def test_network_invocation(self):
hidden_size = 32
sequence_length = 21
vocab_size = 57
num_types = 7
# Create a small TransformerEncoder for testing.
test_network = albert_transformer_encoder.AlbertTransformerEncoder(
vocab_size=vocab_size,
embedding_width=8,
hidden_size=hidden_size,
<<<<<<< HEAD
sequence_length=sequence_length,
=======
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types)
self.assertTrue(
test_network._position_embedding_layer._use_dynamic_slicing)
# Create the inputs (note that the first dimension is implicit).
word_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
mask = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
type_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
data, pooled = test_network([word_ids, mask, type_ids])
# Create a model based off of this network:
model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
# Invoke the model. We can't validate the output data here (the model is too
# complex) but this will catch structural runtime errors.
batch_size = 3
word_id_data = np.random.randint(
vocab_size, size=(batch_size, sequence_length))
mask_data = np.random.randint(2, size=(batch_size, sequence_length))
type_id_data = np.random.randint(
num_types, size=(batch_size, sequence_length))
_ = model.predict([word_id_data, mask_data, type_id_data])
# Creates a TransformerEncoder with max_sequence_length != sequence_length
max_sequence_length = 128
test_network = albert_transformer_encoder.AlbertTransformerEncoder(
vocab_size=vocab_size,
embedding_width=8,
hidden_size=hidden_size,
<<<<<<< HEAD
sequence_length=sequence_length,
=======
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
max_sequence_length=max_sequence_length,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types)
self.assertTrue(test_network._position_embedding_layer._use_dynamic_slicing)
model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
_ = model.predict([word_id_data, mask_data, type_id_data])
def test_serialize_deserialize(self):
tf.keras.mixed_precision.experimental.set_policy("mixed_float16")
# Create a network object that sets all of its config options.
kwargs = dict(vocab_size=100,
embedding_width=8,
hidden_size=32,
num_layers=3,
num_attention_heads=2,
max_sequence_length=21,
type_vocab_size=12,
intermediate_size=1223,
activation="relu",
dropout_rate=0.05,
attention_dropout_rate=0.22,
initializer="glorot_uniform")
network = albert_transformer_encoder.AlbertTransformerEncoder(**kwargs)
expected_config = dict(kwargs)
expected_config["activation"] = tf.keras.activations.serialize(
tf.keras.activations.get(expected_config["activation"]))
expected_config["initializer"] = tf.keras.initializers.serialize(
tf.keras.initializers.get(expected_config["initializer"]))
self.assertEqual(network.get_config(), expected_config)
# Create another network object from the first object's config.
new_network = (
albert_transformer_encoder.AlbertTransformerEncoder.from_config(
network.get_config()))
# Validate that the config can be forced to JSON.
_ = new_network.to_json()
# If the serialization was successful, the new config should match the old.
self.assertAllEqual(network.get_config(), new_network.get_config())
def test_network_creation(self, expected_dtype):
hidden_size = 32
sequence_length = 21
kwargs = dict(
vocab_size=100,
hidden_size=hidden_size,
<<<<<<< HEAD
sequence_length=sequence_length,
=======
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
num_attention_heads=2,
num_layers=3)
if expected_dtype == tf.float16:
tf.keras.mixed_precision.experimental.set_policy("mixed_float16")
# Create a small TransformerEncoder for testing.
test_network = albert_transformer_encoder.AlbertTransformerEncoder(**kwargs)
# Create the inputs (note that the first dimension is implicit).
word_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
mask = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
type_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
data, pooled = test_network([word_ids, mask, type_ids])
expected_data_shape = [None, sequence_length, hidden_size]
expected_pooled_shape = [None, hidden_size]
self.assertAllEqual(expected_data_shape, data.shape.as_list())
self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
# If float_dtype is set to float16, the data output is float32 (from a layer
# norm) and pool output should be float16.
self.assertEqual(tf.float32, data.dtype)
self.assertEqual(expected_dtype, pooled.dtype)
# ALBERT has additonal 'embedding_hidden_mapping_in' weights and
# it shares transformer weights.
self.assertNotEmpty(
[x for x in test_network.weights if "embedding_projection/" in x.name])
self.assertNotEmpty(
[x for x in test_network.weights if "transformer/" in x.name])
self.assertEmpty(
[x for x in test_network.weights if "transformer/layer" in x.name])
def test_network_creation(self, expected_dtype, float_dtype=None):
hidden_size = 32
sequence_length = 21
kwargs = dict(vocab_size=100,
hidden_size=hidden_size,
sequence_length=sequence_length,
num_attention_heads=2,
num_layers=3)
if float_dtype is not None:
kwargs["float_dtype"] = float_dtype
# Create a small TransformerEncoder for testing.
test_network = albert_transformer_encoder.AlbertTransformerEncoder(
**kwargs)
# Create the inputs (note that the first dimension is implicit).
word_ids = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
mask = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
type_ids = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
data, pooled = test_network([word_ids, mask, type_ids])
expected_data_shape = [None, sequence_length, hidden_size]
expected_pooled_shape = [None, hidden_size]
self.assertAllEqual(expected_data_shape, data.shape.as_list())
self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
self.assertEqual(expected_dtype, data.dtype)
self.assertEqual(expected_dtype, pooled.dtype)
# ALBERT has additonal 'embedding_hidden_mapping_in' weights and
# it shares transformer weights.
self.assertNotEmpty([
x for x in test_network.weights
if "embedding_projection/" in x.name
])
self.assertNotEmpty(
[x for x in test_network.weights if "transformer/" in x.name])
self.assertEmpty(
[x for x in test_network.weights if "transformer/layer" in x.name])
def test_network_invocation(self):
hidden_size = 32
sequence_length = 21
vocab_size = 57
num_types = 7
# Create a small TransformerEncoder for testing.
test_network = albert_transformer_encoder.AlbertTransformerEncoder(
vocab_size=vocab_size,
embedding_width=8,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types)
# Create the inputs (note that the first dimension is implicit).
word_ids = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
mask = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
type_ids = tf.keras.Input(shape=(sequence_length, ), dtype=tf.int32)
data, pooled = test_network([word_ids, mask, type_ids])
# Create a model based off of this network:
model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
# Invoke the model. We can't validate the output data here (the model is too
# complex) but this will catch structural runtime errors.
batch_size = 3
word_id_data = np.random.randint(vocab_size,
size=(batch_size, sequence_length))
mask_data = np.random.randint(2, size=(batch_size, sequence_length))
type_id_data = np.random.randint(num_types,
size=(batch_size, sequence_length))
list_outputs = model.predict([word_id_data, mask_data, type_id_data])
# Creates a TransformerEncoder with max_sequence_length != sequence_length
max_sequence_length = 128
test_network = albert_transformer_encoder.AlbertTransformerEncoder(
vocab_size=vocab_size,
embedding_width=8,
hidden_size=hidden_size,
max_sequence_length=max_sequence_length,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types)
model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
_ = model.predict([word_id_data, mask_data, type_id_data])
# Tests dictionary outputs.
test_network_dict = albert_transformer_encoder.AlbertTransformerEncoder(
vocab_size=vocab_size,
embedding_width=8,
hidden_size=hidden_size,
max_sequence_length=max_sequence_length,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types,
dict_outputs=True)
_ = test_network_dict([word_ids, mask, type_ids])
test_network_dict.set_weights(test_network.get_weights())
list_outputs = test_network([word_id_data, mask_data, type_id_data])
dict_outputs = test_network_dict(
dict(input_word_ids=word_id_data,
input_mask=mask_data,
input_type_ids=type_id_data))
self.assertAllEqual(list_outputs[0], dict_outputs["sequence_output"])
self.assertAllEqual(list_outputs[1], dict_outputs["pooled_output"])
