5-poissonCNN.R

# Fix structure for 2d CNN
train_array <- t(trainData$X)
dim(train_array) <- c(50, 50, nrow(trainData$X), 1)
# Reorder dimensions
train_array <- aperm(train_array, c(3,1,2,4))

test_array <- t(testData)
dim(test_array) <- c(50, 50, nrow(testData), 1)
# Reorder dimensions
test_array <- aperm(test_array, c(3,1,2,4))

# Check cat again
testCat <- train_array[2,,,]
image(t(apply(testCat, 2, rev)), col = gray.colors(12),
      axes = F)

# Build CNN model
model <- keras_model_sequential() 
model %>% 
      layer_conv_2d(kernel_size = c(3, 3), filter = 32,
                    activation = "relu", padding = "same",
                    input_shape = c(50, 50, 1),
                    data_format = "channels_last") %>%
      layer_conv_2d(kernel_size = c(3, 3), filter = 32,
                    activation = "relu", padding = "valid") %>%
      layer_max_pooling_2d(pool_size = 2) %>%
      layer_dropout(rate = 0.25) %>%
      
      layer_conv_2d(kernel_size = c(3, 3), filter = 64, strides = 2,
                    activation = "relu", padding = "same") %>%
      layer_conv_2d(kernel_size = c(3, 3), filter = 64,
                    activation = "relu", padding = "valid") %>%
      layer_max_pooling_2d(pool_size = 2) %>%
      layer_dropout(rate = 0.25) %>%
      
      layer_flatten() %>%
      layer_dense(units = 50, activation = "relu") %>% 
      layer_dropout(rate = 0.25) %>%
      layer_dense(units = 1, activation = "sigmoid")

summary(model)

model %>% compile(
      loss = 'binary_crossentropy',
      optimizer = "adam",
      metrics = c('accuracy')
)

history <- model %>% fit(
      x = train_array, y = as.numeric(trainData$y), 
      epochs = 30, batch_size = 100, 
      validation_split = 0.2
)

plot(history)