Expected / IntegratedGradients by PyTorch

Overview

• Simple implementation of IntegratedGradients and ExpectedGradients by PyTorch
• Integrated Gradients is Feature Attribution methods for Neural Networks (Sundararajan et al., 2017)
  • paper
• Expected Gradients is an extension of IG, which samples baseline inputs from the given dataset (Erion et al., 2019)
  • paper

Usage

from attr import integrated_gradients, expected_gradients

model = ...
target = ...
n_iter = ...
baseline = ...
data = ...

ig_attr = integrated_gradients(model, data, target, n_iter)
eg_attr = expected_gradients(model, data, baseline, target, n_iter)

• n_iter : the number of iterations used by the approximation method
  • the higher n_iter is, the more accurate approximation but more memory usage

Model and Data

• Model
  • simple 2 CNN layers
• Data
  • used MNIST data

requirements

numpy
torch
matplotlib