Skip to content

j-cb/Breaking_Down_OOD_Detection

Repository files navigation

Code for the ICML 2022 paper

Julian Bitterwolf, Alexander Meinke, Max Augustin and Matthias Hein

University of Tübingen

https://arxiv.org/abs/2206.09880

Paper Summary

gitlab_rendered_summary_readme.png Breaking_Down_OOD_Detection_ICML_Table1.png

Training

For training new models, run the train_ood_aware.py script with the appropriate arguments. For example, the shared classifier/binary discriminator for CIFAR-100 can be reproduced by running

python train_ood_aware.py --dataset_in CIFAR100 --dataset_out OpenImages --gpu 0 --method SharedCD --save ./my_saves .

Evaluation

The models from Table 1 in the paper can be evaluated with

python evaluate_ood_detection_methods.py --gpu 0 trained_models/eval_OpenImages.json .

To evaluate other models, add them to a .json file with their location and relevant hyperparameters, if applicable.

Setup

The paths in utils/paths_config.py need to be set to the locations of your datasets.

If a dataset is not available for evaluation, it will be skipped for the evaluation.

The code was tested with Python 3.8.5, PyTorch 1.6.0 and torchvision 0.7.0.

Releases

No releases published

Packages

No packages published

Languages