library(keras)
library(densenet)

input_img <- layer_input(shape = c(28, 28, 1))
model <- application_densenet(input_tensor = input_img, classes = 10L)

model %>% compile(
  optimizer = "adam",
  loss = "categorical_crossentropy",
  metrics = "accuracy"
)

mnist <- dataset_mnist()

dim(mnist$train$x) <- c(dim(mnist$train$x), 1)
dim(mnist$test$x) <- c(dim(mnist$test$x), 1)

y <- sapply(0:9, function(x) as.numeric(x == mnist$train$y))
y_test <- sapply(0:9, function(x) as.numeric(x == mnist$test$y))


model %>% fit(
  x = mnist$train$x,
  y = y,
  batch_size = 32,
  validation_data = list(mnist$test$x, y_test)
)