SAL

This is the source code accompanying the paper How Does Unlabeled Data Provably Help Out-of-Distribution Detection? by Xuefeng Du, Zhen Fang, Ilias Diakonikolas, and Yixuan Li

Ads

Check out our ICML'23 SCONE on using wild data for both OOD detection and generalization if you are interested!

Dataset Preparation

CIFAR-10/CIFAR-100

• The dataloader will download it automatically when first running the programs.

OOD datasets

• The OOD datasets with CIFAR-100 as in-distribution are 5 OOD datasets, i.e., SVHN, PLACES365, LSUN-C, LSUN-R, TEXTURES.
• Please refer to Part 1 and 2 of the codebase here.

Training and Inference

Please execute the following in the command shell:

python main.py --dataset cifar 10 --aux_out_dataset lsun_c --test_out_dataset lsun_c --pi 0.1 --num_class 10

"dataset" denotes the in-distribution training data.

"aux_out_dataset" determines the type of OOD data in the unlabeled wild data

Toy data

Please execute the following in the command shell:

python toy_data.py  --N_id 1000  --circle_ood 1 --ood_rate 0.1 --no_conditional 1 --use_thres 1  --num_epochs 3

It will reproduce the figures in the paper. Feel free to adjust the number of data, training epochs, ood rates in the unlabeled data, etc.

Limitations

Calculating the score is kind of slow for a large unlabeled wild data right now. I will need to think about how to speed up this procedure. Please consider use a small OOD dataset to construct the wild data, such as Textures/LSUN-C/LSUN-R for quick verification.

Citation

If you found any part of this code is useful in your research, please consider citing our paper:

@inproceedings{du2024sal,
  title={How Does Wild Data Provably Help OOD Detection?},
  author={Du, Xuefeng and Fang, Zhen and  Diakonikolas, Ilias and Li, Yixuan},
  booktitle={Proceedings of the International Conference on Learning Representations},
  year={2024}
}