Fit mxnet on a spiral

spiralmxnet.R

library("mlbench")
library("ggplot2")
library("mxnet")


plot_mxmodel <- function(model, data) {
  x <- seq(from = min(data), to = max(data), length.out = 500)
  d2 <- as.matrix(expand.grid(x, x))

  mx_pred <- predict(model, d2)
  pred_label <- t(mx_pred)

  df_d2 <- expand.grid(x = x, y = x)

  df_d2$p <- pred_label

  ggplot() +
    geom_tile(data = df_d2, aes(x, y, fill = p)) +
    scale_fill_continuous(low = "sandybrown", high = "steelblue2")  +
    geom_point(data = data.frame(data),
               aes(x = X1, y = X2,colour = factor(train_y)), size = 1) +
    scale_colour_manual(values = c("red", "blue"))
}


spiral_data <- mlbench.spirals(1000, 1.5, 0)
plot(spiral_data)

train_x <- data.matrix(spiral_data$x)
train_y <- as.numeric(spiral_data$classes) - 1

mx.set.seed(0)
mx_model <- mx.mlp(train_x,
                   train_y,
                   num.round = 1000,
                   hidden_node = c(7, 7, 7),
                   activation = "relu",
                   out_activation = "logistic",
                   out_node = 1,
                   array.batch.size = 50,
                   learning.rate = 0.03,
                   momentum = 0.1,
                   array.layout = "rowmajor",
                   initializer = mx.init.normal(1),
                   eval.metric = mx.metric.rmse)

plot_mxmodel(mx_model, train_x)
# predict(mx_model, train_x) %>% t %>% summary