This repository contains code used for
Kuan-Chieh Wang, Paul Vicol, Eleni Triantafillou, Richard Zemel.
(2020). “Few-shot Out-of-Distribution Detection.”
International Conference on Machine Learning (ICML)
Workshop on Uncertainty and Robustness in Deep Learning
Python==3.7.3
For packages, see requirements.txt.
Example cmd for creating a Conda Env
conda create -n pnl python=3.7 pip
source activate pnl
pip install -r requirements.txt
Download miniImageNet
# takes a few seconds
./download_scripts/download_miniimagenet.shDownload the out-of-dataset (OOS) datasets
# takes a few seconds
./download_scripts/download_anomaly.sh # Downloads the Texture, Places, and notMNIST datasets
# each of the following takes a few seconds to download, untar takes a few minutes (<5)
./download_scripts/download_isun.sh
./download_scripts/download_lsun.sh
./download_scripts/download_tinyimagenet.shTo train standard (all-way) classifiers
for CIFAR-FS
./run_scripts/classify_classic/cifar.shfor miniImageNet
./run_scripts/classify_classic/miniimagenet.sh./run_scripts/train_confidence/cifar.shfor CIFAR-FS
./run_scripts/eval_ood/cifar.sh./run_scripts/train/submitted.shThis code is built upon many other repos.
An incomplete list includes:
- https://github.com/jakesnell/prototypical-networks
- https://github.com/wyharveychen/CloserLookFewShot
- https://github.com/hendrycks/outlier-exposure
- https://github.com/facebookresearch/odin
- https://github.com/pokaxpoka/deep_Mahalanobis_detector
Also, during development of the private (and ugly repo), my colleagues Paul and Eleni contributed in very significant ways.
If you use this code, consider citing:
@article{wang2020fsood,
title={Few-shot Out-of-Distribution Detection},
author={Kuan-Chieh Wang and Paul Vicol and Eleni Triantafillou and Richard Zemel},
booktitle={{International Conference on Machine Learning (ICML) Workshop on Uncertainty and Robustness in Deep Learning}},
year={2020}
}