HyperModel subclass

Recently, the reticulate library provided with one of the most anticipating functionality — ability to write a python class in R.

We could use a HyperModel subclass instead of a model-building function. So, this makes it easy to share and reuse hypermodels.

A HyperModel subclass only needs to implement a build(self, hp) method. And, again one should return a compiled model inside a build function.

MyHyperModel <- reticulate::PyClass(
  "HyperModel",
  
  inherit = kerastuneR::HyperModel_class(),
  
  list(
    
    `__init__` = function(self, num_classes) {
      
      self$num_classes = num_classes
      NULL
    },
    
    build = function(self,hp) {
      model = keras_model_sequential() 
      model %>% layer_dense(units = hp$Int('units',
                                           min_value=32L,
                                           max_value=512L,
                                           step=32L),
                            activation='relu') %>% 
        layer_dense(as.integer(self$num_classes), activation='softmax') %>% 
        compile(
          optimizer= tf$keras$optimizers$Adam(
            hp$Choice('learning_rate',
                      values=c(1e-2, 1e-3, 1e-4))),
          loss='categorical_crossentropy',
          metrics='accuracy')
    }
  )
)

Full code

# generate some data

x_data <- matrix(data = runif(500,0,1),nrow = 50,ncol = 5)
y_data <-  ifelse(runif(50,0,1) > 0.6, 1L,0L) %>% as.matrix()

x_data2 <- matrix(data = runif(500,0,1),nrow = 50,ncol = 5)
y_data2 <-  ifelse(runif(50,0,1) > 0.6, 1L,0L) %>% as.matrix()

# subclass

MyHyperModel <- reticulate::PyClass(
  "HyperModel",
  
  inherit = kerastuneR::HyperModel_class(),
  
  list(
    
    `__init__` = function(self, num_classes) {
      
      self$num_classes = num_classes
      NULL
    },
    
    build = function(self,hp) {
      model = keras_model_sequential() 
      model %>% layer_dense(units = hp$Int('units',
                                           min_value=32L,
                                           max_value=512L,
                                           step=32L),
                            activation='relu') %>% 
        layer_dense(as.integer(self$num_classes), activation='softmax') %>% 
        compile(
          optimizer= tf$keras$optimizers$Adam(
            hp$Choice('learning_rate',
                      values=c(1e-2, 1e-3, 1e-4))),
          loss='categorical_crossentropy',
          metrics='accuracy')
    }
  )
)

# Random Search

hypermodel = MyHyperModel(num_classes = 10)

tuner = RandomSearch(
    hypermodel,
    objective = 'val_accuracy',
    max_trials = 10,
    directory = 'my_dir',
    project_name = 'helloworld')

# Run

tuner %>% fit_tuner(x_data,y_data,
                    epochs = 5, 
                    validation_data = list(x_data2, y_data2))