IDGI is a framework to eleminate explanation noises from IG (Integrated Gradients) based method.
By eleminating the noise, IDGI outputs better saliency map compare to original IG-based methods.
IDGI works with many IG-based method, e.g., IG, GIG, and BlurIG, which utilizes the Riemann integration for computing the final attribution map.
This repository contains the official implementation of IDGI. For more details, please check our paper:
numpy installed.
We will try to work with other interpretability/explanation packages to extend our method to these packages.