CNN

Para treinar uma rede convolucional, temos que definir um objeto do tipo array. Com isso, temos que definir também a entrada do argumento input_shape.

# Base para testar
dim(x_test)

## [1] 4208 1568

x_test_cnn <- array_reshape(x_test, c(nrow(x_test), 28, 28*2, 1))

y_test <- readRDS("y_test.rds")

Treinando a rede…

model_cnn <- keras_model_sequential() %>%
  layer_conv_2d(
    filters = 32,
    kernel_size = c(3, 3),
    activation = 'relu',
    input_shape = c(28,28*2,1)) %>%
  layer_conv_2d(filters = 64,
                kernel_size = c(3, 3),
                activation = 'relu') %>%
  layer_max_pooling_2d(pool_size = c(2, 2)) %>%
  layer_dropout(rate = 0.25) %>%
  layer_flatten() %>%
  # these are the embeddings (activations) we are going to visualize
  layer_dense(units = 784*4, activation = 'relu', name = 'features1') %>%
  layer_dropout(rate = 0.4) %>%
  layer_dense(units = 784*2, activation = 'relu', name = 'features2') %>%
  layer_dropout(rate = 0.2) %>%
  layer_dense(units = 784, activation = 'relu', name = 'features3') %>%
  layer_dropout(rate = 0.1) %>%
  layer_dense(units = 10, activation = 'softmax')

# Compile model
model_cnn %>% compile(
  loss = "categorical_crossentropy",
  optimizer = optimizer_adadelta(),
  metrics = c('accuracy')
)

history <- model_cnn %>% fit(
  x_train_cnn, rbind(y_train,y_train,y_train), 
  epochs = 2, batch_size = 128, 
  validation_split = 0.2
)

Testando…

# Evaluate
model_cnn %>% evaluate(x_test_cnn, y_test)

## $loss
## [1] 0.04116193
## 
## $acc
## [1] 0.9888308