Information-Theoretic Visual Explanation

A tensorflow implementation of "Information-Theoretic Visual Explanation for Black-Box Classifiers"

Example

An parachute image was classified as a balloon.

• The IG map provides a class-independent explanation: the classifier made a decision based on the highlighted object (the parachute and the rope).
• The PMI map provides a class-specific explanation for the balloon class: the orange fabric looks a balloon, but the rope doesn't.

Compatibility

The code runs on python 3.7 and tensorflow 1.13.1.

Installation

Step 1. Install libraries.

• Install required libraries.

Step 2. Download model checkpoints.

• Download ckpts.zip and unzip the file.
• The zip file contains model checkpoints VGG19 (converted from pytorch model zoo to tensorflow) and trained PatchSampler.

Code examples

Step 1. Obtain the PMI and IG maps.

python main.py --image_path="data/parachute.png"

• image_path denotes the path to the image to explain.

Step 2. Check out the saved results in data/results/.

The examples of the PMI and IG maps are provided in data/results/parachute_K8_N8_S1.png