Removing input features via a generative model to explain their attributions to classifier's decisions

This repository contains source code necessary to reproduce some of the main results in the paper:

If you use this software, please consider citing:

@misc{agarwal2019removing,
title={Removing input features via a generative model to explain their attributions to classifier's decisions},
author={Chirag Agarwal and Dan Schonfeld and Anh Nguyen},
year={2019},
eprint={1910.04256},
archivePrefix={arXiv},
primaryClass={cs.LG}

1. Setup

Installing software

This repository is built using PyTorch. You can install the necessary libraries by pip installing the requirements text file pip install -r ./requirements.txt

Note: We ran our codes using tensorflow_gpu-1.12.0 with CUDA 9.0. Add the following lines in your ~/.bashrc file if you get "ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory" on running the code.

export PATH=/usr/local/cuda-9.0/bin${PATH:+:${PATH}}

export LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

2. Usage

The main scripts for SP/SP-G, LIME/LIME-G, and MP/MP-G are in formal_SP_single_image.py, formal_LIME_single_image.py, and formal_MP_single_image.py. In addition, after installing the LIME library you will have to replace the lime_image.py script with our lime_image.py. Three shell scripts have been provided which for a given an image and target class generates its respective attribution maps for an algorithm and its generative version.

Examples

SP_test.sh: Generating the attribution map for the class "freight car" using SP and SP-G algorithm.

• Running source SP_test.sh produces this result:

(left-->right) The real image followed by five random intermediate perturbed images and the resultant attribution map for SP (top) and SP-G (bottom). For each intermediate perturbed image, the top and bottom row labels shows the target and top-1 class predictions with their respective probabilities.

LIME_test.sh: Generating the attribution map for the class "kuvasz" using LIME and LIME-G algorithm.

• Running source LIME_test.sh produces this result:

(left-->right) The real image followed by five random intermediate perturbed images and the resultant attribution map for LIME (top) and LIME-G (bottom). For each intermediate perturbed image, the top and bottom row labels shows the target and top-1 class predictions with their respective probabilities.

MP_test.sh: Generating the attribution map for the class "freight car" using MP and MP-G algorithm.

• Running source MP_test.sh produces this result:

(left-->right) The real image followed by five random intermediate perturbed images and the resultant attribution map for MP (top) and MP-G (bottom). For each intermediate perturbed image, the top and bottom row labels shows the target and top-1 class predictions with their respective probabilities.

4. Licenses

Note that the code in this repository is licensed under MIT License, but, the pre-trained condition models used by the code have their own licenses. Please carefully check them before use.

5. Questions?

If you have questions/suggestions, please feel free to email or create github issues.