This folder contains the source code and weights for the experiments of our work, Probability Trust Intervals for Out of Distribution Input Detection.

Directory Structure

1. The files Cn.py (n denotes the architecture number) contains the code for the Perturbed-NNs.
2. The files, Cn_Eval.py (n denotes the architecture number) contains the code for evaluation of all the considered baselines and our proposed approach.

Experimental Setup

All the experiments are done on Google Colaboratory. The links to original files have been removed for the purpose of double blind review. If you are interested in executing our code, you may follow the steps given below,

1. Upload the code files to the above website.
2. Create a weights folder in your Google drive. Under this folder, create, mnist subfolder for weights of mnist, cifar10 subfolder for weights of cifar10 and so on for other datasets.
3. Mount Google drive to Google colaboratory. You may use these instructions in order to do the same.
4. Change the prefix variable (search for prefix = "") in the uploaded code file to the path of parent folder of weights in your Google drive. For example if your weights folder is somewhere in the path, ./drive/My Drive/reviewCode/PNN/weights then value of prefix should be ./drive/My Drive/reviewCode/PNN/ (observe that the path is to PNN, the parent folder of weights).
5. Now you should be able to run the notebook and re-generate the results.

Note

We are unable to provide weights of CNNs and PNNs due to the maximum file size limit imposed for the supplementary material.