tests/testthat/test-freeze.R

In keras: R Interface to 'Keras'

context("freeze")




define_freeze_model <- function() {
  model <- keras_model_sequential()
  model %>%
    layer_dense(32, input_shape = 784, kernel_initializer = initializer_ones(), name = "input") %>%
    layer_activation('relu', name = "relu_activation") %>%
    layer_dense(10, name = "dense") %>%
    layer_activation('softmax', name = "softmax_activation")
  model
}

test_succeeds("freeze_weights can freeze an entire model", {
  model <- define_freeze_model()
  freeze_weights(model)
  expect_length(model$trainable_weights, 0)
})

test_succeeds("model can be unfrozen after freezing", {
  model <- define_freeze_model()
  freeze_weights(model)
  unfreeze_weights(model)
  expect_length(model$trainable_weights, 4)
})

test_succeeds("freeze_weights can work on indexes", {
  model <- define_freeze_model()
  freeze_weights(model, from = 2, to = 3)
  expect_length(model$trainable_weights, 2)
})

test_succeeds("freeze_weights can work on names", {
  model <- define_freeze_model()
  freeze_weights(model, from = "dense")
  expect_length(model$trainable_weights, 2)
})


test_succeeds("freeze_weights which", {
  # model <- define_freeze_model()

  model <- keras_model_sequential(input_shape = 32) %>%
    layer_dense(32, name = "input") %>%
    layer_activation('relu', name = "relu_activation") %>%
    layer_dense(10, name = "dense1") %>%
    layer_dense(10, name = "dense2") %>%
    layer_dense(10, name = "dense3") %>%
    layer_dense(10, name = "dense4") %>%
    layer_dense(10, name = "dense5") %>%
    layer_activation('softmax', name = "softmax_activation")

  freeze_weights(model, from = -2)
  expect_length(model$non_trainable_weights, 2) # kernel + bias from last dense layer
  freeze_weights(model, from = -3)
  expect_length(model$non_trainable_weights, 4)

  freeze_weights(model, from = -4, to = -3)
  expect_length(model$non_trainable_weights, 4)

  freeze_weights(model, which = ~ grepl("^dense", .x$name))
  expect_length(model$trainable_weights, 2) # just the first dense layer

  unfreeze_weights(model, which = ~ grepl("^dense", .x$name))
  expect_length(model$non_trainable_weights, 2) # just the first dense layer

  unfreeze_weights(model, which = c(6, 7))
  expect_length(model$trainable_weights, 4) # just the first dense layer

  unfreeze_weights(model, which = c("dense1", "dense2"))
  expect_length(model$trainable_weights, 4) # just the first dense layer

  freeze_weights(model, which = c(6, 7))
  expect_length(model$non_trainable_weights, 4) # just the first dense layer

  freeze_weights(model, which = c("dense1", "dense2"))
  expect_length(model$non_trainable_weights, 4) # just the first dense layer

  freeze_weights(model, which = list(3, "dense3"))
  expect_length(model$non_trainable_weights, 4) # just the first dense layer
})