R/gradient_ascent.R

In primaj/deepDreameR: Using Keras to Demonstrate Deep Dream on Images

eval_loss_and_grads <- function(x, loss_and_grads) {
  outs <- loss_and_grads(list(x))
  loss_value <- outs[[1]]
  grad_values <- outs[[2]]
  list(loss_value, grad_values)
}

gradient_ascent <- function(x, loss_and_grads,  iterations, step, max_loss = NULL) {
  for (i in 1:iterations) {
    c(loss_value, grad_values) %<-% eval_loss_and_grads(x, loss_and_grads)
    if (!is.null(max_loss) && loss_value > max_loss)
      break
    cat("...Loss value at", i, ":", loss_value, "\n")
    x <- x + (step * grad_values)
  }
  x
}