Building a simple generative adversarial net with neuralnet in R - r

Question

I am trying to reproduce the example given in Goodfellow, I. et al.: Generative Adversarial Nets
The pseudocode is given on page 4 as "Algorithm 1". I am trying to rebuild it with the neuralnet package in R:
library(neuralnet)
train_iter <- 10
steps <- 1
m <- 100
# initialize D and G
z <- sort(runif(m))
x <- sort(rnorm(m))
data <- cbind(z, x)
D <- neuralnet( , data = data, hidden = 11) # unclear how to define formula
G <- neuralnet(x ~ z, data = data, hidden = 11)
for (i in 1:train_iter) {
for (k in 1:steps) {
z <- sort(runif(m))
x <- sort(rnorm(m))
data <- cbind(z, x)
err_fct_d <- function(x, z) {
-log(compute(D, x)$net.result + log(1 - compute(D, compute(G, z)$net.result)$net.result))
}
D <- neuralnet( , data = data, hidden = 11, err.fct = err_fct_d, startweights = D$weights) # unclear how to define formula
}
z <- sort(runif(m))
data <- cbind(z, x)
err_fct_g <- function(x, z) {
log(1 - compute(D, compute(G, z)$net.result)$net.result)
}
G <- neuralnet(x ~ z, data = data, hidden = 11, err.fct = err_fct_g, startweights = G$weights)
}
My questions
My first question is whether it is possible to use the neuralnet package with these customized error functions in the above way at all. My second question concerns the discriminator network: I don't know how to train it, i.e. how to define the formula part of the neuralnet function.

Answer 1

Unfortunately this does not work out of the box because err.fct has to be an analytically differentiable function and the compute function prevents this.
Edit: I contacted the author of the package and he wrote:
I checked your GAN example and I think that is not possible with neuralnet because the way error functions are handled is not flexible enough. [...]