Robust Explainability

A Tutorial on Gradient-Based Attribution Methods for Deep Neural Networks

Full paper can be found here: https://arxiv.org/abs/2107.11400

Install

1. Clone repo.

git clone https://github.com/nielseni6/ShiftSmoothedAttributions.git

2. Make sure to have the following required libraries if you do not already.

Python 3.8.5

PyTorch 1.4.0

torchvision 0.5.0

matplotlib 3.5.1

numpy 1.16.3

robustness 1.2.1.post2

scikit-image 0.19.1

opencv-python 4.5.5.62

captum 0.4.1

numpy 1.22.1

3. Pretrained models can be found here: https://drive.google.com/drive/u/0/folders/1KdJ0aK0rPjmowS8Swmzxf8hX6gU5gG2U and here: https://www.dropbox.com/s/knf4uimlqsi1yz8/imagenet_l2_3_0.pt?dl=0

Add these files to the \model folder.

4. Run the following scripts to generate the figures which can be found in the paper:

Visualize_Saliency_ImageNet.py: fig2.pdf

Visualize_Target_Class.py: fig3.pdf

Visualize_Robust_Grad.py: fig4.pdf

For questions contact [email protected] to get more information on the paper