Introduction to VisuaLIME

Getting Started

To install VisuaLIME, run:

pip install visualime

VisuaLIME provides two convenience functions (explain_classification and render_explanation) that package its building blocks into a reference explanation pipeline. While you’ll likely want to create your own version later on, they’re a great starting point.

Note

If you’re new to LIME, you might want to check out our Grokking LIME talk/tutorial for a general introduction.

It covers the fundamentals of loading and processing images, classifying images using a pre-trained deep learning model, and walks through the essential steps that LIME uses to generate an explanation. The simplified version of LIME you’ll create as part of tutorial is directly derived from VisuaLIME.

Under the hood, VisuaLIME uses numpy. Hence, you need to provide the image whose classification you want to explain as a numpy array of shape (width, height, 3) representing an RGB image.

Since LIME is a model-agnostic explanation method, it does not make any assumptions about the classifier you’re using. It’s on you to provide a function that takes a numpy array of shape (num_of_samples, width, height, 3) representing a collection of num_of_samples RGB images and returns a numpy array of shape (num_of_samples, num_of_classes) where each entry corresponds to the classifiers output for the respective image:

import numpy as np

def predict_fn(images: np.ndarray) -> np.ndarray:
   predictions = ...  # call your classifier
   return predictions

To check that you’ve prepared everything correctly, try to generate a prediction for your image as follows:

image = ...  # load your image as a three-dimensional numpy array

predict_fn(image.reshape(1, -1))

VisuaLIME conceptually structures the explanation process into two steps:

  1. Computing an explanation, which consists of a segment_mask and corresponding segment_weights

  2. Rendering a visualization of the explanation

This is what generating an explanation with VisuaLIME looks like in code:

from visualime.explain import explain_classification, render_explanation

segment_mask, segment_weights = explain_classification(image, predict_fn)

explanation = render_explanation(
        image,
        segment_mask,
        segment_weights,
        positive="green",
        negative="red",
        coverage=0.2,
    )

For a full, interactive example with more detailed instructions, see the example notebook on GitHub.

Why VisuaLIME?

We initially used the original LIME implementation in our XAI Demonstrator project. After a while, we faced the issue that this version gives you very little control over how the explanation is rendered. In fact, the way segments are colored is somewhat misleading, as the opacity is based not on the weight of a segment but the maximum value of any color channel of the original image within the segment.

The original implementation is a classic example of “research code”: Written while conceiving and exploring a new algorithm, with lots of hard-to-follow data manipulation and cluttered with abandoned experiments. While it serves its purpose as a research tool, the tight coupling between the different components and its poor test coverage make it hard to extend or adapt.

To our knowledge, there is only one other LIME implementation that is not directly based on the original implementation. The popular PyTorch interpretability library Captum contains a version which (in line with the general approach of the library) is relatively low-level. For example, users are expected to provide their own similarity functions and take care of input segmentation. Further, it depends on PyTorch as the computational backend.

Hence, we decided to write our own version, specifically tailored towards computer vision applications and the generation of explanations for end-users. Similar to the original implementation, we’ve opted for numpy and scikit-learn as the foundation. We have structured the library around the idea of an “explanation pipeline”, a chain of small, exchangeable building blocks that can be selected according to the particular use case.

Package overview

VisuaLIME is currently structured into five modules: