NMIST1

####Deep Learning with R Chapter 1 ####

####Loading in the MNIST dataset into Keras####
library(keras)
mnist <- dataset_mnist() #Note - there are several other data sets in Keras
train_images <- mnist$train$x
train_labels <- mnist$train$y
test_images <- mnist$test$x
test_labels <- mnist$test$y

#characteristics of our data sets#
dim(train_images)
class(train_images)
typeof(train_images)

#### View the four data sets ####
View(train_images)
View(train_labels)
View(test_images)
View(test_labels)

#### Setting up our first network ####
network <- keras_model_sequential() %>%
  layer_dense(units = 512, activation = "relu", input_shape = c(28 * 28)) %>%
  layer_dense(units = 10, activation = "softmax")

summary(network) #a look at our network

#### Setting up an optimizer, loss function, and our metrics ####
network %>% compile(
  optimizer = "rmsprop",
  loss = "categorical_crossentropy",
  metrics = c("accuracy")
)

#### Prepare the image data ####
train_images <- array_reshape(train_images, c(60000, 28 * 28))
train_images <- train_images / 255
test_images <- array_reshape(test_images, c(10000, 28 * 28))
test_images <- test_images / 255

#### Preparing the labels ####
train_labels <- to_categorical(train_labels)
test_labels <- to_categorical(test_labels)

#View(train_images[1:10])
#View(test_labels)

#### Train the network ####
network %>% fit(train_images, train_labels, epochs = 5, batch_size = 128) #notice the call to the fit function

#### Evaluate how the network did on the *test* data
metrics <- network %>% evaluate(test_images, test_labels)
metrics

#### How to generate predictions ####
network %>% predict_classes(test_images[1:10,])