Evaluating and improving the faithfulness of the interpretations offered by Neural Module Networks - View it on GitHub