal.
If you want to cite this work, please use
@inproceedings{
Ciosek2020Conservative,
title={Conservative Uncertainty Estimation By Fitting Prior Networks},
author={Kamil Ciosek and Vincent Fortuin and Ryota Tomioka and Katja Hofmann and Richard Turner},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=BJlahxHYDS}
}