This repository contains the code needed to replicate the experiments in our paper: A Confusion Matrix for Evaluating Feature Attribution Methods.

The four feature attribution techniques used are:

• Layer-wise Relevance Propagation (LRP) [1]: the implementation used is based on the work of Nakashima et al.
• GradCAM [6]: the implementation used is based on the work of Gildenblat et al.
• LIME [5]: the implementation used is based on the work of Tulio et al.
• Integrated Gradients (IG) [7]: the implementation used is based on the work of Kokhlikyan et al. [2].

This code runs under Python 3.7.1. The python dependencies are defined in requirements.txt.

These are the mosaics used in our experiments:

• Dogs vs. Cats mosaics can be downloaded here
• MIT67 [4] mosaics can be downloaded here.
• MAMe [3] mosaics can be downloaded here.

These are the bash scripts needed to compute the different scores.

cd $PROJECT_PATH/explainability/scripts/explainability_dataset/

sh explainability_dataset_architecture_method.sh

cd $PROJECT_PATH/evaluation/scripts/evaluation_dataset/

sh evaluation_dataset_architecture_method.sh

cd $PROJECT_PATH/plots/scripts/plot_dataset/

sh plot_dataset_architecture_method.sh

where:

• dataset must be exchanged by one of the following datasets: catsdogs, mit67 or mame.
• architecture must be vgg16 or resnet18.
• method must be lrp, gradcam, lime or intgrad.

For example, to get the scores for the Dogs vs. Cats dataset, using the ResNet18 architecture and the GradCAM method, run the following:

cd $PROJECT_PATH/explainability/scripts/explainability_catsdogs/

sh explainability_catsdogs_resnet18_gradcam.sh

cd $PROJECT_PATH/evaluation/scripts/evaluation_dataset/

sh evaluation_catsdogs_resnet18_gradcam.sh

cd $PROJECT_PATH/plots/scripts/evaluation_dataset/

sh plot_catsdogs_resnet18_gradcam.sh

Please cite our paper when using this code.

[1] Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K. R., & Samek, W. (2015). On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one, 10(7), e0130140.

[2] Kokhlikyan, N., Miglani, V., Martin, M., Wang, E., Alsallakh, B., Reynolds, J., ... & Reblitz-Richardson, O. (2020). Captum: A unified and generic model interpretability library for pytorch. arXiv preprint arXiv:2009.07896.

[3] Parés, F., Arias-Duart, A., Garcia-Gasulla, D., Campo-Francés, G., Viladrich, N., Ayguadé, E., & Labarta, J. (2020). A Closer Look at Art Mediums: The MAMe Image Classification Dataset. arXiv preprint arXiv:2007.13693.

[4] Quattoni, A., & Torralba, A. (2009, June). Recognizing indoor scenes. In 2009 IEEE Conference on Computer Vision and Pattern Recognition (pp. 413-420). IEEE.

[5] Ribeiro, M. T., Singh, S., & Guestrin, C. (2016, August). " Why should i trust you?" Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (pp. 1135-1144).

[6] Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626).

[7] Sundararajan, M., Taly, A., & Yan, Q. (2017, July). Axiomatic attribution for deep networks. In International Conference on Machine Learning (pp. 3319-3328). PMLR.