ascend: Perform A Single Gradient Ascent
In tpq/caress: Caress, a Gentler Introduction to Keras

Description Usage Arguments Value Examples

Perform A Single Gradient Ascent

1	ascend(model, input, loss, gradient, lr = 1)

`model`	A keras model.
`input`	The input to perturb.
`loss`	The output from `get_layer_loss`.
`gradient`	The output from `get_layer_gradient`.
`lr`	The learning rate used to update input.

The perturbed input.

library(keras)
library(caress)
data(iris)
x <- as.matrix(iris[,1:4])
y <- to_categorical(as.numeric(iris[,5])-1)
input <- from_input(x, name = "input")
output <- input %>%
  layer_dense(units = 2, activation = "linear", name = "middle") %>%
  to_output(y, name = "output")
model <- prepare(input, output)
build(model, x, y, epochs = 10, batch_size = 8)
loss <- get_layer_loss(model, "middle")
gradient <- get_layer_gradient(model, loss)
x <- matrix(runif(4), 1, 4)
for(i in 1:20){
  x <- ascend(model, x, loss, gradient)
}