Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/jiin99/Generlized-OutlierExposure
This repository is the official implementation of the Generalized Outlier Exposure:
https://github.com/jiin99/Generlized-OutlierExposure
Last synced: 7 days ago
JSON representation
This repository is the official implementation of the Generalized Outlier Exposure:
- Host: GitHub
- URL: https://github.com/jiin99/Generlized-OutlierExposure
- Owner: jiin99
- License: mit
- Created: 2022-02-21T01:23:49.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2024-02-17T10:22:18.000Z (9 months ago)
- Last Synced: 2024-08-02T15:31:03.578Z (3 months ago)
- Language: Python
- Size: 9.11 MB
- Stars: 2
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Generlized-OutlierExposure
This repository is the official implementation of [Generalized Outlier Exposure: Towards a trustworthy out-of-distribution detector without sacrificing accuracy](https://www.sciencedirect.com/science/article/abs/pii/S0925231224001425 ), accepted as a Journal article in Neurocomputing, 2024.
A part of code has been based on the public code of
[Outlier Exposure](https://github.com/hendrycks/outlier-exposure), [SC-OOD](https://github.com/jingkang50/ICCV21_SCOOD), [Unknown Detection](https://github.com/daintlab/unknown-detection-benchmarks), [Mixup](https://github.com/facebookresearch/mixup-cifar10).
# Overview
Generalized OE is an effective training algorithm to obtain a reliable OoD detector in real-world scenarios. Existing OE-based methods have shown strong OoD detection performance, however, they have the performance degradation issue on ID. G-OE successfully overcomes the drawbacks of current OE-based methods by utilizing RandomMixup and a filtering strategy simultaneously.
## Visualize the effect of G-OE
A confidence surface of each experiment on a 2D spiral dataset. The color barrepresents how a model is confident in its predictions: the higher the confidence, the closerto yellow, and the lower the confidence, the closer to dark blue.
## Environment
* Python >= 3.6
* Pytorch >= 1.9
* CUDA >= 10.2
## In-distribution Dataset
* CIFAR10, CIFAR100
## Outlier Dataset for train
* [**80 Million Tiny Images**](http://www.archive.org/download/80-million-tiny-images-2-of-2/tiny_images.bin)
## Outlier Datasets for test
SC-OOD dataset can be downloaded by the following link: [SC-OOD dataset download](https://drive.google.com/file/d/1cbLXZ39xnJjxXnDM7g2KODHIjE0Qj4gu/view).
* [SC-OOD dataset](https://github.com/jingkang50/ICCV21_SCOOD)
* Blobs
* Gaussian
## Train a model
```Python
python main.py --dataset cifar100 --model res34 --exp-type G-OE --filtered_num 20 --trial 01 --gpu-id 0 --save-path ./save-path/
```
## Evaluate a model
```Python
python test.py --dataset cifar100 --model res34 --conf mcp --gpu-id 0 --datapath --save-path ./save-path/
```
### Arguments
| name |type | available values | help |
|:----------------:|:---:|:--------------------:|:---------------------------------------------------------------|
| dataset | str | cifar10, cifar100 | choose ID dataset between cifar10 and cifar100|
| model | str | wrn, dense, res34 | choose architecture |
| batch-size | int | int | batch size for ID |
| oe-batch-size | int | int | batch size for OoD |
| droprate | float | float | dropout probability |
| weight | float | float | final loss weight for RandomMixup term |
| oe-weight | float | float | final loss weight for OE term |
| filtered_num | int | int | number of samples to filter out, which indicates hyperparameter k in paper |
| save | str | save directory | directory where the results will be saved |
| trial | str | str | the number of trial experiments |
### Evaluation metrics
* ACC
* AURC
* FRR at 95% TPR
* AUROC
* AUPR
## Results
## Citation
```
@article{koo2024generalized,
title={Generalized Outlier Exposure: Towards a trustworthy out-of-distribution detector without sacrificing accuracy},
author={Koo, Jiin and Choi, Sungjoon and Hwang, Sangheum},
journal={Neurocomputing},
pages={127371},
year={2024}
}
```
### Contact for issues
- Jiin Koo, [email protected]