Bayesian Optimization

Subclassing Tuner for Custom Training Loops

The Tuner class at Tuner_class() can be subclassed to support advanced uses such as:

• Custom training loops (GANs, reinforement learning, etc.)
• Adding hyperparameters outside of the model builing function (preprocessing, data augmentation, test time augmentation, etc.)

Understanding the search process.

tuner %>% fit_tuner() can be passed any arguments. These arguments will be passed directly to Tuner$run_trial, along with a Trial object that contains information about the current trial, including hyperparameters and the status of the trial. Typically, Tuner$run_trial is the only method that users need to override when subclassing Tuner.

Thanks to Daniel Falbel from RStudio, the Bayesian Optimization example was successfully adapted.

library(keras)
library(tensorflow)
library(dplyr)
library(tfdatasets)
library(kerastuneR)
library(reticulate)


conv_build_model = function(hp) {
  'Builds a convolutional model.'
  inputs = tf$keras$Input(shape=c(28L, 28L, 1L))
  
  x = inputs
  
  for (i in 1:hp$Int('conv_layers', 1L, 3L, default=3L)) {
    x = tf$keras$layers$Conv2D(filters = hp$Int(paste('filters_', i, sep = ''), 4L, 32L, step=4L, default=8L),
                               kernel_size = hp$Int(paste('kernel_size_', i, sep = ''), 3L, 5L),
                               activation ='relu',
                               padding='same')(x)
    if (hp$Choice(paste('pooling', i, sep = ''), c('max', 'avg')) == 'max') {
      x = tf$keras$layers$MaxPooling2D()(x)
    } else {
      x = tf$keras$layers$AveragePooling2D()(x)
    }
    x = tf$keras$layers$BatchNormalization()(x) 
    x =  tf$keras$layers$ReLU()(x)
    
  }
  if (hp$Choice('global_pooling', c('max', 'avg')) == 'max') {
    x =  tf$keras$layers$GlobalMaxPooling2D()(x)
  } else {
    x = tf$keras$layers$GlobalAveragePooling2D()(x)
  }
  
  outputs = tf$keras$layers$Dense(10L, activation='softmax')(x)
  model = tf$keras$Model(inputs, outputs)
  optimizer = hp$Choice('optimizer', c('adam', 'sgd'))
  model %>% compile(optimizer, loss='sparse_categorical_crossentropy', metrics='accuracy')
  return(model)
}

MyTuner = PyClass(
  'Tuner',
  inherit = Tuner_class(),
  list(
    run_trial = function(self, trial, train_ds){
      hp = trial$hyperparameters
      train_ds = train_ds$batch(hp$Int('batch_size', 32L, 128L, step=32L, default=64L))
      model = self$hypermodel$build(trial$hyperparameters)
      lr = hp$Float('learning_rate', 1e-4, 1e-2, sampling='log', default=1e-3)
      optimizer = tf$keras$optimizers$Adam(lr)
      epoch_loss_metric = tf$keras$metrics$Mean()
      
      
      run_train_step = function(data){
        images = data[[1]]
        labels = data[[2]]
        
        
        with (tf$GradientTape() %as% tape,{
          logits = model(images)
          loss = tf$keras$losses$sparse_categorical_crossentropy(labels, logits)
          if(length(model$losses) > 0){
            loss = loss + tf$math$add_n(model$losses)
          }
          gradients = tape$gradient(loss, model$trainable_variables)
        })
        optimizer$apply_gradients(purrr::transpose(list(gradients, model$trainable_variables)))
        epoch_loss_metric$update_state(loss)
        loss
      }
      
      for (epoch in 1:1) {
        print(paste('Epoch',epoch))
        self$on_epoch_begin(trial, model, epoch, logs= list())
        intializer = make_iterator_one_shot(train_ds)
        
        for (batch in 1:length(iterate(train_ds))) {
          
          init_next = iter_next(intializer)
          
          self$on_batch_begin(trial, model, batch, logs=list())
          batch_loss = as.numeric(run_train_step(init_next))
          self$on_batch_end(trial, model, batch, logs=list(paste('loss', batch_loss)))
          
          if (batch %% 100L == 0L){
            loss = epoch_loss_metric$result()$numpy()
            print(paste('Batch',batch,  'Average loss', loss))
          }
        }
        
        epoch_loss = epoch_loss_metric$result()$numpy()
        self$on_epoch_end(trial, model, epoch, logs=list('loss'= epoch_loss))
        epoch_loss_metric$reset_states()
      }
    }
  )
)


main = function () {
  tuner = MyTuner(
    oracle=BayesianOptimization(
      objective=Objective(name='loss', direction = 'min'),
      max_trials=1),
    hypermodel=conv_build_model,
    directory='results2',
    project_name='mnist_custom_training2')
  
  mnist_data = dataset_fashion_mnist()
  c(mnist_train, mnist_test) %<-%  mnist_data
  rm(mnist_data)
  
  mnist_train$x = tf$dtypes$cast(mnist_train$x, 'float32') / 255.
  
  mnist_train$x = keras::k_reshape(mnist_train$x,shape = c(6e4,28,28,1))
  
  mnist_train = tensor_slices_dataset(mnist_train) %>% dataset_shuffle(1e3)
    
    tuner %>% fit_tuner(train_ds = mnist_train)
    
    best_model = tuner %>% get_best_models(1L)
  
}

main()