Unboxer is the supporting tool for the paper: An Empirical Study on Low- and High-Level Explanations of Deep Learning Misbehaviours.
First, you should install the environment and set the configurations based on the case study (MNIST or IMDB) you want to run:
📲 Install 📲
⚙️ Configure ⚙️
You should run the following command to generate the inputs for corresponding case study.
python -m utls.generate_inputs
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.
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.
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.
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 **
