Project page PASTA (Perceptual Assessment System for explanaTion of Artificial Intelligence)

This repository contains code for two main components of the paper: Benchmarking XAI Explanations with Human-Aligned Evaluations. For more details, see the paper: Benchmarking XAI Explanations with Human-Aligned Evaluations.

The PASTA dataset is available on Hugging Face: PASTA Dataset.

PASTA Score Calculation

To compute the PASTA score for a new XAI method:

1. Download the test set images
2. Generate explanations for all images in the test set (test_imgs_PASTA_score.zip in the PASTA Dataset link):
   • For saliency-based explanations, save the resulting heatmap images
   • For CBM-based explanations, save JSON dictionaries containing concept values
   • File format should be: {ROOT}/{DATASET}_{ID_IMG}_{LABEL}.{EXTENTION}
3. Run the PASTA score script:

python pasta_score.py --path_explanation PATH_TO_THE_EXPLANATIONS --id_question ID_PASTA_QUESTION --out_format TYPE_OF_EXPLANATION

Explanation Computation Pipeline

To generate explanations using the XAI methods included in the PASTA dataset:

1. Download the classifier's weights from: PASTA Weights on Hugging Face
2. Place the downloaded weights in the models_pkl folder
3. Run the explanation computation script:

python compute_explanation.py --network CLASSIFIER_NAME --expl_method XAI_METHOD --input_dataset CLASSIFICATION_DATASET --input_image_path PATH_OF_THE_IMG_TO_EXPLAIN

Example:

python compute_explanation.py --network resnet50 --expl_method GradCAM --input_dataset monumai --input_image_path data_samples/Monumai/Images/MonuMAI_6_Baroque.png

Implemented Methods

The following table summarizes the XAI methods included in the PASTA dataset:

Name	Functioning	Attribution on	Stage	Applied on
BCos	Interpretable latent space	Image	Ante-hoc	ResNet50-BCos
GradCAM	Gradient	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
HiResCAM	Gradient	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
GradCAMElementWise	Gradient	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
GradCAM++	Gradient	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
XGradCAM	Gradient	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
AblationCAM	Perturbation	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
ScoreCAM	Perturbation	Image	Post-hoc	ViT, ResNet50
EigenCAM	Factorization	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
EigenGradCAM	Gradient+Factorization	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
FullGrad	Gradient	Image	Post-hoc	ViT, ResNet50
Deep Feature Factorizations	Factorization	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
SHAP	Perturbation	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
LIME	Perturbation	Image	Post-hoc	ViT, ResNet50, CLIP (zero-shot)
X-NeSyL	Interpretable latent space	Concepts	Ante-hoc	X-NeSyL
CLIP-linear-sample	Interpretable latent space	Concepts	Ante-hoc	CLIP-linear
CLIP-QDA-sample	Counterfactual	Concepts	Ante-hoc	CLIP-QDA
LIME-CBM	Perturbation	Concepts	Post-hoc	CLIP-QDA, ConceptBottleneck
SHAP-CBM	Perturbation	Concepts	Post-hoc	CLIP-QDA, ConceptBottleneck
RISE-CBM	Perturbation	Concepts	Post-hoc	ConceptBottleneck