tests/testthat/test-initializers.R

In keras: R Interface to 'Keras'

context("initializers")



test_initializer <- function(name, required_version = NULL) {
  initializer_fn <- eval(parse(text = paste0("initializer_", name)))
  test_call_succeeds(name, required_version = required_version, {
    keras_model_sequential() %>%
      layer_dense(32, input_shape = c(32,32), kernel_initializer = initializer_fn()) %>%
      compile(
        optimizer = 'sgd',
        loss='binary_crossentropy',
        metrics='accuracy'
      )
  })
}


test_initializer("zeros")
test_initializer("ones")
test_initializer("constant")
test_initializer("random_normal")
test_initializer("random_uniform")
test_initializer("truncated_normal")
test_initializer("variance_scaling")
test_initializer("orthogonal")
if (is_keras_available())
  if (is_keras_implementation() || (is_tensorflow_implementation() && (tensorflow::tf_version() <= "1.1")))
    test_initializer("identity") # don't know why this test fails on v1.2
test_initializer("glorot_normal")
test_initializer("glorot_uniform")
test_initializer("he_uniform")
test_initializer("he_normal")
test_initializer("lecun_uniform")
test_initializer("lecun_normal", required_version = "2.0.6")