ID Label

This is the source code accompanying the paper When and How Does In-distribution Label Help Out-of-Distribution Detection? by Xuefeng Du, Yiyou Sun, and Yixuan Li

Ads

Check out our ICLR'24 SAL on analyzing the effect of the unlabeled data for OOD detection 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

Please execute the following in the command shell for the unsupervised case (cifar100 as ID):

python run.py --config-file configs/my_resnet_mlp1000_norelu_cifar100.yaml --add none --gamma_u 1 --gamma_l 1

and

python run.py --config-file configs/my_resnet_mlp1000_norelu_cifar100.yaml --add combine --gamma_u 3 --gamma_l 0.0225

for the supervised case.

Please execute the following in the command shell for the unsupervised case (cifar10 as ID):

python run.py --config-file configs/my_resnet_mlp1000_norelu_cifar10.yaml --add none --gamma_u 1 --gamma_l 1

and

python run.py --config-file configs/my_resnet_mlp1000_norelu_cifar10.yaml --add combine --gamma_u 0.5 --gamma_l 0.25

for the supervised case.

Linear Probing

To run linear probing when the test ood distribution is the same as the training outliers, run:

python lp_same_dis.py --load_ckpt c100_sup.pth --ood_name svhn --config-file configs/my_resnet_mlp1000_norelu_cifar100.yaml

"ood_name" denotes the type of OOD data, and "load_ckpt" denotes the pretrained model.

To run linear probing when the test ood distribution is different from the training outliers, run:

python lp.py --load_ckpt c100_sup.pth --config-file configs/my_resnet_mlp1000_norelu_cifar100.yaml

Pretrained models

Please check the models here

Citing

If you find our code useful, please consider citing:

@inproceedings{du2024when,
      title={When and How Does In-Distribution Label Help Out-of-Distribution Detection?}, 
      author={Xuefeng Du and Yiyou Sun and Yixuan Li},
      booktitle = {International Conference on Machine Learning},
      year = {2024}
}