Welcome to TorchRay

The TorchRay package implements several visualization methods for deep convolutional neural networks using PyTorch. In this release, TorchRay focuses on attribution, namely the problem of determining which part of the input, e.g., an image, is responsible for the value computed by a neural network.

TorchRay is research-oriented. In addition to implementing well known techniques form the literature, it provides code for reproducing results that appear in several papers, and can thus be a tool for reproducible research.

For downloads, installation instructions, and access to the source code use the GitHub repository.

• Attribution
  • Backpropagation methods
    • Changing the backward propagation rules
    • Probing intermediate activations and gradients
    • Limitations
  • Perturbation methods
  • DeConvNet
    • Theory
  • Excitation backprop
    • Contrastive variant
    • Theory
  • Extremal perturbation
    • Perturbation types
    • Extremal perturbation variants
  • Gradient
  • Grad-CAM
    • Theory
  • Guided backprop
    • Theory
  • Linear approximation
  • RISE
  • Common code
• Benchmarking
  • Pointing Game
  • Datasets
  • Reference models
  • Logging with MongoDB
• Utilities

Indices

• Index

• Module Index

• Search Page