Skip to content

GitLab

Unverified Commit 7a45b513 authored by Vishnu Banna's avatar Vishnu Banna Committed by GitHub
Browse files

Merge branch 'tensorflow:master' into exp_pr2

parents 54115e16 12bbefce
Expand all Show whitespace changes
Inline Side-by-side
Showing with 2128 additions and 1123 deletions
official/nlp/keras_nlp/README.md deleted 100644 → 0
View file @ 54115e16
# keras-nlp
## Layers
Layers are the fundamental building blocks for NLP models. They can be used to
assemble new layers, networks, or models.
* [TransformerEncoderBlock](layers/transformer_encoder_block.py) implements
an optionally masked transformer as described in
["Attention Is All You Need"](https://arxiv.org/abs/1706.03762).
* [OnDeviceEmbedding](layers/on_device_embedding.py) implements efficient
embedding lookups designed for TPU-based models.
* [PositionalEmbedding](layers/position_embedding.py) creates a positional
embedding as described in ["BERT: Pre-training of Deep Bidirectional
Transformers for Language Understanding"](https://arxiv.org/abs/1810.04805).
* [SelfAttentionMask](layers/self_attention_mask.py) creates a 3D attention
mask from a 2D tensor mask.
* [MaskedLM](layers/masked_lm.py) implements a masked language model. It
assumes the embedding table variable is passed to it.
## Encoders
Encoders are combinations of layers (and possibly other encoders). They are
sub-units of models that would not be trained alone. It encapsulates common
network structures like a classification head or a transformer encoder into an
easily handled object with a standardized configuration.
* [BertEncoder](encoders/bert_encoder.py) implements a bi-directional
Transformer-based encoder as described in
["BERT: Pre-training of Deep Bidirectional Transformers for Language
Understanding"](https://arxiv.org/abs/1810.04805). It includes the embedding
lookups, transformer layers and pooling layer.
official/nlp/keras_nlp/__init__.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Keras-NLP package definition."""
# pylint: disable=wildcard-import
from official.nlp.keras_nlp import encoders
from official.nlp.keras_nlp import layers
official/nlp/keras_nlp/contributing.md deleted 100644 → 0
View file @ 54115e16
## Contributing to KerasNLP
Patches to KerasNLP are welcome!
The source-of-truth repository lives under
[TF Model Garden NLP](https://github.com/tensorflow/models/tree/master/official/nlp/keras_nlp),
and is mirrored as a read-only repository under
[keras-team/keras-nlp](https://github.com/keras-team/keras-nlp).
Contributions should be made as PRs to the TF Model Garden repository.
This is to ensure the codebase is rigorously tested with state-of-art models
on different accelerators.
In the long run, we will move development to the current repository `keras-team/keras-nlp`.
## :heavy_check_mark: Contributor checklist
1. Ensure you have signed the [Contributor License Agreement](https://cla.developers.google.com/about/google-individual?csw=1).
* All code contributors are required to sign a Contributor License Agreement.
* Please read this [troubleshooting guide](Contributor-License-Agreements#troubleshooting-clas)
if you encounter an issue.
2. Please review the [contribution guidelines](https://github.com/tensorflow/models/wiki/How-to-contribute).
3. Check if your changes are consistent with the [TensorFlow coding style](https://www.tensorflow.org/community/contribute/code_style).
official/nlp/keras_nlp/encoders/bert_encoder_test.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Tests for transformer-based bert encoder network."""
from absl.testing import parameterized
import numpy as np
import tensorflow as tf
from tensorflow.python.keras import keras_parameterized # pylint: disable=g-direct-tensorflow-import
from official.nlp.keras_nlp.encoders import bert_encoder
# This decorator runs the test in V1, V2-Eager, and V2-Functional mode. It
# guarantees forward compatibility of this code for the V2 switchover.
@keras_parameterized.run_all_keras_modes
class BertEncoderTest(keras_parameterized.TestCase):
def tearDown(self):
super(BertEncoderTest, self).tearDown()
tf.keras.mixed_precision.set_global_policy("float32")
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)
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_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)
# 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])
all_encoder_outputs = dict_outputs["encoder_outputs"]
pooled = dict_outputs["pooled_output"]
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_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)
# 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)
@parameterized.named_parameters(
("all_sequence", None, 21),
("output_range", 1, 1),
)
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)
dict_outputs = test_network([word_ids, mask, type_ids])
data = dict_outputs["sequence_output"]
pooled = dict_outputs["pooled_output"]
# 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)
dict_outputs = test_network([word_ids, mask, type_ids])
data = dict_outputs["sequence_output"]
pooled = dict_outputs["pooled_output"]
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)
dict_outputs = test_network([word_ids, mask, type_ids])
data = dict_outputs["sequence_output"]
pooled = dict_outputs["pooled_output"]
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_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)
expected_config = dict(kwargs)
expected_config["inner_activation"] = tf.keras.activations.serialize(
tf.keras.activations.get(expected_config["inner_activation"]))
expected_config["initializer"] = tf.keras.initializers.serialize(
tf.keras.initializers.get(expected_config["initializer"]))
# 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)
if __name__ == "__main__":
tf.test.main()
official/nlp/keras_nlp/layers/__init__.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Keras-NLP layers package definition."""
from official.nlp.keras_nlp.layers.masked_lm import MaskedLM
from official.nlp.keras_nlp.layers.on_device_embedding import OnDeviceEmbedding
from official.nlp.keras_nlp.layers.position_embedding import PositionEmbedding
from official.nlp.keras_nlp.layers.self_attention_mask import SelfAttentionMask
from official.nlp.keras_nlp.layers.transformer_encoder_block import TransformerEncoderBlock
official/nlp/keras_nlp/layers/on_device_embedding_test.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Tests for Keras-based one-hot embedding layer."""
import numpy as np
import tensorflow as tf
from tensorflow.python.keras import keras_parameterized # pylint: disable=g-direct-tensorflow-import
from official.nlp.keras_nlp.layers import on_device_embedding
# This decorator runs the test in V1, V2-Eager, and V2-Functional mode. It
# guarantees forward compatibility of this code for the V2 switchover.
@keras_parameterized.run_all_keras_modes
class OnDeviceEmbeddingTest(keras_parameterized.TestCase):
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_creation_with_mixed_precision(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size, embedding_width=embedding_width,
dtype="mixed_float16")
# 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.float16)
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)
def test_layer_invocation_with_mixed_precision(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size, embedding_width=embedding_width,
dtype="mixed_float16")
# 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.float16, output.dtype)
def test_one_hot_layer_creation(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size,
embedding_width=embedding_width,
use_one_hot=True)
# 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_one_hot_layer_creation_with_mixed_precision(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size,
embedding_width=embedding_width,
dtype="mixed_float16",
use_one_hot=True)
# 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.float16)
def test_one_hot_layer_invocation(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size,
embedding_width=embedding_width,
use_one_hot=True)
# 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)
def test_one_hot_layer_invocation_with_mixed_precision(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size,
embedding_width=embedding_width,
dtype="mixed_float16",
use_one_hot=True)
# 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.float16, output.dtype)
def test_use_scale_layer_invocation(self):
vocab_size = 31
embedding_width = 27
test_layer = on_device_embedding.OnDeviceEmbedding(
vocab_size=vocab_size, embedding_width=embedding_width,
scale_factor=embedding_width**0.5)
# 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)
if __name__ == "__main__":
tf.test.main()
official/nlp/keras_nlp/layers/position_embedding_test.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Tests for Keras-based positional embedding layer."""
import numpy as np
import tensorflow as tf
from tensorflow.python.keras import keras_parameterized # pylint: disable=g-direct-tensorflow-import
from official.nlp.keras_nlp.layers import position_embedding
# This decorator runs the test in V1, V2-Eager, and V2-Functional mode. It
# guarantees forward compatibility of this code for the V2 switchover.
@keras_parameterized.run_all_keras_modes
class PositionEmbeddingLayerTest(keras_parameterized.TestCase):
def test_static_layer_output_shape(self):
# Create a 3-dimensional input (the first dimension is implicit).
sequence_length = 21
test_layer = position_embedding.PositionEmbedding(
max_length=sequence_length)
width = 30
input_tensor = tf.keras.Input(shape=(sequence_length, width))
output_tensor = test_layer(input_tensor)
# When using static positional embedding shapes, the output is expected
# to be the same as the input shape in all dimensions save batch.
expected_output_shape = [None, sequence_length, width]
self.assertEqual(expected_output_shape, output_tensor.shape.as_list())
# The default output dtype for this layer should be tf.float32.
self.assertEqual(tf.float32, output_tensor.dtype)
def test_non_default_axis_static(self):
# Create a 3-dimensional input (the first dimension is implicit).
sequence_length = 21
test_layer = position_embedding.PositionEmbedding(
max_length=sequence_length, seq_axis=2)
width = 30
input_tensor = tf.keras.Input(shape=(width, sequence_length, width))
output_tensor = test_layer(input_tensor)
# When using static positional embedding shapes, the output is expected
# to be the same as the input shape in all dimensions save batch.
expected_output_shape = [None, width, sequence_length, width]
self.assertEqual(expected_output_shape, output_tensor.shape.as_list())
# The default output dtype for this layer should be tf.float32.
self.assertEqual(tf.float32, output_tensor.dtype)
def test_float16_dtype(self):
# Create a 3-dimensional input (the first dimension is implicit).
sequence_length = 21
test_layer = position_embedding.PositionEmbedding(
max_length=sequence_length, dtype="float16")
width = 30
input_tensor = tf.keras.Input(shape=(sequence_length, width))
output_tensor = test_layer(input_tensor)
# When using static positional embedding shapes, the output is expected
# to be the same as the input shape in all dimensions save batch.
expected_output_shape = [None, sequence_length, width]
self.assertEqual(expected_output_shape, output_tensor.shape.as_list())
# The default output dtype for this layer should be tf.float32.
self.assertEqual(tf.float16, output_tensor.dtype)
def test_dynamic_layer_output_shape(self):
max_sequence_length = 40
test_layer = position_embedding.PositionEmbedding(
max_length=max_sequence_length)
# Create a 3-dimensional input (the first dimension is implicit).
width = 30
input_tensor = tf.keras.Input(shape=(None, width))
output_tensor = test_layer(input_tensor)
# When using dynamic positional embedding shapes, the output is expected
# to be the same as the input shape in all dimensions - but may be None if
# the input shape is None there.
expected_output_shape = [None, None, width]
self.assertEqual(expected_output_shape, output_tensor.shape.as_list())
def test_non_default_axis_dynamic(self):
max_sequence_length = 60
test_layer = position_embedding.PositionEmbedding(
max_length=max_sequence_length, seq_axis=2)
# Create a 3-dimensional input (the first dimension is implicit).
width = 30
input_tensor = tf.keras.Input(shape=(None, None, width))
output_tensor = test_layer(input_tensor)
# When using dynamic positional embedding shapes, the output is expected
# to be the same as the input shape in all dimensions - but may be None if
# the input shape is None there.
expected_output_shape = [None, None, None, width]
self.assertEqual(expected_output_shape, output_tensor.shape.as_list())
def test_dynamic_layer_slicing(self):
max_sequence_length = 40
test_layer = position_embedding.PositionEmbedding(
max_length=max_sequence_length)
# Create a 3-dimensional input (the first dimension is implicit).
width = 30
input_tensor = tf.keras.Input(shape=(None, width))
output_tensor = test_layer(input_tensor)
model = tf.keras.Model(input_tensor, output_tensor)
# Create input data that is shorter than max_sequence_length, which should
# trigger a down-slice.
input_length = 17
# Note: This test explicitly uses a batch size of 1. This is to get around
# Keras' restriction on Model invocations: inputs are expected to have the
# same batch cardinality as outputs. In practice, this layer should be used
# inside a model, where it can be projected when added to another tensor.
input_data = np.ones((1, input_length, width))
output_data = model.predict(input_data)
self.assertAllEqual([1, input_length, width], output_data.shape)
if __name__ == "__main__":
tf.test.main()
official/nlp/keras_nlp/layers/transformer_encoder_block.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Keras-based TransformerEncoder block layer."""
from official.nlp.modeling import layers
TransformerEncoderBlock = layers.TransformerEncoderBlock
official/nlp/keras_nlp/layers/transformer_encoder_block_test.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Tests for Keras-based transformer block layer."""
from absl.testing import parameterized
import numpy as np
import tensorflow as tf
from tensorflow.python.keras import keras_parameterized # pylint: disable=g-direct-tensorflow-import
from official.nlp.keras_nlp.layers.transformer_encoder_block import TransformerEncoderBlock
@keras_parameterized.run_all_keras_modes
@parameterized.named_parameters(
('base', TransformerEncoderBlock))
class TransformerEncoderBlockLayerTest(keras_parameterized.TestCase):
def tearDown(self):
super(TransformerEncoderBlockLayerTest, self).tearDown()
tf.keras.mixed_precision.set_global_policy('float32')
def test_layer_creation(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048, inner_activation='relu')
sequence_length = 21
width = 80
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
output_tensor = test_layer(data_tensor)
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output_tensor.shape.as_list())
def test_layer_creation_with_mask(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048, inner_activation='relu')
sequence_length = 21
width = 80
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output_tensor.shape.as_list())
def test_layer_invocation(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048, inner_activation='relu')
sequence_length = 21
width = 80
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
output_tensor = test_layer(data_tensor)
# Create a model from the test layer.
model = tf.keras.Model(data_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 = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
_ = model.predict(input_data)
def test_layer_invocation_with_mask(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048, inner_activation='relu')
sequence_length = 21
width = 80
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# Create a model from the test layer.
model = tf.keras.Model([data_tensor, mask_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 = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
# The attention mask should be of shape (batch, from_seq_len, to_seq_len),
# which here is (batch, sequence_length, sequence_length)
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
_ = model.predict([input_data, mask_data])
def test_layer_output_range(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048, inner_activation='relu')
sequence_length = 21
width = 80
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
output_tensor = test_layer([input_data, mask_data])
# The layer only attends to the first token and outputs the first token
# embedding.
new_layer = transformer_cls(
num_attention_heads=10,
inner_dim=2048,
inner_activation='relu',
output_range=1)
_ = new_layer([input_data, mask_data])
new_layer.set_weights(test_layer.get_weights())
new_output_tensor = new_layer([input_data, mask_data])
self.assertAllClose(
new_output_tensor, output_tensor[:, 0:1, :], atol=5e-5, rtol=0.003)
def test_layer_output_range_without_mask(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048,
inner_activation='relu', norm_first=True)
sequence_length = 21
width = 80
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
output_tensor = test_layer(input_data)
# The layer only attends to the first token and outputs the first token
# embedding.
new_layer = transformer_cls(
num_attention_heads=10,
inner_dim=2048,
inner_activation='relu',
output_range=1,
norm_first=True)
_ = new_layer(input_data)
new_layer.set_weights(test_layer.get_weights())
new_output_tensor = new_layer(input_data)
self.assertAllClose(
new_output_tensor, output_tensor[:, 0:1, :], atol=5e-5, rtol=0.003)
def test_layer_output_range_with_pre_norm(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048,
inner_activation='relu', norm_first=True)
sequence_length = 21
width = 80
batch_size = 6
input_data = 10 * np.random.random_sample(
(batch_size, sequence_length, width))
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
output_tensor = test_layer([input_data, mask_data])
# The layer only attends to the first token and outputs the first token
# embedding.
new_layer = transformer_cls(
num_attention_heads=10,
inner_dim=2048,
inner_activation='relu',
output_range=1,
norm_first=True)
_ = new_layer([input_data, mask_data])
new_layer.set_weights(test_layer.get_weights())
new_output_tensor = new_layer([input_data, mask_data])
self.assertAllClose(
new_output_tensor, output_tensor[:, 0:1, :], atol=5e-5, rtol=0.003)
def test_layer_invocation_with_float16_dtype(self, transformer_cls):
tf.keras.mixed_precision.set_global_policy('mixed_float16')
test_layer = transformer_cls(
num_attention_heads=10, inner_dim=2048, inner_activation='relu')
sequence_length = 21
width = 80
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
# Create a 2-dimensional input (the first dimension is implicit).
mask_tensor = tf.keras.Input(shape=(sequence_length, sequence_length))
output_tensor = test_layer([data_tensor, mask_tensor])
# Create a model from the test layer.
model = tf.keras.Model([data_tensor, mask_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 = 6
input_data = (10 * np.random.random_sample(
(batch_size, sequence_length, width)))
# The attention mask should be of shape (batch, from_seq_len, to_seq_len),
# which here is (batch, sequence_length, sequence_length)
mask_data = np.random.randint(
2, size=(batch_size, sequence_length, sequence_length))
_ = model.predict([input_data, mask_data])
def test_transform_with_initializer(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
inner_dim=2048,
inner_activation='relu',
kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
sequence_length = 21
width = 80
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(sequence_length, width))
output = test_layer(data_tensor)
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output.shape.as_list())
def test_dynamic_layer_sequence(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
inner_dim=2048,
inner_activation='relu',
kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
# Create a 3-dimensional input (the first dimension is implicit).
width = 30
input_tensor = tf.keras.Input(shape=(None, width))
output_tensor = test_layer(input_tensor)
model = tf.keras.Model(input_tensor, output_tensor)
input_length = 17
input_data = np.ones((1, input_length, width))
output_data = model.predict(input_data)
self.assertAllEqual([1, input_length, width], output_data.shape)
def test_separate_qkv(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=2,
inner_dim=128,
inner_activation='relu',
kernel_initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02))
# Forward path.
q_tensor = tf.zeros([2, 4, 16], dtype=tf.float32)
kv_tensor = tf.zeros([2, 8, 16], dtype=tf.float32)
dummy_mask = tf.zeros([2, 4, 8], dtype=tf.float32)
inputs = [q_tensor, kv_tensor, dummy_mask]
output = test_layer(inputs)
self.assertEqual(output.shape, q_tensor.shape)
@keras_parameterized.run_all_keras_modes
class TransformerArgumentTest(keras_parameterized.TestCase):
def test_use_bias_norm_first(self):
num_attention_heads = 2
hidden_size = 16
encoder_block = TransformerEncoderBlock(
num_attention_heads=num_attention_heads,
inner_dim=32,
inner_activation='relu',
output_dropout=0.1,
attention_dropout=0.1,
use_bias=False,
norm_first=True,
norm_epsilon=1e-6,
inner_dropout=0.1,
attention_initializer=tf.keras.initializers.RandomUniform(
minval=0., maxval=1.))
# Forward path.
dummy_tensor = tf.zeros([2, 4, 16], dtype=tf.float32)
dummy_mask = tf.zeros([2, 4, 4], dtype=tf.float32)
inputs = [dummy_tensor, dummy_mask]
output = encoder_block(inputs)
self.assertEqual(output.shape, (2, 4, hidden_size))
def test_get_config(self):
num_attention_heads = 2
encoder_block = TransformerEncoderBlock(
num_attention_heads=num_attention_heads,
inner_dim=32,
inner_activation='relu',
output_dropout=0.1,
attention_dropout=0.1,
use_bias=False,
norm_first=True,
norm_epsilon=1e-6,
inner_dropout=0.1,
attention_initializer=tf.keras.initializers.RandomUniform(
minval=0., maxval=1.))
encoder_block_config = encoder_block.get_config()
new_encoder_block = TransformerEncoderBlock.from_config(
encoder_block_config)
self.assertEqual(encoder_block_config, new_encoder_block.get_config())
@parameterized.parameters({'attention_axes': None}, {'attention_axes': [1]},
{'attention_axes': [2]}, {'attention_axes': [1, 2]})
def test_several_attention_axes(self, attention_axes):
test_layer = TransformerEncoderBlock(
inner_dim=32,
inner_activation='relu',
output_dropout=0.1,
attention_dropout=0.1,
use_bias=False,
norm_first=True,
norm_epsilon=1e-6,
inner_dropout=0.1,
num_attention_heads=10,
attention_axes=attention_axes)
num_rows = 21
num_cols = 13
width = 80
# Create a 3-dimensional input (the first dimension is implicit).
data_tensor = tf.keras.Input(shape=(num_rows, num_cols, width))
output_tensor = test_layer(data_tensor)
# The default output of a transformer layer should be the same as the input.
self.assertEqual(data_tensor.shape.as_list(), output_tensor.shape.as_list())
if __name__ == '__main__':
tf.test.main()
official/nlp/keras_nlp/setup.py deleted 100644 → 0
View file @ 54115e16
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Setup script."""
import os
from setuptools import find_packages
from setuptools import setup
version = '0.0.1'
def _get_requirements():
"""Parses requirements.txt file."""
install_requires_tmp = []
dependency_links_tmp = []
with open(
os.path.join(os.path.dirname(__file__), './requirements.txt'), 'r') as f:
for line in f:
package_name = line.strip()
# Skip empty line or comments starting with "#".
if not package_name or package_name[0] == '#':
continue
if package_name.startswith('-e '):
dependency_links_tmp.append(package_name[3:].strip())
else:
install_requires_tmp.append(package_name)
return install_requires_tmp, dependency_links_tmp
install_requires, dependency_links = _get_requirements()
install_requires.append('tf-nightly')
setup(
name='keras-nlp',
version=version,
description='Keras Natural Language Processing Library',
url='https://github.com/keras-team/keras-nlp',
author='The Keras authors',
author_email='keras-team@google.com',
license='Apache License 2.0',
install_requires=install_requires,
classifiers=[
'Programming Language :: Python',
'Programming Language :: Python :: 3.6',
'Operating System :: Unix',
'Operating System :: Microsoft :: Windows',
'Operating System :: MacOS',
'Intended Audience :: Science/Research',
'Topic :: Scientific/Engineering',
'Topic :: Software Development'
],
packages=find_packages(exclude=('tests',)),
exclude_package_data={'': ['*_test.py',],},
dependency_links=dependency_links,
python_requires='>=3.6',
)
official/nlp/modeling/models/README.md
View file @ 7a45b513
......@@ -23,3 +23,12 @@ respectively.
* [`DualEncoder`](dual_encoder.py) implements a dual encoder model, suitbale for
retrieval tasks.
* [`Seq2SeqTransformer`](seq2seq_transformer.py) implements the original
Transformer model for seq-to-seq tasks.
* [`T5Transformer`](t5.py) implements a standalone T5 model for seq-to-seq
tasks. The models are compatible with released T5 architecture and converted
checkpoints. The modules are implemented as `tf.Module`. To use with Keras,
users can wrap them within Keras customized layers, i.e. we can define the
modules inside the `__init__` of Keras layer and call the modules in `call`.
official/nlp/modeling/models/__init__.py
View file @ 7a45b513
......@@ -24,6 +24,8 @@ from official.nlp.modeling.models.bert_token_classifier import BertTokenClassifi
from official.nlp.modeling.models.dual_encoder import DualEncoder
from official.nlp.modeling.models.electra_pretrainer import ElectraPretrainer
from official.nlp.modeling.models.seq2seq_transformer import *
from official.nlp.modeling.models.t5 import T5Transformer
from official.nlp.modeling.models.t5 import T5TransformerParams
from official.nlp.modeling.models.xlnet import XLNetClassifier
from official.nlp.modeling.models.xlnet import XLNetPretrainer
from official.nlp.modeling.models.xlnet import XLNetSpanLabeler
official/nlp/modeling/models/t5.py 0 → 100644
View file @ 7a45b513
This diff is collapsed.
official/nlp/modeling/models/t5_test.py 0 → 100644
View file @ 7a45b513
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Tests for t5."""
from absl.testing import parameterized
import numpy as np
import tensorflow as tf
from tensorflow.python.distribute import combinations
from tensorflow.python.distribute import strategy_combinations
from official.nlp.modeling.models import t5
def _create_cache(batch_size,
init_decode_length,
num_heads,
head_size,
dtype=tf.float32):
if num_heads is None:
kv_shape = [batch_size, init_decode_length, head_size]
else:
kv_shape = [batch_size, init_decode_length, num_heads, head_size]
return {
"key": tf.zeros(kv_shape, dtype=dtype),
"value": tf.zeros(kv_shape, dtype=dtype)
}
class ModulesTest(tf.test.TestCase, parameterized.TestCase):
@parameterized.named_parameters(("bfloat16", tf.bfloat16),
("float32", tf.float32))
def test_embed(self, dtype):
l = t5.Embed(vocab_size=5, features=4, compute_dtype=dtype, name="foo")
inputs = np.array([[2, 3], [1, 2]], dtype=np.int32)
inputs = tf.convert_to_tensor(inputs)
one_hot_outputs = l(inputs, one_hot=True)
gather_outputs = l(inputs, one_hot=False)
self.assertEqual(one_hot_outputs.shape, (2, 2, 4))
self.assertLen(l.trainable_variables, 1)
self.assertAllClose(one_hot_outputs, gather_outputs)
outputs = l.attend(query=tf.zeros((2, 2, 4), dtype))
self.assertEqual(outputs.shape, (2, 2, 5))
# Test initializers.
l = t5.Embed(
vocab_size=5,
features=4,
compute_dtype=dtype,
name="foo",
embeddings_initializer=tf.keras.initializers.Zeros())
self.assertAllClose(l(inputs), tf.zeros((2, 2, 4), dtype))
@parameterized.named_parameters(("bfloat16", tf.bfloat16),
("float32", tf.float32))
def test_rms_norm(self, dtype):
l = t5.RMSNorm(hidden_size=4, epsilon=0.0, name="foo")
inputs = tf.ones((2, 4), dtype=dtype)
outputs = l(inputs)
self.assertAllEqual(l(inputs), inputs)
self.assertEqual(outputs.dtype, dtype)
self.assertLen(l.trainable_variables, 1)
self.assertIn("foo/scale", l.trainable_variables[0].name)
@parameterized.named_parameters(("bfloat16", tf.bfloat16),
("float32", tf.float32))
def test_linear(self, dtype):
l = t5.Linear(
in_features=4,
out_features=4,
w_init=tf.keras.initializers.Ones(),
name="foo")
inputs = tf.ones((2, 4), dtype=dtype)
outputs = l(inputs)
self.assertEqual(outputs.shape, inputs.shape)
self.assertEqual(outputs.dtype, dtype)
self.assertLen(l.trainable_variables, 2)
def test_linear3d(self):
batch_size = 2
l = t5.Linear3D(
in_features=4,
out_features=4,
num_heads=2,
to_3d=True,
w_init=tf.keras.initializers.Ones(),
name="foo")
inputs = np.ones((batch_size, 2, 4), dtype=np.float32)
self.assertEqual(l(inputs).shape, (batch_size, 2, 2, 4))
l = t5.Linear3D(
in_features=2,
out_features=4,
num_heads=2,
to_3d=False,
w_init=tf.keras.initializers.Ones(),
name="foo")
inputs = np.ones((batch_size, 2, 2, 2), dtype=np.float32)
self.assertEqual(l(inputs).shape, (batch_size, 2, 4))
def test_ffn(self):
inputs = np.ones((2, 4), dtype=np.float32)
for activation in ["relu", "linear", "gelu", "swish"]:
l = t5.FFN(
d_model=4,
d_ff=8,
use_bias=True,
dropout_rate=0.1,
activations=[activation],
name="foo")
self.assertEqual(l(inputs).shape, inputs.shape)
self.assertLen(l.trainable_variables, 4)
l = t5.FFN(
d_model=4,
d_ff=8,
dropout_rate=0.1,
activations=["linear", "gelu"],
name="bar")
self.assertLen(l.trainable_variables, 3)
self.assertEqual(l(inputs).shape, inputs.shape)
@parameterized.named_parameters(("bfloat16", tf.bfloat16),
("float32", tf.float32))
def test_relative_position(self, dtype):
l = t5.RelativePositionEmbedding(
num_heads=4,
bidirectional=False,
embeddings_initializer=tf.keras.initializers.Ones(),
compute_dtype=dtype,
name="foo")
self.assertEqual(l(4, 2).shape, (1, 4, 4, 2))
l = t5.RelativePositionEmbedding(
num_heads=4,
bidirectional=True,
embeddings_initializer=tf.keras.initializers.Ones(),
compute_dtype=dtype,
name="bar")
outputs = l(4, 2)
self.assertEqual(outputs.shape, (1, 4, 4, 2))
self.assertEqual(outputs.dtype, dtype)
def test_masks(self):
causal_mask = t5.make_causal_mask(np.zeros((2, 5)))
self.assertEqual(causal_mask.shape, (2, 1, 5, 5))
@combinations.generate(
combinations.combine(
distribution=[
strategy_combinations.default_strategy,
strategy_combinations.cloud_tpu_strategy,
],
mode="eager"))
def test_attention(self, distribution):
num_heads, head_size = 2, 4
from_seq_length, to_seq_length = 4, 6
batch_size = 2
pos_embed = t5.RelativePositionEmbedding(
num_heads=4,
bidirectional=False,
embeddings_initializer=tf.keras.initializers.Ones(),
name="pos_embed")
position_bias = pos_embed(from_seq_length, from_seq_length)
l = t5.MultiHeadAttention(d_model=4, d_kv=2, num_heads=4, dropout_rate=0.1)
query = tf.convert_to_tensor(
np.ones((batch_size, from_seq_length, 4), dtype=np.float32))
self.assertEqual(
l(query, position_bias=position_bias)["context"].shape, query.shape)
kv = tf.convert_to_tensor(
np.ones((batch_size, to_seq_length, 4), dtype=np.float32))
position_bias = pos_embed(from_seq_length, to_seq_length)
outputs = l(query, kv=kv, position_bias=position_bias)
self.assertEqual(outputs["context"].shape, query.shape)
with distribution.scope():
l = t5.MultiHeadAttention(
d_model=4, d_kv=head_size, num_heads=num_heads, dropout_rate=0.1)
@tf.function
def step(inputs):
def _step_fn(inputs):
cache = _create_cache(batch_size, from_seq_length, num_heads,
head_size)
mask = t5.make_causal_mask(tf.ones((batch_size, 1)))
return l(
query=inputs,
mask=mask,
cache=cache,
decode_position=decode_position)
outputs = distribution.run(_step_fn, args=(inputs,))
return tf.nest.map_structure(distribution.experimental_local_results,
outputs)
decode_position = 2
query = tf.convert_to_tensor(np.ones((2, 1, 4), dtype=np.float32))
local_outputs = step(query)
self.assertEqual(local_outputs["context"][0].shape, (2, 1, 4))
self.assertNotEqual(
np.sum(local_outputs["cache"]["key"][0][:, decode_position,
...].numpy()), 0.0)
class T5Test(tf.test.TestCase, parameterized.TestCase):
@combinations.generate(
combinations.combine(
distribution=[
strategy_combinations.default_strategy,
strategy_combinations.cloud_tpu_strategy,
],
mode="eager"))
def test_attention_layers(self, distribution):
num_heads, head_size = 2, 2
from_seq_length = 4
# TPU decoding should pre-allocate the entire sequence.
batch_size = 2
with distribution.scope():
pos_embed = t5.RelativePositionEmbedding(
num_heads=head_size,
bidirectional=False,
embeddings_initializer=tf.keras.initializers.Ones(),
name="pos_embed")
l = t5.SelfAttention(
d_model=4, d_kv=head_size, num_heads=num_heads, dropout_rate=0.1)
decode_position = 2
@tf.function
def step(inputs):
def _step_fn(inputs):
cache = _create_cache(batch_size, from_seq_length, num_heads,
head_size)
mask = t5.make_causal_mask(tf.ones((batch_size, 1)))
position_bias = pos_embed(from_seq_length, from_seq_length)
return l(
hidden_states=inputs,
cache=cache,
attention_mask=mask,
decode_position=decode_position,
position_bias=position_bias)
outputs = distribution.run(_step_fn, args=(inputs,))
return tf.nest.map_structure(distribution.experimental_local_results,
outputs)
query = tf.convert_to_tensor(np.ones((2, 1, 4), dtype=np.float32))
local_outputs = step(query)
self.assertEqual(local_outputs["layer_output"][0].shape, (2, 1, 4))
self.assertNotEqual(
np.sum(
local_outputs["cache"]["key"][0][:,
decode_position, :, :].numpy()),
0.0)
l = t5.CrossAttention(
d_model=4, d_kv=head_size, num_heads=num_heads, dropout_rate=0.1)
to_seq_length = 6
query = tf.convert_to_tensor(
np.ones((2, from_seq_length, 4), dtype=np.float32))
kv = tf.convert_to_tensor(
np.ones((2, to_seq_length, 4), dtype=np.float32))
@tf.function
def step_cross_attn(inputs):
def _step_fn(inputs):
query, kv = inputs
mask = t5.make_attention_mask(
tf.ones((batch_size, from_seq_length)),
tf.ones((batch_size, to_seq_length)))
return l(hidden_states=query, kv=kv, attention_mask=mask)
outputs = distribution.run(_step_fn, args=(inputs,))
return tf.nest.map_structure(distribution.experimental_local_results,
outputs)
local_outputs = step_cross_attn((query, kv))
self.assertEqual(local_outputs["layer_output"][0].shape,
(2, from_seq_length, 4))
def test_encoder_block(self):
batch_size = 2
from_seq_length = 5
d_model = 4
l = t5.EncoderBlock(d_model=4, d_kv=3, num_heads=2, d_ff=8, name="foo")
pos_embed = t5.RelativePositionEmbedding(
num_heads=2,
bidirectional=True,
embeddings_initializer=tf.keras.initializers.Ones(),
name="bar")
attention_mask = t5.make_attention_mask(
tf.ones((batch_size, from_seq_length)),
tf.ones((batch_size, from_seq_length)))
position_bias = pos_embed(from_seq_length, from_seq_length)
inputs = tf.ones((batch_size, from_seq_length, d_model), dtype=tf.float32)
outputs = l(
inputs, attention_mask=attention_mask, position_bias=position_bias)
self.assertEqual(outputs.shape, (batch_size, from_seq_length, d_model))
def test_encdec_block(self):
batch_size = 2
from_seq_length = 5
to_seq_length = 3
d_model = 4
l = t5.EncDecoderBlock(d_model=4, d_kv=3, num_heads=2, d_ff=8, name="foo")
pos_embed = t5.RelativePositionEmbedding(
num_heads=2,
bidirectional=True,
embeddings_initializer=tf.keras.initializers.Ones(),
name="bar")
encoder_decoder_mask = t5.make_attention_mask(
tf.ones((batch_size, from_seq_length)),
tf.ones((batch_size, to_seq_length)))
position_bias = pos_embed(from_seq_length, from_seq_length)
inputs = tf.ones((batch_size, from_seq_length, d_model), dtype=tf.float32)
encoder_hidden_states = tf.ones((batch_size, to_seq_length, d_model),
dtype=tf.float32)
outputs = l(
inputs,
encoder_hidden_states,
encoder_decoder_mask=encoder_decoder_mask,
position_bias=position_bias)
self.assertEqual(outputs[0].shape, (batch_size, from_seq_length, d_model))
@parameterized.named_parameters(("bfloat16", tf.bfloat16),
("float32", tf.float32))
def test_encoder(self, dtype):
config = t5.T5TransformerParams(
num_layers=2,
d_model=4,
d_kv=3,
num_heads=4,
d_ff=16,
vocab_size=10,
vocab_embeddings_initializer=tf.keras.initializers.Ones(),
relative_embeddings_initializer=tf.keras.initializers.Ones())
encoder = t5.Encoder(config, compute_dtype=dtype)
encoded = encoder(tf.zeros((4, 8), dtype=tf.int32))
self.assertEqual(encoded.shape, (4, 8, config.d_model))
def test_decoder(self):
max_decode_len = 10
config = t5.T5TransformerParams(
num_layers=2,
d_model=4,
d_kv=3,
num_heads=4,
d_ff=16,
vocab_size=10,
vocab_embeddings_initializer=tf.keras.initializers.Ones(),
relative_embeddings_initializer=tf.keras.initializers.Ones())
decoder = t5.Decoder(config)
batch_size = 4
targets = tf.zeros((4, 8), dtype=tf.int32)
encoded = tf.zeros((4, 8, config.d_model), dtype=tf.float32)
logits, cache = decoder(targets, encoded)
self.assertEqual(logits.shape, (4, 8, config.vocab_size))
cache = {}
cache[0] = _create_cache(batch_size, max_decode_len, config.num_heads,
config.d_kv)
cache[1] = _create_cache(batch_size, max_decode_len, config.num_heads,
config.d_kv)
targets = tf.zeros((4, 1), dtype=tf.int32)
logits, cache = decoder(
targets,
encoded,
decode_position=2,
cache=cache,
decode=True,
max_decode_len=max_decode_len)
self.assertEqual(logits.shape, (batch_size, 1, config.vocab_size))
for entry in cache.values():
for tensor in entry.values():
self.assertNotAllEqual(tensor.numpy()[:, 2, :, :], 0.0)
@parameterized.named_parameters(
("t5_10", ("relu",), True, 26, False, tf.float32),
("t5_11", ("gelu", "linear"), False, 29, False, tf.float32),
("t5_10_bfloat16", ("relu",), True, 26, False, tf.bfloat16),
("t5_11_bfloat16", ("gelu", "linear"), False, 29, False, tf.bfloat16),
("t5_10_layer_sharing", ("relu",), True, 26, True, tf.float32),
("t5_11_layer_sharing", ("gelu", "linear"), False, 29, True, tf.float32),
("t5_10_bfloat16_layer_sharing", ("relu",), True, 26, True, tf.bfloat16),
("t5_11_bfloat16_layer_sharing",
("gelu", "linear"), False, 29, True, tf.bfloat16))
def test_transformer(self, ffn_activations, logits_via_embedding,
expect_num_variables, layer_sharing, dtype):
max_decode_len = 10
config = t5.T5TransformerParams(
num_layers=1,
d_model=8,
d_kv=4,
num_heads=4,
d_ff=32,
vocab_size=10,
shared_embedding=True,
layer_sharing=layer_sharing,
ffn_activations=ffn_activations,
logits_via_embedding=logits_via_embedding)
transformer = t5.T5Transformer(config, compute_dtype=dtype)
self.assertLen(transformer.trainable_variables, expect_num_variables)
inputs = tf.convert_to_tensor(
np.array([[2, 2, 1, 3, 1, 0], [3, 3, 1, 2, 2, 1]]))
segments = tf.convert_to_tensor(
np.array([[1, 1, 1, 2, 2, 0], [1, 1, 1, 2, 2, 2]]))
outputs = transformer(
encoder_input_tokens=inputs,
decoder_input_tokens=inputs,
decoder_target_tokens=inputs,
encoder_segment_ids=segments,
decoder_segment_ids=segments)
cache = {}
batch_size = 2
cache[0] = _create_cache(
batch_size, max_decode_len, config.num_heads, config.d_kv, dtype=dtype)
outputs = transformer.decode(
encoder_input_tokens=inputs,
encoded=outputs["encoded"],
decoder_target_tokens=tf.ones((batch_size, 1), dtype=tf.int32),
decode_position=1,
decode=True,
max_decode_len=max_decode_len,
cache=cache)
self.assertEqual(outputs["logits"].shape,
(batch_size, 1, config.vocab_size))
for v in transformer.trainable_variables:
print(v.name, v.shape)
self.assertEqual(v.dtype, tf.float32)
@parameterized.named_parameters(
("t5_10", ("relu",), True, 39, tf.float32, 2),
("t5_10_bfloat16", ("relu",), True, 39, tf.bfloat16, 2))
def test_transformer_different_num_decoder_layers(self, ffn_activations,
logits_via_embedding,
expect_num_variables, dtype,
num_decoder_layers):
max_decode_len = 10
config = t5.T5TransformerParams(
num_decoder_layers=num_decoder_layers,
num_layers=1,
d_model=8,
d_kv=4,
num_heads=4,
d_ff=32,
vocab_size=10,
shared_embedding=True,
ffn_activations=ffn_activations,
logits_via_embedding=logits_via_embedding)
transformer = t5.T5Transformer(config, compute_dtype=dtype)
self.assertLen(transformer.trainable_variables, expect_num_variables)
inputs = tf.convert_to_tensor(
np.array([[2, 2, 1, 3, 1, 0], [3, 3, 1, 2, 2, 1]]))
segments = tf.convert_to_tensor(
np.array([[1, 1, 1, 2, 2, 0], [1, 1, 1, 2, 2, 2]]))
outputs = transformer(
encoder_input_tokens=inputs,
decoder_input_tokens=inputs,
decoder_target_tokens=inputs,
encoder_segment_ids=segments,
decoder_segment_ids=segments)
cache = {}
batch_size = 2
for i in range(num_decoder_layers):
cache[i] = _create_cache(
batch_size,
max_decode_len,
config.num_heads,
config.d_kv,
dtype=dtype)
outputs = transformer.decode(
encoder_input_tokens=inputs,
encoded=outputs["encoded"],
decoder_target_tokens=tf.ones((batch_size, 1), dtype=tf.int32),
decode_position=1,
decode=True,
max_decode_len=max_decode_len,
cache=cache)
self.assertEqual(outputs["logits"].shape,
(batch_size, 1, config.vocab_size))
for v in transformer.trainable_variables:
print(v.name, v.shape)
self.assertEqual(v.dtype, tf.float32)
if __name__ == "__main__":
tf.test.main()
official/nlp/modeling/networks/funnel_transformer.py
View file @ 7a45b513
......@@ -14,6 +14,7 @@
"""Funnel Transformer network."""
# pylint: disable=g-classes-have-attributes
from typing import Union, Sequence
from absl import logging
import numpy as np
......@@ -21,6 +22,10 @@ import tensorflow as tf
from official.nlp.modeling import layers
_MAX = 'max'
_AVG = 'avg'
_TRUNCATED_AVG = 'truncated_avg'
def _pool_and_concat(mask, unpool_length: int, strides: Union[Sequence[int],
int],
......@@ -63,6 +68,94 @@ def _pool_and_concat(mask, unpool_length: int, strides: Union[Sequence[int],
return mask
def _create_truncated_avg_transforms(seq_length: int,
pool_strides: Sequence[int]):
"""Computes pooling transforms.
The pooling_transform is of shape [seq_length,
seq_length//pool_stride] and
pooling_transform[i,j] = 1.0/pool_stride if i//pool_stride == j
0.0 otherwise.
It's in essense average pooling but truncate the final window if it
seq_length % pool_stride != 0.
For seq_length==6 and pool_stride==2, it is
[[ 0.5, 0.0, 0.0 ],
[ 0.5, 0.0, 0.0 ],
[ 0.0, 0.5, 0.0 ],
[ 0.0, 0.5, 0.0 ],
[ 0.0, 0.0, 0.5 ],
[ 0.0, 0.0, 0.5 ]]
Args:
seq_length: int, sequence length.
pool_strides: Sequence of pooling strides for each layer.
Returns:
pooling_transforms: Sequence of pooling transforms (Tensors) for each layer.
"""
pooling_transforms = []
for pool_stride in pool_strides:
if pool_stride == 1:
pooling_transforms.append(None)
else:
pooled_seq_length = seq_length // pool_stride
pfac, sl, psl = pool_stride, seq_length, pooled_seq_length
transform = [[1.0 if (i // pfac) == j else 0.0
for j in range(psl)]
for i in range(sl)]
transform = tf.constant(
transform,
dtype=tf.keras.mixed_precision.global_policy().compute_dtype)
pooling_transforms.append(transform / pool_stride)
seq_length = pooled_seq_length
return pooling_transforms
def _create_truncated_avg_masks(input_mask: tf.Tensor,
pool_strides: Sequence[int],
transforms: Sequence[tf.Tensor]):
"""Computes attention masks.
For [1,1,1,0,0]
Args:
input_mask: Tensor of shape [batch_size, seq_length].
pool_strides: Sequence of pooling strides for each layer.
transforms: Sequnce of off-diagonal matrices filling with 0.0 and
1/pool_stride.
Returns:
attention_masks: Sequence of attention masks for each layer.
"""
def create_2d_mask(from_length, mask):
return tf.einsum('F,BT->BFT', tf.ones([from_length], dtype=mask.dtype),
mask)
attention_masks = []
seq_length = tf.shape(input_mask)[-1]
layer_mask = tf.cast(
input_mask, dtype=tf.keras.mixed_precision.global_policy().compute_dtype)
for pool_stride, transform in zip(pool_strides, transforms):
if pool_stride == 1:
attention_masks.append(create_2d_mask(seq_length, layer_mask))
else:
pooled_seq_length = seq_length // pool_stride
attention_masks.append(create_2d_mask(pooled_seq_length, layer_mask))
layer_mask = tf.cast(
tf.einsum('BF,FT->BT', layer_mask, transform) > 0.0,
dtype=layer_mask.dtype)
seq_length = pooled_seq_length
del seq_length
return attention_masks
@tf.keras.utils.register_keras_serializable(package='Text')
class FunnelTransformerEncoder(tf.keras.layers.Layer):
"""Funnel Transformer-based encoder network.
......@@ -90,7 +183,7 @@ class FunnelTransformerEncoder(tf.keras.layers.Layer):
dropout.
attention_dropout: The dropout rate to use for the attention layers within
the transformer layers.
pool_type: Pooling type. Choose from ['max', 'avg'].
pool_type: Pooling type. Choose from ['max', 'avg', 'truncated_avg'].
pool_stride: An int or a list of ints. Pooling stride(s) to compress the
sequence length. If set to int, each layer will have the same stride size.
If set to list, the number of elements needs to match num_layers.
......@@ -124,7 +217,7 @@ class FunnelTransformerEncoder(tf.keras.layers.Layer):
inner_activation=lambda x: tf.keras.activations.gelu(x, approximate=True),
output_dropout=0.1,
attention_dropout=0.1,
pool_type='max',
pool_type=_MAX,
pool_stride=2,
unpool_length=0,
initializer=tf.keras.initializers.TruncatedNormal(stddev=0.02),
......@@ -207,12 +300,21 @@ class FunnelTransformerEncoder(tf.keras.layers.Layer):
raise ValueError('Lengths of pool_stride and num_layers are not equal.')
pool_strides = pool_stride
# TODO(crickwu): explore tf.keras.layers.serialize method.
if pool_type == 'max':
if pool_type == _MAX:
pool_cls = tf.keras.layers.MaxPooling1D
elif pool_type == 'avg':
elif pool_type == _AVG:
pool_cls = tf.keras.layers.AveragePooling1D
elif pool_type == _TRUNCATED_AVG:
# TODO(b/203665205): unpool_length should be implemented.
if unpool_length != 0:
raise ValueError('unpool_length is not supported by truncated_avg now.')
# Compute the attention masks and pooling transforms.
self._pooling_transforms = _create_truncated_avg_transforms(
max_sequence_length, pool_strides)
else:
raise ValueError('pool_type not supported.')
if pool_type in (_MAX, _AVG):
self._att_input_pool_layers = []
for layer_pool_stride in pool_strides:
att_input_pool_layer = pool_cls(
......@@ -224,6 +326,7 @@ class FunnelTransformerEncoder(tf.keras.layers.Layer):
self._pool_strides = pool_strides # This is a list here.
self._unpool_length = unpool_length
self._pool_type = pool_type
self._config = {
'vocab_size': vocab_size,
......@@ -280,11 +383,13 @@ class FunnelTransformerEncoder(tf.keras.layers.Layer):
encoder_outputs = []
x = embeddings
# TODO(b/195972228): attention_mask can be co-generated with pooling.
if self._pool_type in (_MAX, _AVG):
attention_mask = _pool_and_concat(
attention_mask,
unpool_length=self._unpool_length,
strides=self._pool_strides[0],
axes=[1])
for i, layer in enumerate(self._transformer_layers):
# Bypass no pooling cases.
if self._pool_strides[i] == 1:
......@@ -307,12 +412,36 @@ class FunnelTransformerEncoder(tf.keras.layers.Layer):
strides=[self._pool_strides[i + 1], self._pool_strides[i]],
axes=[1, 2])
encoder_outputs.append(x)
elif self._pool_type == _TRUNCATED_AVG:
attention_masks = _create_truncated_avg_masks(mask, self._pool_strides,
self._pooling_transforms)
for i, layer in enumerate(self._transformer_layers):
attention_mask = attention_masks[i]
# Bypass no pooling cases.
if self._pool_strides[i] == 1:
x = layer([x, x, attention_mask])
else:
pooled_inputs = tf.einsum(
'BFD,FT->BTD',
tf.cast(x[:, self._unpool_length:, :],
tf.keras.mixed_precision.global_policy().compute_dtype
), # extra casting for faster mixed computation.
self._pooling_transforms[i])
query_inputs = tf.concat(
values=(tf.cast(
x[:, :self._unpool_length, :],
dtype=pooled_inputs.dtype), pooled_inputs),
axis=1)
x = layer([query_inputs, x, attention_mask])
encoder_outputs.append(x)
last_encoder_output = encoder_outputs[-1]
first_token_tensor = last_encoder_output[:, 0, :]
pooled_output = self._pooler_layer(first_token_tensor)
return dict(
word_embeddings=word_embeddings,
embedding_output=embeddings,
sequence_output=encoder_outputs[-1],
pooled_output=pooled_output,
encoder_outputs=encoder_outputs)
......
official/nlp/modeling/networks/funnel_transformer_test.py
View file @ 7a45b513
......@@ -38,6 +38,8 @@ class FunnelTransformerEncoderTest(parameterized.TestCase, tf.test.TestCase):
tf.keras.mixed_precision.set_global_policy("float32")
@parameterized.named_parameters(
("mix_truncated_avg", "mixed_float16", tf.float16, "truncated_avg"),
("float32_truncated_avg", "float32", tf.float32, "truncated_avg"),
("mix_max", "mixed_float16", tf.float16, "max"),
("float32_max", "float32", tf.float32, "max"),
("mix_avg", "mixed_float16", tf.float16, "avg"),
......@@ -57,6 +59,7 @@ class FunnelTransformerEncoderTest(parameterized.TestCase, tf.test.TestCase):
num_layers=num_layers,
pool_stride=pool_stride,
pool_type=pool_type,
max_sequence_length=sequence_length,
unpool_length=0)
# Create the inputs (note that the first dimension is implicit).
word_ids = tf.keras.Input(shape=(sequence_length,), dtype=tf.int32)
......@@ -71,8 +74,14 @@ class FunnelTransformerEncoderTest(parameterized.TestCase, tf.test.TestCase):
self.assertIsInstance(test_network.pooler_layer, tf.keras.layers.Dense)
# Stride=2 compresses sequence length to half the size at each layer.
# This configuration gives each layer of seq length: 21->11->6->3.
# For pool_type = max or avg,
# this configuration gives each layer of seq length: 21->11->6->3.
# For pool_type = truncated_avg,
# seq length: 21->10->5->2.
if pool_type in ["max", "avg"]:
expected_data_shape = [None, 3, hidden_size]
else:
expected_data_shape = [None, 2, hidden_size]
expected_pooled_shape = [None, hidden_size]
self.assertAllEqual(expected_data_shape, data.shape.as_list())
......
official/nlp/projects/teams/experiments/base/glue_mnli.yaml
View file @ 7a45b513
......@@ -16,6 +16,7 @@ task:
seq_length: 128
trainer:
checkpoint_interval: 1000
continuous_eval_timeout: 7200
optimizer_config:
learning_rate:
polynomial:
......
official/nlp/projects/teams/experiments/base/squad_v1.yaml
View file @ 7a45b513
......@@ -23,6 +23,7 @@ task:
vocab_file: ''
trainer:
checkpoint_interval: 500
continuous_eval_timeout: 7200
max_to_keep: 5
optimizer_config:
learning_rate:
......
official/nlp/projects/teams/experiments/base/squad_v2.yaml
View file @ 7a45b513
......@@ -23,6 +23,7 @@ task:
vocab_file: ''
trainer:
checkpoint_interval: 500
continuous_eval_timeout: 7200
max_to_keep: 5
optimizer_config:
learning_rate:
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment