def test_model_input_output_shape(self):
vocab_size = 5
sequence_length = 7
model = char_prediction_models.create_recurrent_model(
vocab_size, sequence_length)
self.assertEqual(model.input_shape, (None, sequence_length))
self.assertEqual(model.output_shape, (None, sequence_length, vocab_size))
def model_fn() -> model.Model:
return keras_utils.from_keras_model(
keras_model=char_prediction_models.create_recurrent_model(
vocab_size=VOCAB_LENGTH, sequence_length=sequence_length),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
input_spec=task_datasets.element_type_structure,
metrics=[
keras_metrics.NumTokensCounter(masked_tokens=[pad_token]),
keras_metrics.MaskedCategoricalAccuracy(masked_tokens=[pad_token])
])
def test_no_mask_zero_results_in_correct_mask(self):
mask_model = char_prediction_models.create_recurrent_model(
vocab_size=3, sequence_length=3, mask_zero=False)
data = tf.constant([[0, 1, 1]])
output_mask = mask_model.compute_mask(data, mask=None)
self.assertIsNone(output_mask)
def test_create_recurrent_model_raises_on_nonpositive_sequence_length(self):
with self.assertRaisesRegex(ValueError,
'sequence_length must be a positive integer'):
char_prediction_models.create_recurrent_model(
vocab_size=3, sequence_length=0)