official/nlp/modeling/layers/tn_transformer_expand_condense.py
# Copyright 2024 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.
"""TN-BERT TNTransformerExpandCondense employing Expand-Condense layer instead of Dense."""
# pylint: disable=g-classes-have-attributes
# Import libraries
import gin
import tensorflow as tf, tf_keras
from official.modeling import tf_utils
from official.nlp.modeling.layers.tn_expand_condense import TNExpandCondense
@tf_keras.utils.register_keras_serializable(package="Text")
@gin.configurable
class TNTransformerExpandCondense(tf_keras.layers.Layer):
"""Transformer layer using tensor network Expand-Condense layer.
This layer implements the Transformer from transformer.py, with a single
tensor network layer replacing the usual intermediate and output Dense
layers.
Args:
num_attention_heads: Number of attention heads.
intermediate_size: Size of the intermediate layer.
intermediate_activation: Activation for the intermediate layer.
dropout_rate: Dropout probability for the post-attention and output dropout.
attention_dropout_rate: Dropout probability for within the attention layer.
output_range: the sequence output range, [0, output_range) by slicing the
target sequence. `None` means the target sequence is not sliced.
kernel_initializer: Initializer for dense layer kernels.
bias_initializer: Initializer for dense layer biases.
kernel_regularizer: Regularizer for dense layer kernels.
bias_regularizer: Regularizer for dense layer biases.
activity_regularizer: Regularizer for dense layer activity.
kernel_constraint: Constraint for dense layer kernels.
bias_constraint: Constraint for dense layer kernels.
use_bias: Whether to enable use_bias in attention layer. If set to False,
use_bias in attention layer is disabled.
norm_first: Whether to normalize inputs to attention and intermediate dense
layers. If set False, output of attention and intermediate dense layers is
normalized.
norm_epsilon: Epsilon value to initialize normalization layers.
intermediate_dropout: Dropout probability for intermediate_dropout_layer.
attention_initializer: Initializer for kernels of attention layers. If set
`None`, attention layers use kernel_initializer as initializer for kernel.
"""
def __init__(self,
num_attention_heads,
intermediate_size,
intermediate_activation,
dropout_rate=0.0,
attention_dropout_rate=0.0,
output_range=None,
kernel_initializer="glorot_uniform",
bias_initializer="zeros",
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
use_bias=True,
norm_first=False,
norm_epsilon=1e-12,
intermediate_dropout=0.0,
attention_initializer=None,
**kwargs):
super().__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 = tf_utils.clone_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(
"TNTransformerExpandCondense 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 TNTransformerExpandCondense, 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(
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,
bias_initializer=tf_utils.clone_initializer(self._bias_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))
# Substitute Dense layers with a single Expand-Condense layer.
self._output_dense = TNExpandCondense(
4,
use_bias=True,
activation=self._intermediate_activation,
kernel_initializer=self._kernel_initializer,
bias_initializer=self._bias_initializer)
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().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().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)
layer_output = self._output_dense(attention_output)
layer_output = self._output_dropout(layer_output)
# During mixed precision training, attention_output is from layer norm and
# is always fp32 for now. Cast layer_output to fp32 for the subsequent
# add.
layer_output = tf.cast(layer_output, tf.float32)
if self._norm_first:
layer_output = source_attention_output + layer_output
else:
layer_output = self._output_layer_norm(layer_output + attention_output)
return layer_output