Unboxer is the supporting tool for the paper: An Empirical Study on Low- and High-Level Explanations of Deep Learning Misbehaviours.

🥡 How to run the unboxer 🥡

First, you should install the environment and set the configurations based on the case study (MNIST or IMDB) you want to run:

📲 Install 📲

⚙️ Configure ⚙️

Generate inputs

You should run the following command to generate the inputs for corresponding case study.

python -m utls.generate_inputs

🥵 Generate the heatmaps 🥵

You should run the following command to generate the heatmaps.

python -m steps.process_heatmaps

The tool will experiment with the different explainers, find the best configuration for the dimensionality reduction, and export the data collected during the experiment.

🗺 Generate the featuremaps 🗺

You can run the following command to generate the featuremaps.

python -m steps.process_featuremaps

The tool will generate the featuremaps, and export the data collected during the experiment.

📊 Export the insights 📊

You can run the following command to generate the insights about the data.

python -m steps.insights.insights

!!! IMPORTANT !!!
Remember to generate the heatmaps and the featuremaps before running this command.

The tool with prompt a menu with a set of options, and will guide you through the process.

🤔 Export the data for the human evaluation 🤔

You can run the following command to export the data for the human evaluation.

python -m steps.human_evaluation.export_samples

!!! IMPORTANT !!!
Remember to generate the heatmaps and the featuremaps before running this command.

The tool will generate samples for human study in out/human_evaluation.

** Data generated for the corresponding paper is available in out folder **

DOI