Detecting Out-of-Distribution Examples with In-distribution Examples and Gram Matrices

ICML 2020: paper, supplementary material and bibtex available at http://proceedings.mlr.press/v119/sastry20a.html

Dependencies

The code is written in Python 3 with Pytorch 1.1.

Results

(Please refer to this repository for the results of Baseline/ODIN/Mahalanobis on dataset-pairs not presented in the Mahalanobis paper)

Combining Outlier Exposure (OE) and Ours

Downloading Out-of-Distribution Datasets and Pre-trained Models

We used the out-of-distribution datasets presented in odin-pytorch

We used pre-trained neural networks open-sourced by Mahalanobis and odin-pytorch. The DenseNets trained on CIFAR-10 and CIFAR-100 are by ODIN; remaining are by Mahalanobis.

For experiments on OE-trained networks, we used the pre-trained networks open-sourced by OE

Running the setup.sh downloads the Out-of-Distribution Datasets and pre-trained models.