library(keras)
library(mrsdeploy)

#Keras objects cannot be saved and restored across R sessions, so load a keras model and convert it to a r model
kerasModel = load_model_hdf5('my_model.h5')
rModel = serialize_model(kerasModel, include_optimizer = TRUE)

kerasFunc <- function(x) {
  library(keras)
  kerasModel = unserialize_model(rModel, custom_objects = NULL, compile = TRUE)
  prediction <- predict_classes(kerasModel, x)
  answer = as.list(prediction)
}

serviceName <- paste0("kerasService", round(as.numeric(Sys.time()), 0))
api <- publishService(
  serviceName,
  code = kerasFunc,
  model = rModel,
  inputs = list(x="matrix"),
  outputs = list(answer="vector"),
  v = "v1.0.0"
)

swagger <- api$swagger()
cat(swagger, file = "swagger.json", append = FALSE)

# Load test data
mnist <- dataset_mnist()
x_test <- mnist$test$x
x_test <- array_reshape(x_test, c(nrow(x_test), 784))
x_test <- x_test / 255
x = matrix(c(x_test[1,]), nrow=1, ncol=784)

#consume the keras model api
result<-api$kerasFunc(x)
#output is the predicted number
print(result$output("answer"))