def test_layer_creation(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size, embedding_width=embedding_width)
# Create a 2-dimensional input (the first dimension is implicit).
sequence_length = 23
input_tensor = tf.keras.Input(shape=(sequence_length), dtype=tf.int32)
output_tensor = test_layer(input_tensor)
# The output should be the same as the input, save that it has an extra
# embedding_width dimension on the end.
expected_output_shape = [None, sequence_length, embedding_width]
self.assertEqual(expected_output_shape, output_tensor.shape.as_list())
self.assertEqual(output_tensor.dtype, tf.float32)
def test_layer_invocation(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size, embedding_width=embedding_width)
# Create a 2-dimensional input (the first dimension is implicit).
sequence_length = 23
input_tensor = tf.keras.Input(shape=(sequence_length), dtype=tf.int32)
output_tensor = test_layer(input_tensor)
# Create a model from the test layer.
model = tf.keras.Model(input_tensor, output_tensor)
# Invoke the model on test data. We can't validate the output data itself
# (the NN is too complex) but this will rule out structural runtime errors.
batch_size = 3
input_data = np.random.randint(vocab_size,
size=(batch_size, sequence_length))
output = model.predict(input_data)
self.assertEqual(tf.float32, output.dtype)
