signxai2 0.15.8

Extended variant of ZeNNit with SIGNed explanations - Unveiling relevant features by reducing bias

SIGNed XAI for Image and Time Series Models

SIGN (Sign-based Improvement of Gradient-based explaNations) is a novel XAI method intended to reduce bias in explanations that are intrinsically induced by several state-of-the-art XAI methods. The SIGN-XAI-2 package enables simple application of this method in your projects and is based on Zennit (https://pypi.org/project/zennit/) and LRP for Transformers (https://pypi.org/project/lxt/). If your are using TensorFlow instead of PyTorch, have a look at https://pypi.org/project/signxai/.

SIGN-based explanations are particularly well suited for generating bias-reduced heatmaps for both image and time series data, enhancing interpretability by more reliably uncovering relevant features.

If you use this package or parts of it in your own work, please consider citing our paper:

 @article{Gumpfer2023SIGN,
    title = {SIGNed explanations: Unveiling relevant features by reducing bias},
    author = {Nils Gumpfer and Joshua Prim and Till Keller and Bernhard Seeger and Michael Guckert and Jennifer Hannig},
    journal = {Information Fusion},
    pages = {101883},
    year = {2023},
    issn = {1566-2535},
    doi = {https://doi.org/10.1016/j.inffus.2023.101883}
}

Documentation

The latest documentation is available from timexaigroup.github.io/signxai2.

Install

To install the package directly from PyPI using pip, use:

$ pip install signxai2

Supported Methods

• Gradient
• Gradient $\times$ Input
• Gradient $\times$ SIGN
• LRP (all variants contained in Zennit. )
• LRP-$\epsilon$ with SIGN as Input Layer Rule

Usage

SIGN-XAI-2 is based on Zennit and works with PyTorch. If you want to know more about Zennit and its usage, visit github.com/chr5tphr/zennit. There is also a version of SIGN available for usage with TensorFlow environments pypi.org/project/signxai/.

The below example code (see vgg16_simple.py) demonstrates the basic usage of a SIGN-based composite as an extension to LRP-Epsilon in Zennit. To run this code, have a look at the requirements and the setup script.

import matplotlib.pyplot as plt
import numpy as np
import torch
from torchvision.models import vgg16, VGG16_Weights
from torchvision.transforms import Compose, Resize, ToTensor, Normalize
from zennit.attribution import Gradient
from signxai2.misc import get_example_image
from signxai2.composites import EpsilonStdXSIGN

# Define preprocessing pipeline
transform = Compose([
    Resize((224, 224)),
    ToTensor(),
    Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])

# Load and preprocess image
image = get_example_image(1)
data = transform(image)[None]

# Load pretrained VGG16 model
weights = VGG16_Weights.IMAGENET1K_V1
model = vgg16(weights=weights).eval()

# Get model prediction
output = model(data)
pred = output.argmax(1)[0].item()
target = torch.eye(1000)[[pred]]

# Get the class label
label = weights.meta['categories'][pred]
print('Predicted class: {}'.format(label))

# Compute attribution
composite = EpsilonStdXSIGN(mu=0, stdfactor=0.3, signstdfactor=0.3)
with Gradient(model=model, composite=composite) as attributor:
    _, attribution = attributor(data, target)

# Prepare relevance map
attribution = np.nan_to_num(attribution)[0]
relevance = attribution.sum(0)
R = relevance / np.abs(relevance).max()

# Visualize the image and relevance map
fig, axs = plt.subplots(1, 2, figsize=(10, 6))
axs[0].imshow(Compose([Resize((224, 224))])(image))
axs[0].set_title('Image')
axs[1].matshow(R, cmap='seismic', clim=(-1, 1))
axs[1].set_title('LRP-Epsilon-SIGN')

# Switch off axes and labels
for ax in axs:
    ax.axis('off')

# Plot to screen
plt.tight_layout()
plt.show()

For more details and examples, have a look at our documentation.

Examples

The above code can be used to generate heatmaps as shown for the example images below. Comparing the baseline LRP-Epsilon method with its SIGN-augmented counterpart highlights notable differences in pixel-level attributions and image-induced contrast, with SIGN mitigating this explanatory bias and distortion inherent to the standard variant.

Image Source: Coastal Roamer, CC BY-SA 4.0, via Wikimedia Commons

Image Source: randomwild, CC0, via Wikimedia Commons

License

SIGN-XAI-2 and Zennit are licensed under the GNU LESSER GENERAL PUBLIC LICENSE VERSION 3 OR LATER -- see the COPYING and COPYING.LESSER files for details.