Awesome

Training OOD Detectors in their Natural Habitats

This is the official repository of Training OOD Detectors in their Natural Habitats by Julian Katz-Samuels, Julia Nakhleh, Rob Nowak, and Yixuan Li. This method trains OOD detectors effectively using auxiliary data that may be a mixture of both inlier and outlier examples.

Pretrained models

You can find the pretrained models in

./CIFAR/snapshots/pretrained

Datasets

Download the data in the folder

./data

Run

To run the code, execute

bash run.sh score in_distribution aux_distribution test_distribution 

For example, to run woods on cifar10 using dtd as the mixture distribution and the test_distribution, execute

bash run.sh woods cifar10 dtd dtd 

pi is set to 0.1 as default. See the run.sh for more details and options.

Main Files

• CIFAR/train.py contains the main code used to train model(s) under our framework.
• CIFAR/make_datasets.py contains the code for reading datasets into PyTorch.
• CIFAR/plot_results.py contains code for loading and analyzing experimental results.
• CIFAR/test.py contains code for testing experimental results in OOD setting.

Datasets

Here are links for the less common outlier datasets used in the paper: Textures, Places365, LSUN, LSUN-R, iSUN, and 300K Random Images.