def test_network_invocation(self, output_range, out_seq_len):
hidden_size = 32
sequence_length = 21
vocab_size = 57
num_types = 7
# Create a small BertEncoder for testing.
test_network = bert_encoder.BertEncoder(vocab_size=vocab_size,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types,
output_range=output_range)
# 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))
outputs = model.predict([word_id_data, mask_data, type_id_data])
self.assertEqual(outputs[0].shape[1], out_seq_len)
# Creates a BertEncoder with max_sequence_length != sequence_length
max_sequence_length = 128
test_network = bert_encoder.BertEncoder(
vocab_size=vocab_size,
hidden_size=hidden_size,
max_sequence_length=max_sequence_length,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types)
data, pooled = test_network([word_ids, mask, type_ids])
model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
outputs = model.predict([word_id_data, mask_data, type_id_data])
self.assertEqual(outputs[0].shape[1], sequence_length)
# Creates a BertEncoder with embedding_width != hidden_size
test_network = bert_encoder.BertEncoder(
vocab_size=vocab_size,
hidden_size=hidden_size,
max_sequence_length=max_sequence_length,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types,
embedding_width=16)
data, pooled = test_network([word_ids, mask, type_ids])
model = tf.keras.Model([word_ids, mask, type_ids], [data, pooled])
outputs = model.predict([word_id_data, mask_data, type_id_data])
self.assertEqual(outputs[0].shape[-1], hidden_size)
self.assertTrue(hasattr(test_network, "_embedding_projection"))
def test_network_creation(self):
hidden_size = 32
sequence_length = 21
# Create a small BertEncoder for testing.
test_network = bert_encoder.BertEncoder(
vocab_size=100,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3)
# 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])
self.assertIsInstance(test_network.transformer_layers, list)
self.assertLen(test_network.transformer_layers, 3)
self.assertIsInstance(test_network.pooler_layer, tf.keras.layers.Dense)
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())
# The default output dtype is float32.
self.assertAllEqual(tf.float32, data.dtype)
self.assertAllEqual(tf.float32, pooled.dtype)
test_network_dict = bert_encoder.BertEncoder(
vocab_size=100,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
dict_outputs=True)
# Create the inputs (note that the first dimension is implicit).
inputs = dict(
input_word_ids=word_ids, input_mask=mask, input_type_ids=type_ids)
_ = test_network_dict(inputs)
test_network_dict.set_weights(test_network.get_weights())
batch_size = 2
vocab_size = 100
num_types = 2
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 = 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"])
def test_all_encoder_outputs_network_creation(self):
hidden_size = 32
sequence_length = 21
# Create a small BertEncoder for testing.
test_network = bert_encoder.BertEncoder(
vocab_size=100,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
return_all_encoder_outputs=True)
# 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)
all_encoder_outputs, pooled = test_network([word_ids, mask, type_ids])
expected_data_shape = [None, sequence_length, hidden_size]
expected_pooled_shape = [None, hidden_size]
self.assertLen(all_encoder_outputs, 3)
for data in all_encoder_outputs:
self.assertAllEqual(expected_data_shape, data.shape.as_list())
self.assertAllEqual(expected_pooled_shape, pooled.shape.as_list())
# The default output dtype is float32.
self.assertAllEqual(tf.float32, all_encoder_outputs[-1].dtype)
self.assertAllEqual(tf.float32, pooled.dtype)
def test_v2_network_creation_with_float16_dtype(self):
hidden_size = 32
sequence_length = 21
tf.keras.mixed_precision.set_global_policy("mixed_float16")
# Create a small BertEncoder for testing.
test_network = bert_encoder.BertEncoder(vocab_size=100,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
dict_outputs=True)
# 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)
dict_outputs = test_network([word_ids, mask, type_ids])
data = dict_outputs["sequence_output"]
pooled = dict_outputs["pooled_output"]
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.assertAllEqual(tf.float32, data.dtype)
self.assertAllEqual(tf.float16, pooled.dtype)
def test_v2_network_creation(self):
hidden_size = 32
sequence_length = 21
# Create a small BertEncoder for testing.
test_network = bert_encoder.BertEncoder(vocab_size=100,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
dict_outputs=True)
# 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)
dict_outputs = test_network([word_ids, mask, type_ids])
data = dict_outputs["sequence_output"]
pooled = dict_outputs["pooled_output"]
self.assertIsInstance(test_network.transformer_layers, list)
self.assertLen(test_network.transformer_layers, 3)
self.assertIsInstance(test_network.pooler_layer, tf.keras.layers.Dense)
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())
# The default output dtype is float32.
self.assertAllEqual(tf.float32, data.dtype)
self.assertAllEqual(tf.float32, pooled.dtype)
def test_serialize_deserialize(self):
# Create a network object that sets all of its config options.
kwargs = dict(vocab_size=100,
hidden_size=32,
num_layers=3,
num_attention_heads=2,
max_sequence_length=21,
type_vocab_size=12,
inner_dim=1223,
inner_activation="relu",
output_dropout=0.05,
attention_dropout=0.22,
initializer="glorot_uniform",
output_range=-1,
embedding_width=16,
embedding_layer=None,
norm_first=False)
network = bert_encoder.BertEncoder(**kwargs)
# Validate that the config can be forced to JSON.
_ = network.to_json()
# Tests model saving/loading.
model_path = self.get_temp_dir() + "/model"
network.save(model_path)
_ = tf.keras.models.load_model(model_path)
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,
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",
return_all_encoder_outputs=False,
output_range=-1,
embedding_width=16)
network = bert_encoder.BertEncoder(**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 = bert_encoder.BertEncoder.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_weights_forward_compatible(self):
batch_size = 3
hidden_size = 32
sequence_length = 21
vocab_size = 57
num_types = 7
kwargs = dict(
vocab_size=vocab_size,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types,
output_range=None)
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))
data = dict(
input_word_ids=word_id_data,
input_mask=mask_data,
input_type_ids=type_id_data)
# Create small BertEncoders for testing.
new_net = bert_encoder.BertEncoderV2(**kwargs)
_ = new_net(data)
kwargs["dict_outputs"] = True
old_net = bert_encoder.BertEncoder(**kwargs)
_ = old_net(data)
new_net._embedding_layer.set_weights(old_net._embedding_layer.get_weights())
new_net._position_embedding_layer.set_weights(
old_net._position_embedding_layer.get_weights())
new_net._type_embedding_layer.set_weights(
old_net._type_embedding_layer.get_weights())
new_net._embedding_norm_layer.set_weights(
old_net._embedding_norm_layer.get_weights())
# embedding_dropout has no weights.
if hasattr(old_net, "_embedding_projection"):
new_net._embedding_projection.set_weights(
old_net._embedding_projection.get_weights())
# attention_mask_layer has no weights.
new_net._pooler_layer.set_weights(old_net._pooler_layer.get_weights())
for otl, ntl in zip(old_net._transformer_layers,
new_net._transformer_layers):
ntl.set_weights(otl.get_weights())
def check_output_close(data, net1, net2):
output1 = net1(data)
output2 = net2(data)
for key in output1:
self.assertAllClose(output1[key], output2[key])
check_output_close(data, old_net, new_net)
def test_layer_invocation_with_external_logits(self):
vocab_size = 100
sequence_length = 32
hidden_size = 64
num_predictions = 21
xformer_stack = bert_encoder.BertEncoder(
vocab_size=vocab_size,
num_layers=1,
hidden_size=hidden_size,
num_attention_heads=4,
)
test_layer = self.create_layer(vocab_size=vocab_size,
hidden_size=hidden_size,
xformer_stack=xformer_stack,
output='predictions')
logit_layer = self.create_layer(vocab_size=vocab_size,
hidden_size=hidden_size,
xformer_stack=xformer_stack,
output='logits')
# Create a model from the masked LM layer.
lm_input_tensor = tf.keras.Input(shape=(sequence_length, hidden_size))
masked_positions = tf.keras.Input(shape=(num_predictions, ),
dtype=tf.int32)
output = test_layer(lm_input_tensor, masked_positions)
logit_output = logit_layer(lm_input_tensor, masked_positions)
logit_output = tf.keras.layers.Activation(
tf.nn.log_softmax)(logit_output)
logit_layer.set_weights(test_layer.get_weights())
model = tf.keras.Model([lm_input_tensor, masked_positions], output)
logits_model = tf.keras.Model(([lm_input_tensor, masked_positions]),
logit_output)
# Invoke the masked LM on some fake data to make sure there are no runtime
# errors in the code.
batch_size = 3
lm_input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, hidden_size))
masked_position_data = np.random.randint(sequence_length,
size=(batch_size,
num_predictions))
# ref_outputs = model.predict([lm_input_data, masked_position_data])
# outputs = logits_model.predict([lm_input_data, masked_position_data])
ref_outputs = model([lm_input_data, masked_position_data])
outputs = logits_model([lm_input_data, masked_position_data])
# Ensure that the tensor shapes are correct.
expected_output_shape = (batch_size, num_predictions, vocab_size)
self.assertEqual(expected_output_shape, ref_outputs.shape)
self.assertEqual(expected_output_shape, outputs.shape)
self.assertAllClose(ref_outputs, outputs)
def test_keras_model_checkpoint_forward_compatible(self):
batch_size = 3
hidden_size = 32
sequence_length = 21
vocab_size = 57
num_types = 7
kwargs = dict(
vocab_size=vocab_size,
hidden_size=hidden_size,
num_attention_heads=2,
num_layers=3,
type_vocab_size=num_types,
output_range=None)
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))
data = dict(
input_word_ids=word_id_data,
input_mask=mask_data,
input_type_ids=type_id_data)
kwargs["dict_outputs"] = True
old_net = bert_encoder.BertEncoder(**kwargs)
inputs = old_net.inputs
outputs = old_net(inputs)
old_model = tf.keras.Model(inputs=inputs, outputs=outputs)
old_model_outputs = old_model(data)
ckpt = tf.train.Checkpoint(net=old_model)
path = ckpt.save(self.get_temp_dir())
del kwargs["dict_outputs"]
new_net = bert_encoder.BertEncoderV2(**kwargs)
inputs = new_net.inputs
outputs = new_net(inputs)
new_model = tf.keras.Model(inputs=inputs, outputs=outputs)
new_ckpt = tf.train.Checkpoint(net=new_model)
status = new_ckpt.restore(path)
status.assert_existing_objects_matched()
new_model_outputs = new_model(data)
self.assertAllEqual(old_model_outputs.keys(), new_model_outputs.keys())
for key in old_model_outputs:
self.assertAllClose(old_model_outputs[key], new_model_outputs[key])
def create_layer(self,
vocab_size,
hidden_size,
output='predictions',
xformer_stack=None):
# First, create a transformer stack that we can use to get the LM's
# vocabulary weight.
if xformer_stack is None:
xformer_stack = bert_encoder.BertEncoder(
vocab_size=vocab_size,
num_layers=1,
hidden_size=hidden_size,
num_attention_heads=4,
)
# Create a maskedLM from the transformer stack.
test_layer = masked_lm.MaskedLM(
embedding_table=xformer_stack.get_embedding_table(), output=output)
return test_layer
def test_serialize_deserialize(self):
# Create a network object that sets all of its config options.
kwargs = dict(vocab_size=100,
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",
return_all_encoder_outputs=False,
output_range=-1,
embedding_width=16,
dict_outputs=True,
embedding_layer=None,
norm_first=False)
network = bert_encoder.BertEncoder(**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 = bert_encoder.BertEncoder.from_config(
network.get_config())
# Validate that the config can be forced to JSON.
_ = network.to_json()
# If the serialization was successful, the new config should match the old.
self.assertAllEqual(network.get_config(), new_network.get_config())
# Tests model saving/loading.
model_path = self.get_temp_dir() + "/model"
network.save(model_path)
_ = tf.keras.models.load_model(model_path)
