InterpretDL

InterpretDL is an open source toolkit for interpretation algorithms based on PaddlePaddle. This toolkit contains three kinds of interpreters: input feature interpreters, intermediate-layer feature interpreters, and dataset-level interpreters. InterpretDL also provides the evaluation metrics to verify the trustworthiness of interpretation algorithms.

Contents:

• Interpreters
  • Base Interpreter
    • Abstract Interpreter
    • Sub-abstract: Input Gradient Interpreter
    • Sub-abstract: Input Output Interpreter
    • Sub-abstract: Intermediate-Layer Interpreter
    • Sub-abstract: Transformer Interpreter
  • Input Feature Based Interpreters
    • Consensus
    • Gradient Shap
    • Integrated Gradients
    • LIME
    • GLIME
    • LIME With Global Prior
    • LRP
    • Occlusion
    • Smooth Gradients
    • Smooth Gradients V2
    • NormLIME
    • TAM
    • Generic Attention
    • Bidirectional Transformer Interpreter
  • Intermediate-Layer Feature Interpreters
    • Grad-CAM
    • Score CAM
    • Rollout
  • Dataset-Level Interpreters
    • Training Dynamics
    • Beyond Hand-designed Feature Interpreter
    • Forgetting Events
    • SGDNoise
    • TrainIng Data analYzer (TIDY)
• Interpreter Trustworthiness Evaluation Metrics
  • Abstract Evaluator
  • DeletionInsertion
  • Perturbation
  • Infidelity
• Model Interpretability Evaluation Metrics
  • PointGame
  • PointGameSegmentation

Indices and tables

• Index
• Module Index
• Search Page