Skip to content

GitLab

Commit ed38358f authored by Zhenyu Tan's avatar Zhenyu Tan Committed by A. Unique TensorFlower
Browse files

Internal change 

PiperOrigin-RevId: 328674302
parent ff22264d
Hide whitespace changes
Inline Side-by-side
Showing with 26 additions and 431 deletions
official/nlp/keras_nlp/layers/transformer_encoder_block.py
View file @ ed38358f
......@@ -158,7 +158,7 @@ class TransformerEncoderBlock(tf.keras.layers.Layer):
axis=-1,
epsilon=self._norm_epsilon,
dtype=tf.float32))
self._inner_dense = tf.keras.layers.experimental.EinsumDense(
self._intermediate_dense = tf.keras.layers.experimental.EinsumDense(
"abc,cd->abd",
output_shape=(None, self._inner_dim),
bias_axes="d",
......@@ -171,7 +171,7 @@ class TransformerEncoderBlock(tf.keras.layers.Layer):
# as well, so we use float32.
# TODO(b/154538392): Investigate this.
policy = tf.float32
self._inner_activation_layer = tf.keras.layers.Activation(
self._intermediate_activation_layer = tf.keras.layers.Activation(
self._inner_activation, dtype=policy)
self._inner_dropout_layer = tf.keras.layers.Dropout(
rate=self._inner_dropout)
......@@ -260,8 +260,8 @@ class TransformerEncoderBlock(tf.keras.layers.Layer):
if self._norm_first:
source_attention_output = attention_output
attention_output = self._output_layer_norm(attention_output)
inner_output = self._inner_dense(attention_output)
inner_output = self._inner_activation_layer(inner_output)
inner_output = self._intermediate_dense(attention_output)
inner_output = self._intermediate_activation_layer(inner_output)
inner_output = self._inner_dropout_layer(inner_output)
layer_output = self._output_dense(inner_output)
layer_output = self._output_dropout(layer_output)
......
official/nlp/modeling/layers/transformer.py
View file @ ed38358f
......@@ -18,13 +18,14 @@
import gin
import tensorflow as tf
from official.nlp.keras_nlp.layers import transformer_encoder_block
from official.nlp.modeling.layers import attention
from official.nlp.modeling.layers import multi_channel_attention
from official.nlp.modeling.layers.util import tf_function_if_eager
@tf.keras.utils.register_keras_serializable(package="Text")
class Transformer(tf.keras.layers.Layer):
class Transformer(transformer_encoder_block.TransformerEncoderBlock):
"""Transformer layer.
This layer implements the Transformer from "Attention Is All You Need".
......@@ -76,194 +77,26 @@ class Transformer(tf.keras.layers.Layer):
intermediate_dropout=0.0,
attention_initializer=None,
**kwargs):
super(Transformer, self).__init__(**kwargs)
self._num_heads = num_attention_heads
self._intermediate_size = intermediate_size
self._intermediate_activation = intermediate_activation
self._attention_dropout_rate = attention_dropout_rate
self._dropout_rate = dropout_rate
self._output_range = output_range
self._kernel_initializer = tf.keras.initializers.get(kernel_initializer)
self._bias_initializer = tf.keras.initializers.get(bias_initializer)
self._kernel_regularizer = tf.keras.regularizers.get(kernel_regularizer)
self._bias_regularizer = tf.keras.regularizers.get(bias_regularizer)
self._activity_regularizer = tf.keras.regularizers.get(activity_regularizer)
self._kernel_constraint = tf.keras.constraints.get(kernel_constraint)
self._bias_constraint = tf.keras.constraints.get(bias_constraint)
self._use_bias = use_bias
self._norm_first = norm_first
self._norm_epsilon = norm_epsilon
self._intermediate_dropout = intermediate_dropout
if attention_initializer:
self._attention_initializer = tf.keras.initializers.get(
attention_initializer)
else:
self._attention_initializer = self._kernel_initializer
def build(self, input_shape):
input_tensor = input_shape[0] if len(input_shape) == 2 else input_shape
input_tensor_shape = tf.TensorShape(input_tensor)
if len(input_tensor_shape.as_list()) != 3:
raise ValueError("TransformerLayer expects a three-dimensional input of "
"shape [batch, sequence, width].")
batch_size, sequence_length, hidden_size = input_tensor_shape
if len(input_shape) == 2:
mask_tensor_shape = tf.TensorShape(input_shape[1])
expected_mask_tensor_shape = tf.TensorShape(
[batch_size, sequence_length, sequence_length])
if not expected_mask_tensor_shape.is_compatible_with(mask_tensor_shape):
raise ValueError("When passing a mask tensor to TransformerLayer, the "
"mask tensor must be of shape [batch, "
"sequence_length, sequence_length] (here %s). Got a "
"mask tensor of shape %s." %
(expected_mask_tensor_shape, mask_tensor_shape))
if hidden_size % self._num_heads != 0:
raise ValueError(
"The input size (%d) is not a multiple of the number of attention "
"heads (%d)" % (hidden_size, self._num_heads))
self._attention_head_size = int(hidden_size // self._num_heads)
common_kwargs = dict(
bias_initializer=self._bias_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer,
activity_regularizer=self._activity_regularizer,
kernel_constraint=self._kernel_constraint,
bias_constraint=self._bias_constraint)
self._attention_layer = tf.keras.layers.MultiHeadAttention(
num_heads=self._num_heads,
key_dim=self._attention_head_size,
dropout=self._attention_dropout_rate,
use_bias=self._use_bias,
kernel_initializer=self._attention_initializer,
name="self_attention",
**common_kwargs)
self._attention_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
# Use float32 in layernorm for numeric stability.
# It is probably safe in mixed_float16, but we haven't validated this yet.
self._attention_layer_norm = (
tf.keras.layers.LayerNormalization(
name="self_attention_layer_norm",
axis=-1,
epsilon=self._norm_epsilon,
dtype=tf.float32))
self._intermediate_dense = tf.keras.layers.experimental.EinsumDense(
"abc,cd->abd",
output_shape=(None, self._intermediate_size),
bias_axes="d",
kernel_initializer=self._kernel_initializer,
name="intermediate",
**common_kwargs)
policy = tf.keras.mixed_precision.experimental.global_policy()
if policy.name == "mixed_bfloat16":
# bfloat16 causes BERT with the LAMB optimizer to not converge
# as well, so we use float32.
# TODO(b/154538392): Investigate this.
policy = tf.float32
self._intermediate_activation_layer = tf.keras.layers.Activation(
self._intermediate_activation, dtype=policy)
self._intermediate_dropout_layer = tf.keras.layers.Dropout(
rate=self._intermediate_dropout)
self._output_dense = tf.keras.layers.experimental.EinsumDense(
"abc,cd->abd",
output_shape=(None, hidden_size),
bias_axes="d",
name="output",
kernel_initializer=self._kernel_initializer,
**common_kwargs)
self._output_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
# Use float32 in layernorm for numeric stability.
self._output_layer_norm = tf.keras.layers.LayerNormalization(
name="output_layer_norm",
axis=-1,
epsilon=self._norm_epsilon,
dtype=tf.float32)
super(Transformer, self).build(input_shape)
def get_config(self):
config = {
"num_attention_heads":
self._num_heads,
"intermediate_size":
self._intermediate_size,
"intermediate_activation":
self._intermediate_activation,
"dropout_rate":
self._dropout_rate,
"attention_dropout_rate":
self._attention_dropout_rate,
"output_range":
self._output_range,
"kernel_initializer":
tf.keras.initializers.serialize(self._kernel_initializer),
"bias_initializer":
tf.keras.initializers.serialize(self._bias_initializer),
"kernel_regularizer":
tf.keras.regularizers.serialize(self._kernel_regularizer),
"bias_regularizer":
tf.keras.regularizers.serialize(self._bias_regularizer),
"activity_regularizer":
tf.keras.regularizers.serialize(self._activity_regularizer),
"kernel_constraint":
tf.keras.constraints.serialize(self._kernel_constraint),
"bias_constraint":
tf.keras.constraints.serialize(self._bias_constraint),
"use_bias":
self._use_bias,
"norm_first":
self._norm_first,
"norm_epsilon":
self._norm_epsilon,
"intermediate_dropout":
self._intermediate_dropout,
"attention_initializer":
tf.keras.initializers.serialize(self._attention_initializer)
}
base_config = super(Transformer, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs):
if isinstance(inputs, (list, tuple)) and len(inputs) == 2:
input_tensor, attention_mask = inputs
else:
input_tensor, attention_mask = (inputs, None)
if self._output_range:
target_tensor = input_tensor[:, 0:self._output_range, :]
attention_mask = attention_mask[:, 0:self._output_range, :]
else:
if self._norm_first:
source_tensor = input_tensor
input_tensor = self._attention_layer_norm(input_tensor)
target_tensor = input_tensor
attention_output = self._attention_layer(
query=target_tensor, value=input_tensor, attention_mask=attention_mask)
attention_output = self._attention_dropout(attention_output)
if self._norm_first:
attention_output = source_tensor + attention_output
else:
attention_output = self._attention_layer_norm(target_tensor +
attention_output)
if self._norm_first:
source_attention_output = attention_output
attention_output = self._output_layer_norm(attention_output)
intermediate_output = self._intermediate_dense(attention_output)
intermediate_output = self._intermediate_activation_layer(
intermediate_output)
intermediate_output = self._intermediate_dropout_layer(intermediate_output)
layer_output = self._output_dense(intermediate_output)
layer_output = self._output_dropout(layer_output)
if self._norm_first:
return source_attention_output + layer_output
# During mixed precision training, layer norm output is always fp32 for now.
# Casts fp32 for the subsequent add.
layer_output = tf.cast(layer_output, tf.float32)
return self._output_layer_norm(layer_output + attention_output)
super(Transformer, self).__init__(
num_attention_heads=num_attention_heads,
inner_dim=intermediate_size,
inner_activation=intermediate_activation,
output_dropout=dropout_rate,
attention_dropout=attention_dropout_rate,
output_range=output_range,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
kernel_constraint=kernel_constraint,
bias_constraint=bias_constraint,
use_bias=use_bias,
norm_first=norm_first,
norm_epsilon=norm_epsilon,
inner_dropout=intermediate_dropout,
attention_initializer=attention_initializer,
**kwargs)
@tf.keras.utils.register_keras_serializable(package="Text")
......
official/nlp/modeling/layers/transformer_test.py
View file @ ed38358f
......@@ -14,250 +14,12 @@
# ==============================================================================
"""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.modeling.layers import transformer
# 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
@parameterized.named_parameters(('base', transformer.Transformer),
('xla', transformer.CompiledTransformer))
class TransformerLayerTest(keras_parameterized.TestCase):
def tearDown(self):
super(TransformerLayerTest, self).tearDown()
tf.keras.mixed_precision.experimental.set_policy('float32')
def test_layer_creation(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_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,
intermediate_size=2048,
intermediate_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_creation_with_incorrect_mask_fails(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_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 - 3))
with self.assertRaisesRegex(ValueError, 'When passing a mask tensor.*'):
_ = test_layer([data_tensor, mask_tensor])
def test_layer_invocation(self, transformer_cls):
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_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,
intermediate_size=2048,
intermediate_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, _):
# XLA has an obvious numeric issue in this test case.
test_layer = transformer.Transformer(
num_attention_heads=10,
intermediate_size=2048,
intermediate_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
# embeeding.
new_layer = transformer.Transformer(
num_attention_heads=10,
intermediate_size=2048,
intermediate_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_invocation_with_float16_dtype(self, transformer_cls):
tf.keras.mixed_precision.experimental.set_policy('mixed_float16')
test_layer = transformer_cls(
num_attention_heads=10,
intermediate_size=2048,
intermediate_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,
intermediate_size=2048,
intermediate_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,
intermediate_size=2048,
intermediate_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)
@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 = transformer.Transformer(
num_attention_heads=num_attention_heads,
intermediate_size=32,
intermediate_activation='relu',
dropout_rate=0.1,
attention_dropout_rate=0.1,
use_bias=False,
norm_first=True,
norm_epsilon=1e-6,
intermediate_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 = transformer.Transformer(
num_attention_heads=num_attention_heads,
intermediate_size=32,
intermediate_activation='relu',
dropout_rate=0.1,
attention_dropout_rate=0.1,
use_bias=False,
norm_first=True,
norm_epsilon=1e-6,
intermediate_dropout=0.1,
attention_initializer=tf.keras.initializers.RandomUniform(
minval=0., maxval=1.))
encoder_block_config = encoder_block.get_config()
new_encoder_block = transformer.Transformer.from_config(
encoder_block_config)
self.assertEqual(encoder_block_config, new_encoder_block.get_config())
def _create_cache(batch_size, init_decode_length, num_heads, head_size):
return {
'key':
......
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