RelevancePropagation.jl

Julia implementation of Layerwise Relevance Propagation (LRP) and Concept Relevance Propagation (CRP) for use with Flux.jl models.

Installation

To install this package and its dependencies, open the Julia REPL and run

julia> ]add RelevancePropagation

Manual

Basic Usage

• Creating an LRP Analyzer
• • Model preparation
  • • Stripping the output softmax
    • Model canonization
    • Flattening the model
  • LRP rules
  • Computing layerwise relevances
  • Performance tips
  • • Using LRP with a GPU
    • Using LRP without a GPU
• Assigning LRP Rules to Layers
• • Manually assigning rules
  • Custom composites
  • Composite primitives
  • Assigning a rule to a specific layer
  • Composite presets
• Concept Relevance Propagation
• • Step 1: Create LRP analyzer
  • Step 2: Define concepts
  • Step 3: Use CRP analyzer
  • Using CRP on input batches

Advanced Usage

• Supporting New Layers and Activation Functions
• • Model checks
  • Registering layers
  • Registering activation functions
  • Skipping model checks
• Custom LRP Rules
• • Implementing a custom rule
  • • Step 1: Define rule struct
    • Step 2: Implement rule behavior
    • Step 3: Use rule in LRP analyzer
  • Performance tips
  • Advanced layer modification
  • Advanced LRP rules
• Developer Documentation
• • Generic LRP rule implementation
  • • Linear layers
    • The automatic differentiation fallback
  • LRP analyzer struct
  • Forward and reverse pass
  • • Rule calls
    • Specialized implementations

LRP Rule Overview

• LRP Rule Overview
• • Notation
  • Basic rules
  • Lower layer rules
  • Input layer rules
  • Specialized rules

API Reference

• API Reference
• • Basic API
  • LRP analyzer
  • • Model preparation
  • LRP rules
  • Composites
  • • Applying composites
    • Composite primitives
    • Composite presets
    • Manual rule assignment
  • Custom rules
  • CRP
• Index