explainable-ai-image-measures 1.0.1

Compute IAUC, DAUC, IROF scores to measure quality of image attributions

Compute IAUC, DAUC, IROF scores to measure the quality of image attributions (PyTorch)

This package allows to evaluate the quality of attributions wrt. to your original iamge. For that purpose, you can compute IAUC (Insert Area Under Curve, Insertion score), DAUC (Delete Area Under Curve, Deletion score) and IROF (Iterative Removal Of Features).

Install via PyPi: pip install explainable-ai-image-measures

The shared idea behind all of them is to remove the most relevant pixels from the original image and replace it by a baseline color (DAUC, IROF) or to take blank a blank image and add the most relevant pixels again starting with the most relevant (IAUC). In contrast to IAUC and DAUC, IROF does this superpixel-wise.

1. Create new images by removing / adding pixels / entire areas
2. Send the newly created images through the network and save the probabilities.
3. Compute the Area Under the Curve (AUC) for the recorded probabilities.
4. Interpretation: For DAUC lower scores are better, for IROF and IAUC higher scores.

In contrast to other publicly available code,

• all of the heavy computation is running on the GPU (no numpy or CPU in between)
• you can compute the scores for a single image also batch-wise
• there is parametrization possible to suit your needs
• IAUC, DAUC and IROF is possible using a single package
• multiple attributions for multiple images are possible (see example.ipynb)

Here a minimal code for a single image:

import torch
from captum.attr import Saliency
import torchvision.models as models

from explainable_ai_image_measures import Measures

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = models.resnet18(pretrained=True).to(device)

image = torch.rand((1, 3, 224, 224)).to(device)
attribution = Saliency(model).attribute(image, target=17)

# Average the attribution over the color channels
attribution = torch.mean(attribution, dim=1)

measure = Measures(model, batch_size=250, irof_segments=70)
score, y = measure.compute_IROF(image[0], attribution[0], 17)