Zennit Documentation

Zennit (Zennit Explains Neural Networks in Torch) is a python framework using PyTorch to compute local attributions in the sense of eXplainable AI (XAI) with a focus on Layer-wise Relevance Propagation. It works by defining rules which are used to overwrite the gradient of PyTorch modules in PyTorch’s auto-differentiation engine. Rules are mapped to layers with composites, which contain directions to compute the attributions of a full model, which maps rules to modules based on their properties and context.

Zennit is available on PyPI and may be installed using:

$ pip install zennit

Contents

• Getting started
  • Install
  • Basic Usage
  • Example Script
• How-Tos
  • Using Rules, Composites, and Canonizers
  • Using Attributors
  • Visualizing Results
  • Getting Intermediate Relevance
  • Computing Second Order Gradients
  • Writing Custom Composites
  • Writing Custom Canonizers
  • Writing Custom Rules
  • Writing Custom Attributors
• Tutorials
  • Image Classification with VGG and ResNet
• API Reference
  • zennit.attribution
  • zennit.canonizers
  • zennit.cmap
  • zennit.composites
  • zennit.core
  • zennit.image
  • zennit.layer
  • zennit.rules
  • zennit.torchvision
  • zennit.types
• Bibliography

Indices and tables

• Index

• Module Index

• Search Page

Citing

If you find Zennit useful, why not cite our related paper [Anders et al., 2021]:

@article{anders2021software,
      author  = {Anders, Christopher J. and
                 Neumann, David and
                 Samek, Wojciech and
                 Müller, Klaus-Robert and
                 Lapuschkin, Sebastian},
      title   = {Software for Dataset-wide XAI: From Local Explanations to Global Insights with {Zennit}, {CoRelAy}, and {ViRelAy}},
      journal = {CoRR},
      volume  = {abs/2106.13200},
      year    = {2021},
}