keras training is too slow

[berndbischl]

consider this:

set.seed(123)
library(keras)
load_all()

env = mountainCar()
preprocessState = function(state) {
  state[1, , drop = FALSE]
}
m = keras_model_sequential()
m %>% layer_dense(input_shape = 2,   activation = "relu",   units = 10L)
m %>% layer_dense(input_shape = 10L, activation = "linear", units = env$n.actions)

res = qlearning(env, fun.approx = "neural.network", model = m, preprocessState = preprocessState, 
  n.episodes = 5)

this takes minutes just to sweep through one episode?

[berndbischl]

actually it doesnt seem to return at all?

[harryprince]

@berndbischl mxnet might be a better choice for training