Getting started

Install

SIGN-XAI2 can be installed directly from PyPI:

$ pip install signxai2

Basic Usage

SIGN-XAI2 is based on Zennit, which implements propagation-based attribution methods by overwriting the gradient of PyTorch modules in PyTorch’s auto-differentiation engine. This means that Zennit will only work on models which are strictly implemented using PyTorch modules, including activation functions. The following demonstrates a setup to compute SIGN-based Layer-wise Relevance Propagation (LRP) relevance for a simple model and random data.

from zennit.attribution import Gradient
from signxai2.composites import EpsilonStdXSIGN

composite = EpsilonStdXSIGN(mu=0, stdfactor=0.3, signstdfactor=0.3)
with Gradient(model=model, composite=composite) as attributor:
     output, relevance = attributor(data, target)

print('EpsilonStdXSIGN:', relevance)

More information on attributors can be found here:

• Writing Custom Composites

• Writing Custom Canonizers

• Writing Custom Rules

• Writing Custom Attributors

Example Scripts

Ready-to use examples to analyze image and time series models can be found here:

• vgg16_simple.py

• vision_transformer.py

• dftlrp_synthetic.py