Integrated Gradients

The IntegratedGradient class implements the method Integrated Gradients (Sundararajan et al., 2017), which incorporates a reference value $x'$ (also known as baseline value) analogous to the DeepLift method. Integrated Gradients helps to uncover the relative importance of input features in the predictions $y = f(x)$ made by a model compared to the prediction of the reference value $y' = f(x')$. This is achieved through the following formula: $$ (x - x') \times \int_{\alpha=0}^{1} \frac{\partial f(x' + \alpha (x - x'))}{\partial x} d\alpha $$ In simpler terms, it calculates how much each feature contributes to a model's output by tracing a path from a baseline input $x'$ to the actual input $x$ and measuring the average gradients along that path.

Similar to the other gradient-based methods, by default the integrated gradient is multiplied by the input to get an approximate decomposition of $y - y'$. However, with the parameter times_input only the gradient describing the output sensitivity can be returned.

The R6 class can also be initialized using the run_intgrad function as a helper function so that no prior knowledge of R6 classes is required.

References

M. Sundararajan et al. (2017) Axiomatic attribution for deep networks. ICML 2017, PMLR 70, pp. 3319-3328.

Super classes

innsight::InterpretingMethod -> innsight::GradientBased -> IntegratedGradient

Public fields

n: (integer(1))
Number of steps for the approximation of the integration path along $\alpha$ (default: $50$).
x_ref: (list)
The reference input for the IntegratedGradient method. This value is stored as a list of torch_tensors of shape (1, dim_in) for each input layer.

Methods

Inherited methods

Method `new()`

Create a new instance of the IntegratedGradient R6 class. When initialized, the method Integrated Gradient is applied to the given data and baseline value and the results are stored in the field result.

Usage

IntegratedGradient$new(
  converter,
  data,
  x_ref = NULL,
  n = 50,
  times_input = TRUE,
  channels_first = TRUE,
  output_idx = NULL,
  output_label = NULL,
  ignore_last_act = TRUE,
  verbose = interactive(),
  dtype = "float"
)

Arguments

converter

(Converter)
An instance of the Converter class that includes the torch-converted model and some other model-specific attributes. See Converter for details.

data

(array, data.frame, torch_tensor or list)
The data to which the method is to be applied. These must have the same format as the input data of the passed model to the converter object. This means either

an array, data.frame, torch_tensor or array-like format of size (batch_size, dim_in), if e.g., the model has only one input layer, or
a list with the corresponding input data (according to the upper point) for each of the input layers.

x_ref

(array, data.frame, torch_tensor or list)
The reference input for the IntegratedGradient method. This value must have the same format as the input data of the passed model to the converter object. This means either

an array, data.frame, torch_tensor or array-like format of size (1, dim_in), if e.g., the model has only one input layer, or
a list with the corresponding input data (according to the upper point) for each of the input layers.
It is also possible to use the default value NULL to take only zeros as reference input.

n

(integer(1))
Number of steps for the approximation of the integration path along $\alpha$ (default: $50$).

times_input

(logical(1)
Multiplies the integrated gradients with the difference of the input features and the baseline values. By default, the original definition of IntegratedGradient is applied. However, by setting times_input = FALSE only an approximation of the integral is calculated, which describes the sensitivity of the features to the output.

channels_first

(logical(1))
The channel position of the given data (argument data). If TRUE, the channel axis is placed at the second position between the batch size and the rest of the input axes, e.g., c(10,3,32,32) for a batch of ten images with three channels and a height and width of 32 pixels. Otherwise (FALSE), the channel axis is at the last position, i.e., c(10,32,32,3). If the data has no channel axis, use the default value TRUE.

output_idx

(integer, list or NULL)
These indices specify the output nodes for which the method is to be applied. In order to allow models with multiple output layers, there are the following possibilities to select the indices of the output nodes in the individual output layers:

An integer vector of indices: If the model has only one output layer, the values correspond to the indices of the output nodes, e.g., c(1,3,4) for the first, third and fourth output node. If there are multiple output layers, the indices of the output nodes from the first output layer are considered.
A list of integer vectors of indices: If the method is to be applied to output nodes from different layers, a list can be passed that specifies the desired indices of the output nodes for each output layer. Unwanted output layers have the entry NULL instead of a vector of indices, e.g., list(NULL, c(1,3)) for the first and third output node in the second output layer.
NULL (default): The method is applied to all output nodes in the first output layer but is limited to the first ten as the calculations become more computationally expensive for more output nodes.

output_label

(character, factor, list or NULL)
These values specify the output nodes for which the method is to be applied. Only values that were previously passed with the argument output_names in the converter can be used. In order to allow models with multiple output layers, there are the following possibilities to select the names of the output nodes in the individual output layers:

A character vector or factor of labels: If the model has only one output layer, the values correspond to the labels of the output nodes named in the passed Converter object, e.g., c("a", "c", "d") for the first, third and fourth output node if the output names are c("a", "b", "c", "d"). If there are multiple output layers, the names of the output nodes from the first output layer are considered.
A list of charactor/factor vectors of labels: If the method is to be applied to output nodes from different layers, a list can be passed that specifies the desired labels of the output nodes for each output layer. Unwanted output layers have the entry NULL instead of a vector of labels, e.g., list(NULL, c("a", "c")) for the first and third output node in the second output layer.
NULL (default): The method is applied to all output nodes in the first output layer but is limited to the first ten as the calculations become more computationally expensive for more output nodes.

ignore_last_act

(logical(1))
Set this logical value to include the last activation functions for each output layer, or not (default: TRUE). In practice, the last activation (especially for softmax activation) is often omitted.

verbose

(logical(1))
This logical argument determines whether a progress bar is displayed for the calculation of the method or not. The default value is the output of the primitive R function interactive().

dtype

(character(1))
The data type for the calculations. Use either 'float' for torch_float or 'double' for torch_double.

Method `clone()`

The objects of this class are cloneable with this method.

Usage

IntegratedGradient$clone(deep = FALSE)

Arguments

deep: Whether to make a deep clone.

Examples

#----------------------- Example 1: Torch ----------------------------------
library(torch)

# Create nn_sequential model and data
model <- nn_sequential(
  nn_linear(5, 12),
  nn_relu(),
  nn_linear(12, 2),
  nn_softmax(dim = 2)
)
data <- torch_randn(25, 5)
ref <- torch_randn(1, 5)

# Create Converter
converter <- convert(model, input_dim = c(5))

# Apply method IntegratedGradient
int_grad <- IntegratedGradient$new(converter, data, x_ref = ref)

# You can also use the helper function `run_intgrad` for initializing
# an R6 IntegratedGradient object
int_grad <- run_intgrad(converter, data, x_ref = ref)

# Print the result as a torch tensor for first two data points
get_result(int_grad, "torch.tensor")[1:2]
#> torch_tensor
#> (1,.,.) = 
#>  -0.0934  0.2019
#>   0.0002  0.0000
#>   0.0595 -0.0219
#>   0.2179 -0.1043
#>  -0.0006  0.0016
#> 
#> (2,.,.) = 
#>  -0.1249  0.3973
#>  -0.0081  0.0281
#>  -0.0268  0.0117
#>  -0.0248 -0.0819
#>   0.0185  0.0843
#> [ CPUFloatType{2,5,2} ]

# Plot the result for both classes
plot(int_grad, output_idx = 1:2)


# Plot the boxplot of all datapoints and for both classes
boxplot(int_grad, output_idx = 1:2)


# ------------------------- Example 2: Neuralnet ---------------------------
if (require("neuralnet")) {
  library(neuralnet)
  data(iris)

  # Train a neural network
  nn <- neuralnet((Species == "setosa") ~ Petal.Length + Petal.Width,
    iris,
    linear.output = FALSE,
    hidden = c(3, 2), act.fct = "tanh", rep = 1
  )

  # Convert the model
  converter <- convert(nn)

  # Apply IntegratedGradient with a reference input of the feature means
  x_ref <- matrix(colMeans(iris[, c(3, 4)]), nrow = 1)
  int_grad <- run_intgrad(converter, iris[, c(3, 4)], x_ref = x_ref)

  # Get the result as a dataframe and show first 5 rows
  get_result(int_grad, type = "data.frame")[1:5, ]

  # Plot the result for the first datapoint in the data
  plot(int_grad, data_idx = 1)

  # Plot the result as boxplots
  boxplot(int_grad)
}

# ------------------------- Example 3: Keras -------------------------------
if (require("keras") & keras::is_keras_available()) {
  library(keras)

  # Make sure keras is installed properly
  is_keras_available()

  data <- array(rnorm(10 * 32 * 32 * 3), dim = c(10, 32, 32, 3))

  model <- keras_model_sequential()
  model %>%
    layer_conv_2d(
      input_shape = c(32, 32, 3), kernel_size = 8, filters = 8,
      activation = "softplus", padding = "valid") %>%
    layer_conv_2d(
      kernel_size = 8, filters = 4, activation = "tanh",
      padding = "same") %>%
    layer_conv_2d(
      kernel_size = 4, filters = 2, activation = "relu",
      padding = "valid") %>%
    layer_flatten() %>%
    layer_dense(units = 64, activation = "relu") %>%
    layer_dense(units = 2, activation = "softmax")

  # Convert the model
  converter <- convert(model)

  # Apply the IntegratedGradient method with a zero baseline and n = 20
  # iteration steps
  int_grad <- run_intgrad(converter, data,
    channels_first = FALSE,
    n = 20
  )

  # Plot the result for the first image and both classes
  plot(int_grad, output_idx = 1:2)

  # Plot the pixel-wise median of the results
  plot_global(int_grad, output_idx = 1)
}

#------------------------- Plotly plots ------------------------------------
if (require("plotly")) {
  # You can also create an interactive plot with plotly.
  # This is a suggested package, so make sure that it is installed
  library(plotly)
  boxplot(int_grad, as_plotly = TRUE)
}
#> Warning: The `boxplot()` function is only intended for tabular or signal data. It is
#> called `plot_global()` instead.

References

See also

Super classes

Public fields

Methods

Public methods

Method new()

Usage

Arguments

Method clone()

Usage

Arguments

Examples

Method `new()`

Method `clone()`