lygjwy/dos

[ICLR 2024] An official PyTorch implementation of paper 'DOS: Diverse Outlier Sampling for Out-of-Distribution Detection'

DOS

An official PyTorch implementation of the ICLR 2024 paper
"DOS: Diverse Outlier Sampling for Out-of-Distribution Detection"

   

Get Started • Citation

Get Started

Overview

This repository is an official PyTorch implementation of the ICLR 2024 paper 'DOS: Diverse Outlier Sampling for Out-of-Distribution Detection'. The illustration of our algorithm is shown as below:

Requirements

pip install -r requirements.txt

Training

python train_diverse.py

Evaluation

OODs="svhn lsunc dtd places365_10k tinc lsunr tinr isun"
python detect.py --id cifar100 --ood $OODs --score abs --pretrain /path/to/trained/classifier

Results

Citation

If you find our repository useful for your research, please consider citing our paper:

@inproceedings{jiang2024dos,
  title={{DOS}: Diverse Outlier Sampling for Out-of-Distribution Detection},
  author={Wenyu Jiang and Hao Cheng and MingCai Chen and Chongjun Wang and Hongxin Wei},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=iriEqxFB4y}
}

dos's Issues

Image Resolution Inconsistent

I have noticed that in your paper, you said the experiment conducted on ImageNet-1K is based on a pre-trained DenseNet121 model. The pre-trained method you mentioned in Appendix B is trained on a downsampled ImageNet with 3232 resolution. However, your following experiment on ImageNet with IN-10 as ID and IN-990 as OOD is conducted on original resolution 224224? I just wonder if there exists Image Resolution conflict in your experiments?

Request for Detailed README for Code Execution

Hello,

Firstly, I would like to express my gratitude for sharing your code with the community; it's greatly appreciated.

However, I am encountering some challenges with running the code due to the need for dataset preparation, installing specific packages, and executing the training and testing scripts. Could you please provide a detailed README or documentation that outlines the steps for setting up and running the code?

Thank you very much for your assistance!