Out-of-Distribution-Baseline

This is the repository for training and evaluating Out-of-Distribution Detection.

For simplicity, we implement the situation where CIFAR-10 dataset is used as in-distribution dataset and SVHN, CIFAR100, LSUN, ImageNet as out-distribution datasets.

Users can simply refer to datasets/datasets and datasets/ood_datasets to further apply their training scheme to combinations of other datasets.

Dependency

The code is built with following libraries:

• PyTorch 1.2 ~ 1.7.1
• [Torchvision] 0.4.0 ~ 0.8.2 depending on the version of torch.
• scikit-learn

Other torch versions might work but we have not tested.

Training

We provide training example with this repo:

python ood_baseline.py

Different parameters, e.g. Epoch, BatchSize, and etc, can be adjusted with the arguments. Check arguments at the top of ood_baseline.py

References

Some codes are brought from CSI-novelty detection (Neurips 2020). Datasets can also be found in link