Compute integrated gradients — integrated

Computes integrated gradients scores for model and an input sequence. This can be used to visualize what part of the input is import for the models decision. Code is R implementation of python code from here. Tensorflow implementation is based on this paper.

Usage

integrated_gradients(
  m_steps = 50,
  baseline_type = "zero",
  input_seq,
  target_class_idx,
  model,
  pred_stepwise = FALSE,
  num_baseline_repeats = 1
)

Arguments

m_steps: Number of steps between baseline and original input.
baseline_type: Baseline sequence, either "zero" for all zeros or "shuffle" for random permutation of input_seq.
input_seq: Input tensor.
target_class_idx: Index of class to compute gradient for
model: Model to compute gradient for.
pred_stepwise: Whether to do predictions with batch size 1 rather than all at once. Can be used if input is too big to handle at once. Only supported for single input layer.
num_baseline_repeats: Number of different baseline estimations if baseline_type is "shuffle" (estimate integrated gradient repeatedly for different shuffles). Final result is average of num_baseline single calculations.

Value

A tensorflow tensor.

Examples

if (FALSE) { # reticulate::py_module_available("tensorflow")
library(reticulate)
model <- create_model_lstm_cnn(layer_lstm = 8, layer_dense = 3, maxlen = 20, verbose = FALSE)
random_seq <- sample(0:3, 20, replace = TRUE)
input_seq <- array(keras::to_categorical(random_seq), dim = c(1, 20, 4))
integrated_gradients(
  input_seq = input_seq,
  target_class_idx = 3,
  model = model)
  
}