The following papers are covered in this tutorial:

• Maximum Softmax Probability
• ODIN
• Mahalanobis distance-based ODD
• MSP-like Error Detection
• Bayes by backprop. I did not code this part myself and re-used this code written by saxena-mayur.

Finally, the data_and_nn_loader.py file is taken from this repo.

Python >3.6 is required to run this code, along with Pytorch, matplotlib, scikit-learn and ipywidgets (for interactive plots).

Run the following command lines to set the environment using conda:

conda create --name tutorial python=3.6
conda activate tutorial
conda install -c conda-forge matplotlib
conda install pytorch torchvision -c pytorch
conda install -c anaconda scikit-learn
conda install -c conda-forge ipywidgets

Please have special care for the extra steps needed to complete ipywidgets installation. Indeed, depending on the version of jupyter-lab at use, it may not be automatically configured to use widgets.

To download the resized ImageNet test dataset and the DenseNet-10 architecture pre-trained on CIFAR-10, run the following commands:

chmmod +x download.sh
./download.sh