tests/testthat/test-layer-text_vectorization.R

In rstudio/keras: R Interface to 'Keras'

context("layer_text_vectorization")



test_call_succeeds("layer_text_vectorization", {

  if (tensorflow::tf_version() < "2.1")
    skip("TextVectorization requires TF version >= 2.1")

  input <- matrix(c("hello world", "hello world"), ncol = 1)

  layer <- layer_text_vectorization()
  layer %>% adapt(input)
  output <- layer(input)

  expect_s3_class(output, "tensorflow.tensor")
})

test_call_succeeds("layer_text_vectorization", {

  if (tensorflow::tf_version() < "2.1")
    skip("TextVectorization requires TF version >= 2.1")

  x <- matrix(c("hello world", "hello world"), ncol = 1)

  layer <- layer_text_vectorization(output_mode = "binary",
                                    pad_to_max_tokens = FALSE)
  layer %>% adapt(x)

  output <- layer(x)

  expect_s3_class(output, "tensorflow.tensor")
})

test_call_succeeds("can use layer_text_vectorization in a functional model", {

  if (tensorflow::tf_version() < "2.1")
    skip("TextVectorization requires TF version >= 2.1")

  x <- matrix(c("hello world", "hello world"), ncol = 1)

  layer <- layer_text_vectorization()
  layer %>% adapt(x)

  input <- layer_input(shape = 1, dtype = "string")
  output <- layer(input)
  model <- keras_model(input, output)

  pred <- predict(model, x)

})

test_call_succeeds("can set and get the vocabulary of layer_text_vectorization", {

  if (tensorflow::tf_version() < "2.1")
    skip("TextVectorization requires TF version >= 2.1")

  x <- matrix(c("hello world", "hello world"), ncol = 1)

  layer <- layer_text_vectorization()

  # workaround upstream regression, getting an empty vocab throws an exception in 2.5
  if(tf_version() < "2.5")
    layer$get_vocabulary()

  set_vocabulary(layer, vocabulary = c("hello", "world"))

  output <- layer(x)

  vocab <- get_vocabulary(layer)

  expect_s3_class(output, "tensorflow.tensor")
  if (tensorflow::tf_version() < "2.3")
    expect_length(vocab, 2)
  else
    expect_length(vocab, 4) # 0 is used for padding and 1 for unknown.
})


test_call_succeeds("can use layer_text_vectorization", {
  if (tensorflow::tf_version() < "2.1")
    skip("TextVectorization requires TF version >= 2.1")

  x <- matrix(c("hello world", "hello world"), ncol = 1)
  x_ds <- tfdatasets::tensor_slices_dataset(x)

  layer <- layer_text_vectorization()
  layer %>% adapt(x_ds)

  if (tensorflow::tf_version() < "2.3")
    expect_length(get_vocabulary(layer), 2)
  else
    expect_length(get_vocabulary(layer), 4) # 0 is used for padding and 1 for unknown.
})


test_call_succeeds("can create a tf-idf layer", {

  if (tensorflow::tf_version() < "2.1")
    skip("TextVectorization requires TF version >= 2.1")

  num_words <- 10000
  max_length <- 50

  text_vectorization <- layer_text_vectorization(
    max_tokens = num_words,
    output_mode = if(tf_version() >= "2.6") "tf_idf" else "tf-idf"
  )
  with(tf$device("/cpu:0"), {
    text_vectorization %>% adapt(c("hello world", "hello"))
  })
  x <- text_vectorization(matrix(c("hello"), ncol = 1))

  expect_s3_class(x, "tensorflow.tensor")

})



test_call_succeeds("get_vocabulary() returns R character vector", {

  text_vectorization <- layer_text_vectorization()
  with(tf$device("/cpu:0"), {
    text_vectorization %>% adapt(c("hello world", "hello"))
  })
  vocab <- get_vocabulary(text_vectorization)

  expect_type(vocab, "character")
  expect_contains(vocab, c("hello", "world"))

})