R/Convert_neuralnet.R

In innsight: Get the Insights of Your Neural Network

convert_neuralnet_model <- function(model) {

  # Test whether the model has been fitted yet
  if (!("result.matrix" %in% names(model))) {
    stopf("The {.pkg neuralnet} model hasn't been fitted yet!")
  }

  # Get number of best repetition
  if (ncol(model$result.matrix) == 1) {
    best_rep <- 1
  } else {
    best_rep <- which.min(model$result.matrix["error", ])
  }

  # Get the weight matrices and the activation name
  weights <- model$weights[[best_rep]]
  act_name <- attributes(model$act.fct)$type
  if (act_name == "function") {
    stopf("You can't use custom activation functions for a {.pkg neuralnet} model.")
  }

  model_as_list <- list()
  for (i in seq_along(weights)) {
    name <- sprintf("Dense_%s", i)

    # the first row is the bias vector and the rest the weight matrix
    b <- torch_tensor(as.vector(weights[[i]][1, ]))
    w <- torch_tensor(t(matrix(weights[[i]][-1, ], ncol = length(b))))

    # consider the activation of the last layer
    if (i == length(weights) && model$linear.output == TRUE) {
      act_name <- "linear"
    }

    # convert the layer as a list with all important parameters
    # The argument 'input_layers' saves all the previous layers of the current
    # layer. '0' means that the current layer is an input layer. Similarly,
    # the argument 'output_layers' saves all indices of the succeeding layers,
    # whereby '-1' means that the current layer is an output layer
    model_as_list$layers[[name]] <-
      list(
        type = "Dense",
        weight = w,
        bias = b,
        activation_name = act_name,
        dim_in = dim(w)[2],
        dim_out = dim(w)[1],
        input_layers = i - 1,
        output_layers = ifelse(i == length(weights), -1, i + 1)
      )
  }

  # Save other attributes of the model
  model_as_list$input_dim <- list(ncol(model$covariate))
  model_as_list$output_dim <- list(ncol(model$response))
  model_as_list$input_names <- list(list(model$model.list$variables))
  model_as_list$output_names <- list(list(model$model.list$response))
  # Neuralnet implements only sequential models, hence the input node is the
  # first and the output node the last layer
  model_as_list$input_nodes <- c(1)
  model_as_list$output_nodes <- c(i)

  model_as_list
}