ODIN: Out-of-Distribution Detector for Neural Networks

This is a fork of facebookresearch/odin written in a morden way with the power of functorch and TorchMetrics.

The method is described in the paper Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks by S. Liang, Yixuan Li and R. Srikant.

Running the code

Dependencies

Tested on:

• pytorch==1.13.1
• torchmetrics==0.11.1
• tqdm, matplotlib, ...

Downloading

In the root of the repository, run

sh download.sh

Out-of-Distribtion Datasets

facebookresearch/odin provide download links of five out-of-distributin datasets:

• Tiny-ImageNet (crop)
• Tiny-ImageNet (resize)
• LSUN (crop)
• LSUN (resize)
• iSUN

Neural Network Models

We can use any pytorch model.

facebookresearch/odin provide download links of four pre-trained models.

• DenseNet-BC trained on CIFAR-10
• DenseNet-BC trained on CIFAR-100
• Wide ResNet trained on CIFAR-10
• Wide ResNet trained on CIFAR-100

DenseNet-BC models print some warnings, but they work. Wide ResNet models need 3 GPUs and older version of pytorch (NOT tested).

Running

See ODIN.ipynb.