BERT代码解读(2)-模型

google开源的tensorflow版本的bert 源码见 https://github.com/google-research/bert。本文主要对该官方代码的一些关键部分进行解读。

上一节我们分析了从原始数据集到能够输入bert的数据特征的处理过程，最终每条样本都得到的对应的input_ids, input_mask, segment_ids这三个特征和对应的标签label_ids。本节我们分析将这些特征输入bert网络的过程以及解读bert的网络结构。

使用EstimatorAPI的关键是实现model_fn和input_fn函数，input_fn在此不再赘述，我们分析model_fn：

def model_fn_builder(bert_config, num_labels, init_checkpoint, learning_rate,
                     num_train_steps, num_warmup_steps, use_tpu,
                     use_one_hot_embeddings):
    """Returns `model_fn` closure for TPUEstimator."""

    def model_fn(features, labels, mode, params):  # pylint: disable=unused-argument
        """The `model_fn` for TPUEstimator."""
        tf.logging.info("*** Features ***")
        for name in sorted(features.keys()):
            tf.logging.info("  name = %s, shape = %s" % (name, features[name].shape))

        input_ids = features["input_ids"]
        input_mask = features["input_mask"]
        segment_ids = features["segment_ids"]
        label_ids = features["label_ids"]
        is_real_example = None
        if "is_real_example" in features:
            is_real_example = tf.cast(features["is_real_example"], dtype=tf.float32)
        else:
            is_real_example = tf.ones(tf.shape(label_ids), dtype=tf.float32)

        is_training = (mode == tf.estimator.ModeKeys.TRAIN)

        # 核心部分 创建模型
        (total_loss, per_example_loss, logits, probabilities) = create_model(
            bert_config, is_training, input_ids, input_mask, segment_ids, label_ids,
            num_labels, use_one_hot_embeddings)

        tvars = tf.trainable_variables()
        initialized_variable_names = {}
        scaffold_fn = None
        # 从checkpoint中恢复参数
        if init_checkpoint:
            (assignment_map, initialized_variable_names
             ) = modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint)
            if use_tpu:

                def tpu_scaffold():
                    tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
                    return tf.train.Scaffold()

                scaffold_fn = tpu_scaffold
            else:
                tf.train.init_from_checkpoint(init_checkpoint, assignment_map)

        tf.logging.info("**** Trainable Variables ****")
        for var in tvars:
            init_string = ""
            if var.name in initialized_variable_names:
                init_string = ", *INIT_FROM_CKPT*"
            tf.logging.info("  name = %s, shape = %s%s", var.name, var.shape,
                            init_string)

        output_spec = None
        if mode == tf.estimator.ModeKeys.TRAIN:

            train_op = optimization.create_optimizer(
                total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu)

            output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                mode=mode,
                loss=total_loss,
                train_op=train_op,
                scaffold_fn=scaffold_fn)
        elif mode == tf.estimator.ModeKeys.EVAL:

            def metric_fn(per_example_loss, label_ids, logits, is_real_example):
                predictions = tf.argmax(logits, axis=-1, output_type=tf.int32)
                accuracy = tf.metrics.accuracy(
                    labels=label_ids, predictions=predictions, weights=is_real_example)
                loss = tf.metrics.mean(values=per_example_loss, weights=is_real_example)
                return {
                    "eval_accuracy": accuracy,
                    "eval_loss": loss,
                }

            eval_metrics = (metric_fn,
                            [per_example_loss, label_ids, logits, is_real_example])
            output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                mode=mode,
                loss=total_loss,
                eval_metrics=eval_metrics,
                scaffold_fn=scaffold_fn)
        else:
            output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                mode=mode,
                predictions={"probabilities": probabilities},
                scaffold_fn=scaffold_fn)
        return output_spec

    return model_fn

可以看出，核心代码是create_model用来创建模型，其余大部分是模板代码。

def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
                 labels, num_labels, use_one_hot_embeddings):
    """Creates a classification model."""
    model = modeling.BertModel(
        config=bert_config,
        is_training=is_training,
        input_ids=input_ids,
        input_mask=input_mask,
        token_type_ids=segment_ids,
        use_one_hot_embeddings=use_one_hot_embeddings)

    # In the demo, we are doing a simple classification task on the entire segment.
    # If you want to use the token-level output, use model.get_sequence_output() instead.
    output_layer = model.get_pooled_output()

    hidden_size = output_layer.shape[-1].value

    output_weights = tf.get_variable(
        "output_weights", [num_labels, hidden_size],
        initializer=tf.truncated_normal_initializer(stddev=0.02))

    output_bias = tf.get_variable(
        "output_bias", [num_labels], initializer=tf.zeros_initializer())

    with tf.variable_scope("loss"):
        if is_training:
            # I.e., 0.1 dropout
            output_layer = tf.nn.dropout(output_layer, keep_prob=0.9)

        logits = tf.matmul(output_layer, output_weights, transpose_b=True)
        logits = tf.nn.bias_add(logits, output_bias)
        probabilities = tf.nn.softmax(logits, axis=-1)
        log_probs = tf.nn.log_softmax(logits, axis=-1)

        one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)

        per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)
        loss = tf.reduce_mean(per_example_loss)

        return (loss, per_example_loss, logits, probabilities)

该函数通过modeling.BertModel得到bert模型，然后通过get_pooled_output得到最后一层[CLS]token对应的输出向量，之后接一个全连接层得到logits，进而通过softmax得到各类别对应的概率分数，最后计算相应的loss。

因此模型网络结构的关键代码是modeling.BertModel:

class BertModel(object):

  def __init__(self,
               config,
               is_training,
               input_ids,
               input_mask=None,
               token_type_ids=None,
               use_one_hot_embeddings=False,
               scope=None):
    """Constructor for BertModel.

    Args:
      config: `BertConfig` instance.
      is_training: bool. true for training model, false for eval model. Controls
        whether dropout will be applied.
      input_ids: int32 Tensor of shape [batch_size, seq_length].
      input_mask: (optional) int32 Tensor of shape [batch_size, seq_length].
      token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
      use_one_hot_embeddings: (optional) bool. Whether to use one-hot word
        embeddings or tf.embedding_lookup() for the word embeddings.
      scope: (optional) variable scope. Defaults to "bert".

    Raises:
      ValueError: The config is invalid or one of the input tensor shapes
        is invalid.
    """
    config = copy.deepcopy(config)
    if not is_training:
      config.hidden_dropout_prob = 0.0
      config.attention_probs_dropout_prob = 0.0

    input_shape = get_shape_list(input_ids, expected_rank=2)
    batch_size = input_shape[0]
    seq_length = input_shape[1]

    if input_mask is None:
      input_mask = tf.ones(shape=[batch_size, seq_length], dtype=tf.int32)

    if token_type_ids is None:
      token_type_ids = tf.zeros(shape=[batch_size, seq_length], dtype=tf.int32)

    with tf.variable_scope(scope, default_name="bert"):
      with tf.variable_scope("embeddings"):
        # Perform embedding lookup on the word ids.
        # embedding_output维度为[batch_size, seq_length, embedding_size]，比如[4, 128, 768]
        (self.embedding_output, self.embedding_table) = embedding_lookup(
            input_ids=input_ids,
            vocab_size=config.vocab_size,
            embedding_size=config.hidden_size,
            initializer_range=config.initializer_range,
            word_embedding_name="word_embeddings",
            use_one_hot_embeddings=use_one_hot_embeddings)

        # Add positional embeddings and token type embeddings, then layer
        # normalize and perform dropout.
        self.embedding_output = embedding_postprocessor(
            input_tensor=self.embedding_output,
            use_token_type=True,
            token_type_ids=token_type_ids,
            token_type_vocab_size=config.type_vocab_size,
            token_type_embedding_name="token_type_embeddings",
            use_position_embeddings=True,
            position_embedding_name="position_embeddings",
            initializer_range=config.initializer_range,
            max_position_embeddings=config.max_position_embeddings,
            dropout_prob=config.hidden_dropout_prob)

      with tf.variable_scope("encoder"):
        # This converts a 2D mask of shape [batch_size, seq_length] to a 3D
        # mask of shape [batch_size, seq_length, seq_length] which is used
        # for the attention scores.
        attention_mask = create_attention_mask_from_input_mask(
            input_ids, input_mask)

        # Run the stacked transformer.
        # `sequence_output` shape = [batch_size, seq_length, hidden_size].
        self.all_encoder_layers = transformer_model(
            input_tensor=self.embedding_output,
            attention_mask=attention_mask,
            hidden_size=config.hidden_size,
            num_hidden_layers=config.num_hidden_layers,
            num_attention_heads=config.num_attention_heads,
            intermediate_size=config.intermediate_size,
            intermediate_act_fn=get_activation(config.hidden_act),
            hidden_dropout_prob=config.hidden_dropout_prob,
            attention_probs_dropout_prob=config.attention_probs_dropout_prob,
            initializer_range=config.initializer_range,
            do_return_all_layers=True)

      self.sequence_output = self.all_encoder_layers[-1]
      # The "pooler" converts the encoded sequence tensor of shape
      # [batch_size, seq_length, hidden_size] to a tensor of shape
      # [batch_size, hidden_size]. This is necessary for segment-level
      # (or segment-pair-level) classification tasks where we need a fixed
      # dimensional representation of the segment.
      with tf.variable_scope("pooler"):
        # We "pool" the model by simply taking the hidden state corresponding
        # to the first token. We assume that this has been pre-trained
        first_token_tensor = tf.squeeze(self.sequence_output[:, 0:1, :], axis=1)
        self.pooled_output = tf.layers.dense(
            first_token_tensor,
            config.hidden_size,
            activation=tf.tanh,
            kernel_initializer=create_initializer(config.initializer_range))

  def get_pooled_output(self):
    return self.pooled_output

  def get_sequence_output(self):
    """Gets final hidden layer of encoder.

    Returns:
      float Tensor of shape [batch_size, seq_length, hidden_size] corresponding
      to the final hidden of the transformer encoder.
    """
    return self.sequence_output

  def get_all_encoder_layers(self):
    return self.all_encoder_layers

  def get_embedding_output(self):
    """Gets output of the embedding lookup (i.e., input to the transformer).

    Returns:
      float Tensor of shape [batch_size, seq_length, hidden_size] corresponding
      to the output of the embedding layer, after summing the word
      embeddings with the positional embeddings and the token type embeddings,
      then performing layer normalization. This is the input to the transformer.
    """
    return self.embedding_output

  def get_embedding_table(self):
    return self.embedding_table

(1) 代码首先把配置参数config深拷贝一份；当前若不是训练阶段，则把网络层中的dropout置为0；如果input_mask为None，则构造一个shape合适的且值全为1的张量，表示输入的token都来自真实句子，而不是padding的；如果token_type_ids为None，则构造一个shaple合适的且值全为0的张量，表示输入的tokens都属于第一个句子。

(2) 然后得到输入网络的词向量embeddings：

• embedding_lookup根据输入tokens的input_ids找到对应的词向量，得到token/word_embeddings，shape为[batch_size, seq_length, embedding_size]。
• embedding_postprocessor通过token_type_ids得到对应的segment/token_type_embeddings词向量，shape为[batch_size, seq_length, embedding_size]。
• embedding_postprocessor从full_position_embeddings(shape=[max_position_embedding, embedding_size])表中查找得到句子对应的position_embeddings。需要注意的是，position_embeddings与输入句子的内容无关，因此shape是[seq_length, embedding_size]
• 最后将三种词向量相加并过一个layer normalization和dropout层。注意position_embeddings需要broadcasting后再与另两个词向量相加。
  由此得到送入bert网络的词向量，如下图所示：
  BERT input representation

(3) 之后构建Transformer的encoder网络结构。关键代码是transformer_model，将Transformer子层堆叠在一起，得到最终的all_encoder_layers，其为一个长度为num_hidden_layers的列表，列表每个元素对应网络的每一层输出，shape均为[batch_size, seq_length, hidden_size]。我们将Transformer网络结构的讲解放到本文最后。

(4)sequence_output对应网络最后一层的输出，shape为[batch_size, seq_length, hidden_size]。first_token_tensor是网络最后一层的第一个token([CLS])处的输出，shape为[batch_size, hidden_size]，最后通过tf.layers.dense对其加一个维度保持不变的全连接层，得到pooled_output，shape依然是[batch_size, hidden_size]。

Transformer的网络结构

在分析Transformer的网络结构前，我们先来了解下attention_mask。

attention_mask = create_attention_mask_from_input_mask(
            input_ids, input_mask)

上述构造输入网络的词向量时我们用到了之前提取的特征input_ids和segment_ids，而这里构造attention mask矩阵则用到了特征input_mask。在计算self-attention时每个样本都需要一个attention mask矩阵，表示每个时刻可以attend to的范围，input_mask=1表示是来自句子的真实token，可以attend to；input_mask=0表示是padding出来的token，不可以attend to。

例如：当input_ids=[1,2,3,0,0], input_mask=[1,1,1,0,0]时，对应的attention_mask矩阵为：

[1, 1, 1, 0, 0], #它表示第1个词可以attend to 3个词
[1, 1, 1, 0, 0], #它表示第2个词可以attend to 3个词
[1, 1, 1, 0, 0], #它表示第3个词可以attend to 3个词
[1, 1, 1, 0, 0], #无意义，因为输入第4个词是padding的
[1, 1, 1, 0, 0]  #无意义，因为输入第5个词是padding的

encoder

接下来我们来看Transformer encoder部分的网络结构，论文出处 Attention is all your need

      # Run the stacked transformer.
      self.all_encoder_layers = transformer_model(
            input_tensor=self.embedding_output,
            attention_mask=attention_mask,
            hidden_size=config.hidden_size,
            num_hidden_layers=config.num_hidden_layers,
            num_attention_heads=config.num_attention_heads,
            intermediate_size=config.intermediate_size,
            intermediate_act_fn=get_activation(config.hidden_act),
            hidden_dropout_prob=config.hidden_dropout_prob,
            attention_probs_dropout_prob=config.attention_probs_dropout_prob,
            initializer_range=config.initializer_range,
            do_return_all_layers=True)

def transformer_model(input_tensor,
                      attention_mask=None,
                      hidden_size=768,
                      num_hidden_layers=12,
                      num_attention_heads=12,
                      intermediate_size=3072,
                      intermediate_act_fn=gelu,
                      hidden_dropout_prob=0.1,
                      attention_probs_dropout_prob=0.1,
                      initializer_range=0.02,
                      do_return_all_layers=False):
  """Multi-headed, multi-layer Transformer from "Attention is All You Need".

  This is almost an exact implementation of the original Transformer encoder.

  See the original paper:
  https://arxiv.org/abs/1706.03762

  Also see:
  https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/models/transformer.py

  Args:
    input_tensor: float Tensor of shape [batch_size, seq_length, hidden_size].
    attention_mask: (optional) int32 Tensor of shape [batch_size, seq_length,
      seq_length], with 1 for positions that can be attended to and 0 in
      positions that should not be.
    hidden_size: int. Hidden size of the Transformer.
    num_hidden_layers: int. Number of layers (blocks) in the Transformer.
    num_attention_heads: int. Number of attention heads in the Transformer.
    intermediate_size: int. The size of the "intermediate" (a.k.a., feed
      forward) layer.
    intermediate_act_fn: function. The non-linear activation function to apply
      to the output of the intermediate/feed-forward layer.
    hidden_dropout_prob: float. Dropout probability for the hidden layers.
    attention_probs_dropout_prob: float. Dropout probability of the attention
      probabilities.
    initializer_range: float. Range of the initializer (stddev of truncated
      normal).
    do_return_all_layers: Whether to also return all layers or just the final
      layer.

  Returns:
    float Tensor of shape [batch_size, seq_length, hidden_size], the final
    hidden layer of the Transformer.

  Raises:
    ValueError: A Tensor shape or parameter is invalid.
  """
  if hidden_size % num_attention_heads != 0:
    raise ValueError(
        "The hidden size (%d) is not a multiple of the number of attention "
        "heads (%d)" % (hidden_size, num_attention_heads))

  ## 因为最终要输出hidden_size大小，总共有num_attention_heads个head，
  ## 因此每个head输出大小为 (hidden_size/num_attention_heads)
  attention_head_size = int(hidden_size / num_attention_heads)
  input_shape = get_shape_list(input_tensor, expected_rank=3)
  batch_size = input_shape[0]
  seq_length = input_shape[1]
  input_width = input_shape[2]

  # The Transformer performs sum residuals on all layers so the input needs
  # to be the same as the hidden size.
  if input_width != hidden_size:
    raise ValueError("The width of the input tensor (%d) != hidden size (%d)" %
                     (input_width, hidden_size))

  ## 为了避免在2D和3D之间来回reshape，统一把所有3D tensor转成2D。
  ## 虽然reshape在GPU/CPU很快，但在TPU却不是这样，这样做的目的是为了优化TPU
  ## 比如input_tensor为[4, 128, 768]，则prev_output为[4*128, 768]
  prev_output = reshape_to_matrix(input_tensor)

  all_layer_outputs = []
  for layer_idx in range(num_hidden_layers):
    ## 每一层都有自己的variable_scope
    with tf.variable_scope("layer_%d" % layer_idx):
      layer_input = prev_output

      with tf.variable_scope("attention"):
        attention_heads = []
        with tf.variable_scope("self"):
          attention_head = attention_layer(
              from_tensor=layer_input,
              to_tensor=layer_input,
              attention_mask=attention_mask,
              num_attention_heads=num_attention_heads,
              size_per_head=attention_head_size,
              attention_probs_dropout_prob=attention_probs_dropout_prob,
              initializer_range=initializer_range,
              do_return_2d_tensor=True,
              batch_size=batch_size,
              from_seq_length=seq_length,
              to_seq_length=seq_length)
          attention_heads.append(attention_head)

        attention_output = None
        if len(attention_heads) == 1:
          attention_output = attention_heads[0]
        else:
          # In the case where we have other sequences, we just concatenate
          # them to the self-attention head before the projection.
          attention_output = tf.concat(attention_heads, axis=-1)

        ## 使用线性变换把前面的输出变成hidden_size，然后加上layer_input形成残差连接
        with tf.variable_scope("output"):
          attention_output = tf.layers.dense(
              attention_output,
              hidden_size,
              kernel_initializer=create_initializer(initializer_range))
          attention_output = dropout(attention_output, hidden_dropout_prob)
          attention_output = layer_norm(attention_output + layer_input)

      # 全连接层
      # The activation is only applied to the "intermediate" hidden layer.
      with tf.variable_scope("intermediate"):
        intermediate_output = tf.layers.dense(
            attention_output,
            intermediate_size,
            activation=intermediate_act_fn,
            kernel_initializer=create_initializer(initializer_range))

      # Down-project back to `hidden_size` then add the residual.
      with tf.variable_scope("output"):
        layer_output = tf.layers.dense(
            intermediate_output,
            hidden_size,
            kernel_initializer=create_initializer(initializer_range))
        layer_output = dropout(layer_output, hidden_dropout_prob)
        layer_output = layer_norm(layer_output + attention_output)
        prev_output = layer_output
        all_layer_outputs.append(layer_output)

  if do_return_all_layers:
    final_outputs = []
    for layer_output in all_layer_outputs:
      final_output = reshape_from_matrix(layer_output, input_shape)
      final_outputs.append(final_output)
    return final_outputs
  else:
    final_output = reshape_from_matrix(prev_output, input_shape)
    return final_output

当参数do_return_all_layers=True时返回所有层的输出，结果是一个长度为num_hidden_layers的列表，列表里每个元素的shape都是[batch_size, seq_length, hidden_size]；否则只返回网络最后一层的输出，shape为[batch_size, seq_length, hidden_size]。

transformer_model的网络结构即下图左侧Transformer encoder部分的结构：

Transformer architecture
注意这里的Norm采用的是layer normalization。

接下来我们详细分析self attention层：

def attention_layer(from_tensor,
                    to_tensor,
                    attention_mask=None,
                    num_attention_heads=1,
                    size_per_head=512,
                    query_act=None,
                    key_act=None,
                    value_act=None,
                    attention_probs_dropout_prob=0.0,
                    initializer_range=0.02,
                    do_return_2d_tensor=False,
                    batch_size=None,
                    from_seq_length=None,
                    to_seq_length=None):
  """Performs multi-headed attention from `from_tensor` to `to_tensor`.

  This is an implementation of multi-headed attention based on "Attention
  is all you Need". If `from_tensor` and `to_tensor` are the same, then
  this is self-attention. Each timestep in `from_tensor` attends to the
  corresponding sequence in `to_tensor`, and returns a fixed-with vector.

  This function first projects `from_tensor` into a "query" tensor and
  `to_tensor` into "key" and "value" tensors. These are (effectively) a list
  of tensors of length `num_attention_heads`, where each tensor is of shape
  [batch_size, seq_length, size_per_head].

  Then, the query and key tensors are dot-producted and scaled. These are
  softmaxed to obtain attention probabilities. The value tensors are then
  interpolated by these probabilities, then concatenated back to a single
  tensor and returned.

  In practice, the multi-headed attention are done with transposes and
  reshapes rather than actual separate tensors.

  Args:
    from_tensor: float Tensor of shape [batch_size, from_seq_length,
      from_width].
    to_tensor: float Tensor of shape [batch_size, to_seq_length, to_width].
    attention_mask: (optional) int32 Tensor of shape [batch_size,
      from_seq_length, to_seq_length]. The values should be 1 or 0. The
      attention scores will effectively be set to -infinity for any positions in
      the mask that are 0, and will be unchanged for positions that are 1.
    num_attention_heads: int. Number of attention heads.
    size_per_head: int. Size of each attention head.
    query_act: (optional) Activation function for the query transform.
    key_act: (optional) Activation function for the key transform.
    value_act: (optional) Activation function for the value transform.
    attention_probs_dropout_prob: (optional) float. Dropout probability of the
      attention probabilities.
    initializer_range: float. Range of the weight initializer.
    do_return_2d_tensor: bool. If True, the output will be of shape [batch_size
      * from_seq_length, num_attention_heads * size_per_head]. If False, the
      output will be of shape [batch_size, from_seq_length, num_attention_heads
      * size_per_head].
    batch_size: (Optional) int. If the input is 2D, this might be the batch size
      of the 3D version of the `from_tensor` and `to_tensor`.
    from_seq_length: (Optional) If the input is 2D, this might be the seq length
      of the 3D version of the `from_tensor`.
    to_seq_length: (Optional) If the input is 2D, this might be the seq length
      of the 3D version of the `to_tensor`.

  Returns:
    float Tensor of shape [batch_size, from_seq_length,
      num_attention_heads * size_per_head]. (If `do_return_2d_tensor` is
      true, this will be of shape [batch_size * from_seq_length,
      num_attention_heads * size_per_head]).

  Raises:
    ValueError: Any of the arguments or tensor shapes are invalid.
  """

  def transpose_for_scores(input_tensor, batch_size, num_attention_heads,
                           seq_length, width):
    output_tensor = tf.reshape(
        input_tensor, [batch_size, seq_length, num_attention_heads, width])

    output_tensor = tf.transpose(output_tensor, [0, 2, 1, 3])
    return output_tensor

  from_shape = get_shape_list(from_tensor, expected_rank=[2, 3])
  to_shape = get_shape_list(to_tensor, expected_rank=[2, 3])

  if len(from_shape) != len(to_shape):
    raise ValueError(
        "The rank of `from_tensor` must match the rank of `to_tensor`.")

  if len(from_shape) == 3:
    batch_size = from_shape[0]
    from_seq_length = from_shape[1]
    to_seq_length = to_shape[1]
  elif len(from_shape) == 2:
    if (batch_size is None or from_seq_length is None or to_seq_length is None):
      raise ValueError(
          "When passing in rank 2 tensors to attention_layer, the values "
          "for `batch_size`, `from_seq_length`, and `to_seq_length` "
          "must all be specified.")

  # Scalar dimensions referenced here:
  #   B = batch size (number of sequences)
  #   F = `from_tensor` sequence length
  #   T = `to_tensor` sequence length
  #   N = `num_attention_heads`
  #   H = `size_per_head`

  from_tensor_2d = reshape_to_matrix(from_tensor) # e.g. [4*128, 768]
  to_tensor_2d = reshape_to_matrix(to_tensor)  # e.g. [4*128, 768]

  # `query_layer` = [B*F, N*H]  根据from_tensor计算Query
  query_layer = tf.layers.dense(
      from_tensor_2d,
      num_attention_heads * size_per_head,
      activation=query_act,
      name="query",
      kernel_initializer=create_initializer(initializer_range))

  # `key_layer` = [B*T, N*H]  根据to_tensor计算Key
  key_layer = tf.layers.dense(
      to_tensor_2d,
      num_attention_heads * size_per_head,
      activation=key_act,
      name="key",
      kernel_initializer=create_initializer(initializer_range))

  # `value_layer` = [B*T, N*H]  根据to_tensor计算Value
  value_layer = tf.layers.dense(
      to_tensor_2d,
      num_attention_heads * size_per_head,
      activation=value_act,
      name="value",
      kernel_initializer=create_initializer(initializer_range))

  # 把query_layer从 [B*F, N*H] 转为 [B, N, F, H]
  query_layer = transpose_for_scores(query_layer, batch_size,
                                     num_attention_heads, from_seq_length,
                                     size_per_head)

  # 把key_layer从[B*T, N*H] 转为 [B, N, T, H]
  key_layer = transpose_for_scores(key_layer, batch_size, num_attention_heads,
                                   to_seq_length, size_per_head)

  # 计算Query和Key的内积，得到attention_scores. 
  # Take the dot product between "query" and "key" to get the raw attention scores.
  # `attention_scores` = [B, N, F, T] 最后两维[128, 128]表示从from的128个时刻attend到to的128个时刻的score
  attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
  attention_scores = tf.multiply(attention_scores,
                                 1.0 / math.sqrt(float(size_per_head)))

  if attention_mask is not None:
    # `attention_mask` = [B, 1, F, T]
    attention_mask = tf.expand_dims(attention_mask, axis=[1])

    # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
    # masked positions, this operation will create a tensor which is 0.0 for
    # positions we want to attend and -10000.0 for masked positions.
    adder = (1.0 - tf.cast(attention_mask, tf.float32)) * -10000.0

    # Since we are adding it to the raw scores before the softmax, this is
    # effectively the same as removing these entirely.
    attention_scores += adder

  # Normalize the attention scores to probabilities.
  # `attention_probs` = [B, N, F, T]
  attention_probs = tf.nn.softmax(attention_scores)

  # 这里进行drop虽然有点奇怪，但Transformer原始论文就这么干的
  # This is actually dropping out entire tokens to attend to, which might
  # seem a bit unusual, but is taken from the original Transformer paper.
  attention_probs = dropout(attention_probs, attention_probs_dropout_prob)

  # `value_layer` = [B, T, N, H]
  value_layer = tf.reshape(
      value_layer,
      [batch_size, to_seq_length, num_attention_heads, size_per_head])

  # `value_layer` = [B, N, T, H]
  value_layer = tf.transpose(value_layer, [0, 2, 1, 3])

  # `context_layer` = [B, N, F, H]
  context_layer = tf.matmul(attention_probs, value_layer)

  # `context_layer` = [B, F, N, H]
  context_layer = tf.transpose(context_layer, [0, 2, 1, 3])

  if do_return_2d_tensor:
    # `context_layer` = [B*F, N*H]
    context_layer = tf.reshape(
        context_layer,
        [batch_size * from_seq_length, num_attention_heads * size_per_head])
  else:
    # `context_layer` = [B, F, N*H]
    context_layer = tf.reshape(
        context_layer,
        [batch_size, from_seq_length, num_attention_heads * size_per_head])

  return context_layer

该函数将from_tensor作为Query去attent to to_tensor(提供Key和Value)。也就是用from的Query去乘以to的Key，得到weight后，再对to的Value进行加权求和。总共有num_attention_heads个head，每个head对应一组Query、Key和Value，shape都是[batch_size, seq_length, size_per_head]。最后把所有head的输出拼接起来，得到shape为[batch_size, seq_length, size_per_head*num_head_attentions]的输出。

如果from_tensor=to_tensor，则为self-attention。
若do_return_2d_tensor=True，则返回[batch_size*seq_length, size_per_head*num_head_attentions]；否则返回[batch_size, seq_length, size_per_head*num_head_attentions]。

到此，解读完BERT模型的网络结构，下一篇我们将解读在多种下游任务上fine-tune BERT模型的代码。

BERT代码阅读-李理的博客
Attention is all your need
通俗易懂理解BN和LN归一化原理
BERT中，multi-head 768x64x12与直接使用768x768矩阵统一计算的区别？
深度学习中Attention与全连接层的区别何在？ - 知乎 (zhihu.com)