BERT源码解析（上）

解读源码前先大致了解下什么是Bert吧

Bert用了Transform的encoder侧网络，作为一个文本编码器，使用大规模数据进行预训练，预训练使用了两个loss，一个是LM Mask,遮蔽源端的一些字，通过上下文去预测这些字。还有一个是next sentence prediction，判断两个句子是否在文章中互为上下句，然后使用大规模的语料库去预训练。模型结构如下：

Bert base 具有12层上图所示的结构，每一层包含 multi-head Attention、Feed Forward两层子结构。

了解大致模型结构之后，开始快乐的源码解析之路～

一、模型结构（modeling.py）

这一部分是Bert里面的精髓，主要也是transformer encoder的实现

源码地址：https://github.com/google-research/bert/blob/master/modeling.py

上图是文件里一些函数和类，我会挑一些比较重要的函数进行说明

1、BertConfig

关于Bert模型的一些参数配置

class BertConfig(object):
  """Configuration for `BertModel`."""

  def __init__(self,
               vocab_size,
               hidden_size=768,
               num_hidden_layers=12,
               num_attention_heads=12,
               intermediate_size=3072,
               hidden_act="gelu",
               hidden_dropout_prob=0.1,
               attention_probs_dropout_prob=0.1,
               max_position_embeddings=512,
               type_vocab_size=16,
               initializer_range=0.02):
  
    self.vocab_size = vocab_size
    self.hidden_size = hidden_size
    self.num_hidden_layers = num_hidden_layers
    self.num_attention_heads = num_attention_heads
    self.hidden_act = hidden_act
    self.intermediate_size = intermediate_size
    self.hidden_dropout_prob = hidden_dropout_prob
    self.attention_probs_dropout_prob = attention_probs_dropout_prob
    self.max_position_embeddings = max_position_embeddings
    self.type_vocab_size = type_vocab_size
    self.initializer_range = initializer_range

    
  # 以下函数无非就是针对读取不同类型配置文件的适配  
  @classmethod
  def from_dict(cls, json_object):
    """Constructs a `BertConfig` from a Python dictionary of parameters."""
    config = BertConfig(vocab_size=None)
    for (key, value) in six.iteritems(json_object):
      config.__dict__[key] = value
    return config

  @classmethod
  def from_json_file(cls, json_file):
    """Constructs a `BertConfig` from a json file of parameters."""
    with tf.gfile.GFile(json_file, "r") as reader:
      text = reader.read()
    return cls.from_dict(json.loads(text))

  def to_dict(self):
    """Serializes this instance to a Python dictionary."""
    output = copy.deepcopy(self.__dict__)
    return output

  def to_json_string(self):
    """Serializes this instance to a JSON string."""
    return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"

【参数含义】

vocab_size: 输入词汇量大小

hidden_size: 隐藏层大小

num_hidden_layers: 就是上图中结构的个数（Transformer encoder中的隐藏层数）

num_attention_heads：定义multi-header attention 的head数 遵循 hidden_size % num_attention_heads == 0

intermediate_size: feed-forward layer中神经元的个数

hidden_act: 隐藏层的激活函数

hidden_dropout_prob: 隐藏层的神经元失活率

attention_probs_dropout_prob: 注意力部分的dropout

max_position_embeddings: 最大位置编码

type_vocab_size: 其实就是一个样本中输入的句子个数， 比如输入: [cls] 句A [sep] 句B [sep] 那么type_vocab_size=2

initializer_range: truncated_normal_initializer初始化方法的stdev

2、Embedding_lookup(BERT词向量token_embedding)

这个函数其实就是一个查表的功能，当给定一个one-hot的时候将one-hot与矩阵相乘，得到对应位置为1的embedding向量，或者给定词的token_id通过tf.gather()根据索引值查找embedding向量（tf.gather用法参考：https://blog.csdn.net/u012193416/article/details/86516009）

def embedding_lookup(input_ids,  # [batch_size, seq_length]
                     vocab_size,
                     embedding_size=128,
                     initializer_range=0.02,
                     word_embedding_name="word_embeddings",
                     use_one_hot_embeddings=False):
  """Looks up words embeddings for id tensor.
  Args:
    input_ids: int32 Tensor of shape [batch_size, seq_length] containing word
      ids.
    vocab_size: int. Size of the embedding vocabulary.
    embedding_size: int. Width of the word embeddings.
    initializer_range: float. Embedding initialization range.
    word_embedding_name: string. Name of the embedding table.
    use_one_hot_embeddings: bool. If True, use one-hot method for word
      embeddings. If False, use `tf.gather()`.
  Returns:
    float Tensor of shape [batch_size, seq_length, embedding_size].
  """
  # This function assumes that the input is of shape [batch_size, seq_length,
  # num_inputs].
  #
  # If the input is a 2D tensor of shape [batch_size, seq_length], we
  # reshape to [batch_size, seq_length, 1].
  if input_ids.shape.ndims == 2:
    input_ids = tf.expand_dims(input_ids, axis=[-1])
	
  # 初始化embedding向量
  embedding_table = tf.get_variable(
      name=word_embedding_name,
      shape=[vocab_size, embedding_size],
      initializer=create_initializer(initializer_range))
	# 转化为1维向量
  flat_input_ids = tf.reshape(input_ids, [-1])
  if use_one_hot_embeddings:
    one_hot_input_ids = tf.one_hot(flat_input_ids, depth=vocab_size)
    output = tf.matmul(one_hot_input_ids, embedding_table)
  else:
    output = tf.gather(embedding_table, flat_input_ids)
	# 以list的形式返回input_ids的形状，这个是源码自己定义的函数
  input_shape = get_shape_list(input_ids)
	# input_shape[0:-1] = [batch_size, sequence], input_shape[-1]= num_inputs 
  
  output = tf.reshape(output,
                      input_shape[0:-1] + [input_shape[-1] * embedding_size])
  return (output, embedding_table)

3、embedding_postprocessor（其实就是补全bert的输入）

因为BERT模型有三个输入分别是 token embedding ，segment embedding 以
及 position embedding 目前我们已经解决了token embedding，还差segment embedding 和position embedding。

稍微解释下三个embedding的含义

token embedding: 每个字或词对应embedding，可以随机初始化通过模型不断训练

Segment embedding: 如果输入为两个句子让模型知道哪个是句子1哪个是句子2

Position embedding: 输入的位置信息，Bert是随机初始化的，然后通过训练来学，transformer是由sin/cos函数生成的固定的值

def embedding_postprocessor(input_tensor,
                            use_token_type=False,
                            token_type_ids=None,
                            token_type_vocab_size=16,
                            token_type_embedding_name="token_type_embeddings",
                            use_position_embeddings=True,
                            position_embedding_name="position_embeddings",
                            initializer_range=0.02,
                            max_position_embeddings=512,
                            dropout_prob=0.1):
  """Performs various post-processing on a word embedding tensor.
  Args:
    input_tensor: float Tensor of shape [batch_size, seq_length,
      embedding_size].
    use_token_type: bool. Whether to add embeddings for `token_type_ids`.
    token_type_ids: (optional) int32 Tensor of shape [batch_size, seq_length].
      Must be specified if `use_token_type` is True.
    token_type_vocab_size: int. The vocabulary size of `token_type_ids`.
    token_type_embedding_name: string. The name of the embedding table variable
      for token type ids.
    use_position_embeddings: bool. Whether to add position embeddings for the
      position of each token in the sequence.
    position_embedding_name: string. The name of the embedding table variable
      for positional embeddings.
    initializer_range: float. Range of the weight initialization.
    max_position_embeddings: int. Maximum sequence length that might ever be
      used with this model. This can be longer than the sequence length of
      input_tensor, but cannot be shorter.
    dropout_prob: float. Dropout probability applied to the final output tensor.
  Returns:
    float tensor with same shape as `input_tensor`.
  Raises:
    ValueError: One of the tensor shapes or input values is invalid.
  """
  input_shape = get_shape_list(input_tensor, expected_rank=3)
  batch_size = input_shape[0]
  seq_length = input_shape[1]
  # 感觉命名为embedding_size会更好
  width = input_shape[2]

  output = input_tensor
	
  # segment embedding信息
  if use_token_type:
    if token_type_ids is None:
      raise ValueError("`token_type_ids` must be specified if"
                       "`use_token_type` is True.")
    token_type_table = tf.get_variable(
        name=token_type_embedding_name,
        shape=[token_type_vocab_size, width],
        initializer=create_initializer(initializer_range))
    # This vocab will be small so we always do one-hot here, since it is always
    # faster for a small vocabulary. token_type_vocab_size一般为2如果输入两个句子的话
    flat_token_type_ids = tf.reshape(token_type_ids, [-1])
    one_hot_ids = tf.one_hot(flat_token_type_ids, depth=token_type_vocab_size)
    token_type_embeddings = tf.matmul(one_hot_ids, token_type_table)
    token_type_embeddings = tf.reshape(token_type_embeddings,
                                       [batch_size, seq_length, width])
    output += token_type_embeddings
	
  # position embedding信息
  if use_position_embeddings:
    assert_op = tf.assert_less_equal(seq_length, max_position_embeddings)
    with tf.control_dependencies([assert_op]):
      full_position_embeddings = tf.get_variable(
          name=position_embedding_name,
          shape=[max_position_embeddings, width],
          initializer=create_initializer(initializer_range))
       # 这里position embedding是可学习的参数，[max_position_embeddings, width
       #  但是通常实际输入序列没有达到max_position_embeddings 
       #  所以为了提高训练速度，使用t f . s l i c e 取 出句 子长度的e m b e d d i n g
       #  [0,0] 起始位置 [seq_length, -1]终止位置
      position_embeddings = tf.slice(full_position_embeddings, [0, 0],
                                     [seq_length, -1])
      
      num_dims = len(output.shape.as_list())
			
      # word embedding之后的tensor是[batch_size, seq_length, width]  

      #   因为位置编码是与输入内容无关，它的shape总 是[seq_length, width]  
      #   我们无法把位置position embedding加到word embedding上 
      #   因此我们需要扩展位置编码为[1, seq_length,  width]  
      #   然后 就能通过b r o a d c a s t i n g 加上去了。
      position_broadcast_shape = []
      for _ in range(num_dims - 2):
        position_broadcast_shape.append(1)
      position_broadcast_shape.extend([seq_length, width])
      position_embeddings = tf.reshape(position_embeddings,
                                       position_broadcast_shape)
      output += position_embeddings

  output = layer_norm_and_dropout(output, dropout_prob)
  return output

4、attention_mask

一般在一个batch里我们会把每个样本输入的长度固定一个值，那么有些句子长度小于这个固定值那么可以通过padding来补长（大于固定值进行截断），当计算self-attention的时候我们希望去除padding部分和其它词的关联，那么可以通过mask掉这部分计算出的相似度权重。

def create_attention_mask_from_input_mask(from_tensor, to_mask):
  """Create 3D attention mask from a 2D tensor mask.
  Args:
    from_tensor: 2D or 3D Tensor of shape [batch_size, from_seq_length, ...].
    to_mask: int32 Tensor of shape [batch_size, to_seq_length].
  Returns:
    float Tensor of shape [batch_size, from_seq_length, to_seq_length].
  """
  from_shape = get_shape_list(from_tensor, expected_rank=[2, 3])
  batch_size = from_shape[0]
  from_seq_length = from_shape[1]

  to_shape = get_shape_list(to_mask, expected_rank=2)
  to_seq_length = to_shape[1]

  to_mask = tf.cast(
      tf.reshape(to_mask, [batch_size, 1, to_seq_length]), tf.float32)

  # We don't assume that `from_tensor` is a mask (although it could be). We
  # don't actually care if we attend *from* padding tokens (only *to* padding)
  # tokens so we create a tensor of all ones.
  #
  # `broadcast_ones` = [batch_size, from_seq_length, 1]
  broadcast_ones = tf.ones(
      shape=[batch_size, from_seq_length, 1], dtype=tf.float32)

  # Here we broadcast along two dimensions to create the mask.
  mask = broadcast_ones * to_mask # [batch_size, from_seq_length, to_seq_length]

  return mask

5、Mutil-Head Attention

这部分主要是实现《Attention Is All You Need》论文里的self-attention想要了解原理的看官老爷请移步[2]

def attention_layer(from_tensor, # [batch_size, from_seq_length, from_width]
                    to_tensor, # [batch_size, to_seq_length, to_width] to_seq_length==from_seq_lengt
                    attention_mask=None, # [batch_size,from_seq_length,to_seq_length]
                    num_attention_heads=1, # attention head个数
                    size_per_head=512, # 每一个attention head 的维度
                    query_act=None, # query transform的激活函数
                    key_act=None, # key transform的激活函数
                    value_act=None, # value transform的激活函数
                    attention_probs_dropout_prob=0.0, # attention dropout失活率
                    initializer_range=0.02, # embedding随机初始化的范围
                    do_return_2d_tensor=False, # 返回维度是否为2D
                    batch_size=None, 
                    from_seq_length=None,
                    to_seq_length=None):
	# 输入的维度是 [batch_size, seq_length, embedding_size] 转换为[batch_size, num_heads, seq_length, size_per_head]通过拆分矩阵来计算多头注意力
  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`

  #   把from_tensor和to_tensor 压缩成2D 张量，源码中自定义的函数
  from_tensor_2d = reshape_to_matrix(from_tensor)
  to_tensor_2d = reshape_to_matrix(to_tensor)

  # `query_layer` = [B*F, N*H]
  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]
  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]
  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))

  #  query_layer转成多头 ：[B*F, N*H] ==> [B, F, N, H ] ==> [B, N, F,  H]  
  query_layer = transpose_for_scores(query_layer, batch_size,
                                     num_attention_heads, from_seq_length,
                                     size_per_head)

  # key_layer转成多头 ：[B*T, N*H] ==> [B, T, N, H ] ==> [B, N, T, H]  
  key_layer = transpose_for_scores(key_layer, batch_size, num_attention_heads,
                                   to_seq_length, size_per_head)

  # Take the dot product between "query" and "key" to get the raw
  # attention scores.
  # `attention_scores` = [B, N, F, T]
  attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
  # 感觉这里很巧妙把除法转换为乘法，加快运算
  attention_scores = tf.multiply(attention_scores,
                                 1.0 / math.sqrt(float(size_per_head)))
	
  # 将padding部分计算的权重的分设置成非常小的一个负数，因为当输入是一个非常小的负数是经过softmax出来就趋近于0了，相当于mask为0的地方不计算权重，很巧妙
  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)

  # 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

上述代码有点多，总结一下self-attention流程

1）对输入的tensor进行形状的校验

2）from_tensor作为query，to_tensor作为key, value。经过一层全连接后得到query_layer、key_layer、value_layer

3） 将上述query_layer, key_layer经过transpose_for_scores转换为多头

4）根据公式计算query、key之间的点积、缩放、送入softmax得到attention_probs，然后与value加权计算:

​ $(Q,K,V)=softmax(\frac{Q{K}^T}{\sqrt{d_k}})V$

6、Transformer_model

这部分无非就是将attention_layer + layer_norm + feed forward + layer_norm按顺序执行，得到output

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))

  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))

  # We keep the representation as a 2D tensor to avoid re-shaping it back and
  # forth from a 3D tensor to a 2D tensor. Re-shapes are normally free on
  # the GPU/CPU but may not be free on the TPU, so we want to minimize them to
  # help the optimizer.
  prev_output = reshape_to_matrix(input_tensor)

  all_layer_outputs = []
  for layer_idx in range(num_hidden_layers):
    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)

        # Run a linear projection of `hidden_size` then add a residual
        # with `layer_input`.
        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)

      # 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

7、Bert函数入口（init）

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.
        (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))

看懂了上面的代码理解这部分应该是很轻松的

看下源码给的一个例子吧

# Already been converted into WordPiece token ids
  input_ids = tf.constant([[31, 51, 99], [15, 5, 0]])
  input_mask = tf.constant([[1, 1, 1], [1, 1, 0]])
  token_type_ids = tf.constant([[0, 0, 1], [0, 2, 0]])
  config = modeling.BertConfig(vocab_size=32000, hidden_size=512,
    num_hidden_layers=8, num_attention_heads=6, intermediate_size=1024)
  model = modeling.BertModel(config=config, is_training=True,
    input_ids=input_ids, input_mask=input_mask, token_type_ids=token_type_ids)
  label_embeddings = tf.get_variable(...)
  pooled_output = model.get_pooled_output()
  logits = tf.matmul(pooled_output, label_embeddings)

二、数据输入

2.1 分词器（tokenization.py）

先看下这个文件里面有哪些函数和类

tokenization.py主要是对原语句输入进行一些处理，主要有BasicTokenizer和WordpieceTokenizer

2.1.1BasicTokenizer

class BasicTokenizer(object):
  """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""

  def __init__(self, do_lower_case=True):
    """Constructs a BasicTokenizer.
    Args:
      do_lower_case: Whether to lower case the input.
    """
    self.do_lower_case = do_lower_case

  def tokenize(self, text):
    """Tokenizes a piece of text."""
    text = convert_to_unicode(text)
    text = self._clean_text(text)

    # This was added on November 1st, 2018 for the multilingual and Chinese
    # models. This is also applied to the English models now, but it doesn't
    # matter since the English models were not trained on any Chinese data
    # and generally don't have any Chinese data in them (there are Chinese
    # characters in the vocabulary because Wikipedia does have some Chinese
    # words in the English Wikipedia.).
    # 增加中文支持
    text = self._tokenize_chinese_chars(text)

    orig_tokens = whitespace_tokenize(text)
    split_tokens = []
    for token in orig_tokens:
      if self.do_lower_case:
        token = token.lower()
        token = self._run_strip_accents(token)
      split_tokens.extend(self._run_split_on_punc(token))

    output_tokens = whitespace_tokenize(" ".join(split_tokens))
    return output_tokens

  def _run_strip_accents(self, text):
    """Strips accents from a piece of text."""
    text = unicodedata.normalize("NFD", text)
    output = []
    for char in text:
      cat = unicodedata.category(char)
      if cat == "Mn":
        continue
      output.append(char)
    return "".join(output)

  def _run_split_on_punc(self, text):
    """Splits punctuation on a piece of text."""
    # 用标点切分，返回list
    chars = list(text)
    i = 0
    start_new_word = True
    output = []
    while i < len(chars):
      char = chars[i]
      if _is_punctuation(char):
        output.append([char])
        start_new_word = True
      else:
        if start_new_word:
          output.append([])
        start_new_word = False
        output[-1].append(char)
      i += 1

    return ["".join(x) for x in output]

  def _tokenize_chinese_chars(self, text):
    """Adds whitespace around any CJK character."""
    # 按字切分中文句子，就是在字两侧加入空格
    output = []
    for char in text:
      cp = ord(char)
      if self._is_chinese_char(cp):
        output.append(" ")
        output.append(char)
        output.append(" ")
      else:
        output.append(char)
    return "".join(output)

  def _is_chinese_char(self, cp):
    # 判断是否为汉字
    """Checks whether CP is the codepoint of a CJK character."""
    # This defines a "chinese character" as anything in the CJK Unicode block:
    #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
    #
    # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
    # despite its name. The modern Korean Hangul alphabet is a different block,
    # as is Japanese Hiragana and Katakana. Those alphabets are used to write
    # space-separated words, so they are not treated specially and handled
    # like the all of the other languages.
    if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
        (cp >= 0x3400 and cp <= 0x4DBF) or  #
        (cp >= 0x20000 and cp <= 0x2A6DF) or  #
        (cp >= 0x2A700 and cp <= 0x2B73F) or  #
        (cp >= 0x2B740 and cp <= 0x2B81F) or  #
        (cp >= 0x2B820 and cp <= 0x2CEAF) or
        (cp >= 0xF900 and cp <= 0xFAFF) or  #
        (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
      return True

    return False

  def _clean_text(self, text):
    # 去除没有意义的字符以及空格
    """Performs invalid character removal and whitespace cleanup on text."""
    output = []
    for char in text:
      cp = ord(char)
      if cp == 0 or cp == 0xfffd or _is_control(char):
        continue
      if _is_whitespace(char):
        output.append(" ")
      else:
        output.append(char)
    return "".join(output)

2.1.2 WordpieceTokenizer

WordpieceTokenizer 是在WordpieceTokenizer进一步对句子进行更加细致的粒度划分，比如英文里的 taked、taking等，可能会被划分为[tak, ##ed, ##ing] 这样在一定程度上缓解未登录词对模型效果的影响，但是对中文没有影响，因为在前面BasicTokenizer里面已经切分成以字为单位的了。

class WordpieceTokenizer(object):
  """Runs WordPiece tokenziation."""

  def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=200):
    self.vocab = vocab
    self.unk_token = unk_token
    self.max_input_chars_per_word = max_input_chars_per_word

  def tokenize(self, text):
    """Tokenizes a piece of text into its word pieces.
    This uses a greedy longest-match-first algorithm to perform tokenization
    using the given vocabulary.
    For example:
      input = "unaffable"
      output = ["un", "##aff", "##able"]
    Args:
      text: A single token or whitespace separated tokens. This should have
        already been passed through `BasicTokenizer.
    Returns:
      A list of wordpiece tokens.
    """

    text = convert_to_unicode(text)

    output_tokens = []
    for token in whitespace_tokenize(text):
      chars = list(token)
      if len(chars) > self.max_input_chars_per_word:
        output_tokens.append(self.unk_token)
        continue

      is_bad = False
      start = 0
      sub_tokens = []
      while start < len(chars):
        end = len(chars)
        cur_substr = None
        while start < end:
          substr = "".join(chars[start:end])
          if start > 0:
            substr = "##" + substr
          if substr in self.vocab:
            cur_substr = substr
            break
          end -= 1
        if cur_substr is None:
          is_bad = True
          break
        sub_tokens.append(cur_substr)
        start = end

      if is_bad:
        output_tokens.append(self.unk_token)
      else:
        output_tokens.extend(sub_tokens)
    return output_tokens


def _is_whitespace(char):
  """Checks whether `chars` is a whitespace character."""
  # \t, \n, and \r are technically contorl characters but we treat them
  # as whitespace since they are generally considered as such.
  if char == " " or char == "\t" or char == "\n" or char == "\r":
    return True
  cat = unicodedata.category(char)
  if cat == "Zs":
    return True
  return False


def _is_control(char):
  """Checks whether `chars` is a control character."""
  # These are technically control characters but we count them as whitespace
  # characters.
  if char == "\t" or char == "\n" or char == "\r":
    return False
  cat = unicodedata.category(char)
  if cat in ("Cc", "Cf"):
    return True
  return False


def _is_punctuation(char):
  """Checks whether `chars` is a punctuation character."""
  cp = ord(char)
  # We treat all non-letter/number ASCII as punctuation.
  # Characters such as "^", "$", and "`" are not in the Unicode
  # Punctuation class but we treat them as punctuation anyways, for
  # consistency.
  if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or
      (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
    return True
  cat = unicodedata.category(char)
  if cat.startswith("P"):
    return True
  return False

解释下源码中的例子

比如假设输入是”unaffable”。我们跳到while循环部分，这是start=0，end=len(chars)=9，也就是先看看unaffable在不在词典里，如果在，那么直接作为一个WordPiece，如果不再，那么end-=1，也就是看unaffabl在不在词典里，最终发现”un”在词典里，把un加到结果里。接着start=2，看affable在不在，不在再看affabl，…，最后发现 ##aff 在词典里。注意：##表示这个词是接着前面的，这样使得WordPiece切分是可逆的——我们可以恢复出“真正”的词。

2.1.3 FullTokenizer

上述两种分词方式的接口

class FullTokenizer(object):
  """Runs end-to-end tokenziation."""

  def __init__(self, vocab_file, do_lower_case=True):
    self.vocab = load_vocab(vocab_file)
    self.inv_vocab = {v: k for k, v in self.vocab.items()}
    self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
    self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)

  def tokenize(self, text):
    split_tokens = []
    for token in self.basic_tokenizer.tokenize(text):
      for sub_token in self.wordpiece_tokenizer.tokenize(token):
        split_tokens.append(sub_token)

    return split_tokens

  def convert_tokens_to_ids(self, tokens):
    return convert_by_vocab(self.vocab, tokens)

  def convert_ids_to_tokens(self, ids):
    return convert_by_vocab(self.inv_vocab, ids)

2.2 数据的生成（create_pretraining_data.py ）

先看main函数再依次细分

def main(_):
  tf.logging.set_verbosity(tf.logging.INFO)
	# 构造tokenizer对输入语料库进行分词处理
  tokenizer = tokenization.FullTokenizer(
      vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case)

  input_files = []
  for input_pattern in FLAGS.input_file.split(","):
    input_files.extend(tf.gfile.Glob(input_pattern))

  tf.logging.info("*** Reading from input files ***")
  for input_file in input_files:
    tf.logging.info("  %s", input_file)

  rng = random.Random(FLAGS.random_seed)
  # 构造训练训练instances
  instances = create_training_instances(
      input_files, tokenizer, FLAGS.max_seq_length, FLAGS.dupe_factor,
      FLAGS.short_seq_prob, FLAGS.masked_lm_prob, FLAGS.max_predictions_per_seq,
      rng)

  output_files = FLAGS.output_file.split(",")
  tf.logging.info("*** Writing to output files ***")
  for output_file in output_files:
    tf.logging.info("  %s", output_file)
	# TFRecord格式保存数据
  write_instance_to_example_files(instances, tokenizer, FLAGS.max_seq_length,
                                  FLAGS.max_predictions_per_seq, output_files)

2.2.1 create_training_instances(构造训练数据)

def create_training_instances(input_files, tokenizer, max_seq_length,
                              dupe_factor, short_seq_prob, masked_lm_prob,
                              max_predictions_per_seq, rng):
  """Create `TrainingInstance`s from raw text."""
  #   all_documents 是list的list ，第一层list 表示document ， 
  #   第二层list表示document里的多个句子。
  all_documents = [[]]
  

  # 输入文件格式：
  # (1) 每行一个句子。理想情况下，这些应该是实际的句子，而不是
  # 整个段落或任意范围的文本。 （因为我们使用“下一句预测”任务的句子边界）。
  # (2) 文档之间的空行。需要文档边界，所以“下一句预测”任务不跨越文档。
  for input_file in input_files:
    with tf.gfile.GFile(input_file, "r") as reader:
      while True:
        line = tokenization.convert_to_unicode(reader.readline())
        if not line:
          break
        line = line.strip()

        # 空行表示分割文档
        if not line:
          all_documents.append([])
        tokens = tokenizer.tokenize(line)
        if tokens:
          all_documents[-1].append(tokens)

  # Remove empty documents
  all_documents = [x for x in all_documents if x]
  rng.shuffle(all_documents)

  vocab_words = list(tokenizer.vocab.keys())
  instances = []
  # 重复dupe_factor次，一个句子可以mask不同的位置
  for _ in range(dupe_factor):
    for document_index in range(len(all_documents)):
      instances.extend(
          create_instances_from_document(
              all_documents, document_index, max_seq_length, short_seq_prob,
              masked_lm_prob, max_predictions_per_seq, vocab_words, rng))

  rng.shuffle(instances)
  return instances

# 来实现从一个文档中抽取多个训练样本
def create_instances_from_document(
    all_documents, document_index, max_seq_length, short_seq_prob,
    masked_lm_prob, max_predictions_per_seq, vocab_words, rng):
  """Creates `TrainingInstance`s for a single document."""
  document = all_documents[document_index]

  # Account for [CLS], [SEP], [SEP]
  max_num_tokens = max_seq_length - 3

  # We *usually* want to fill up the entire sequence since we are padding
  # to `max_seq_length` anyways, so short sequences are generally wasted
  # computation. However, we *sometimes*
  # (i.e., short_seq_prob == 0.1 == 10% of the time) want to use shorter
  # sequences to minimize the mismatch between pre-training and fine-tuning.
  # The `target_seq_length` is just a rough target however, whereas
  # `max_seq_length` is a hard limit.
  
  target_seq_length = max_num_tokens
  #   以short_seq_prob 的概率随机生成（2 ~ max_num_tokens）的长度
  if rng.random() < short_seq_prob:
    target_seq_length = rng.randint(2, max_num_tokens)

  # We DON'T just concatenate all of the tokens from a document into a long
  # sequence and choose an arbitrary split point because this would make the
  # next sentence prediction task too easy. Instead, we split the input into
  # segments "A" and "B" based on the actual "sentences" provided by the user
  # input.
  instances = []
  current_chunk = []
  current_length = 0
  i = 0
  while i < len(document):
    segment = document[i]
    current_chunk.append(segment)
    current_length += len(segment)
    #   将句子依次加入current_chunk 中，直到加完或者达到限制的最大长度
    if i == len(document) - 1 or current_length >= target_seq_length:
      if current_chunk:
        # ` a_end ` 是第一个句子A结束的下标
        a_end = 1
        if len(current_chunk) >= 2:
          a_end = rng.randint(1, len(current_chunk) - 1)

        tokens_a = []
        for j in range(a_end):
          tokens_a.extend(current_chunk[j])

        tokens_b = []
        # Random next
        # 随机构建下一句
        is_random_next = False
        if len(current_chunk) == 1 or rng.random() < 0.5:
          is_random_next = True
          target_b_length = target_seq_length - len(tokens_a)

            # 随机的挑选另外一篇文档的随机开始的句子 
            # 但是理论上有可 能随机到的文档就是当前文档，因此需要一个w h i l e 循环 
            # 这里只 w h i l e 循环1 0 次，理论上还是有重复 的可 能性 ，但是我们忽 略
          for _ in range(10):
            random_document_index = rng.randint(0, len(all_documents) - 1)
            if random_document_index != document_index:
              break

          random_document = all_documents[random_document_index]
          random_start = rng.randint(0, len(random_document) - 1)
          for j in range(random_start, len(random_document)):
            tokens_b.extend(random_document[j])
            if len(tokens_b) >= target_b_length:
              break
          #   对于上述构建的随机下一句 ，我们并没有真正地使用它们 
          #   所以为了避免数据浪费，我们将其“ 放回”
          num_unused_segments = len(current_chunk) - a_end
          i -= num_unused_segments
        # 构建真实的next
        else:
          is_random_next = False
          for j in range(a_end, len(current_chunk)):
            tokens_b.extend(current_chunk[j])
         # 随机去掉一些
        truncate_seq_pair(tokens_a, tokens_b, max_num_tokens, rng)

        assert len(tokens_a) >= 1
        assert len(tokens_b) >= 1

        tokens = []
        segment_ids = []
        # 处理句子A
        tokens.append("[CLS]")
        segment_ids.append(0)
        for token in tokens_a:
          tokens.append(token)
          segment_ids.append(0)
				# 句子A结束，加上[SEP]
        tokens.append("[SEP]")
        segment_ids.append(0)
				# 处理句子B
        for token in tokens_b:
          tokens.append(token)
          segment_ids.append(1)
       # 句子A结束，加上[SEP]
        tokens.append("[SEP]")
        segment_ids.append(1)
				# #   调用create_masked_lm_predictions来随机对某些T o k e n 进行m a s k 
        (tokens, masked_lm_positions,
         masked_lm_labels) = create_masked_lm_predictions(
             tokens, masked_lm_prob, max_predictions_per_seq, vocab_words, rng)
        instance = TrainingInstance(
            tokens=tokens,
            segment_ids=segment_ids,
            is_random_next=is_random_next,
            masked_lm_positions=masked_lm_positions,
            masked_lm_labels=masked_lm_labels)
        instances.append(instance)
      current_chunk = []
      current_length = 0
    i += 1

  return instances

描述下上诉代码过程

1）算法首先会维护一个chunk，不断加入document中的元素，也就是句子（segment），
直到加载完或者chunk中token数大于等于最大限制，这样做的目的是使得padding的尽
量少，训练效率更高。

2）现在chunk建立完毕之后，假设包括了前三个句子，算法会随机选择一个切分点，比如
2。接下来构建 predict next 判断： (1) 如果是正样本，前两个句子当成是句子A，
后一个句子当成是句子B； (2) 如果是负样本，前两个句子当成是句子A，无关的句子从
其他文档中随机抽取。

3）得到句子A和句子B之后，对其填充tokens和segment_ids，这里会加入特殊的[CLS]和
[SEP]标记。

4）对句子进行mask操作。

2.2.2create_masked_lm_predictions（随机mask）

对Tokens进行随机mask是BERT的一大创新点。使用mask的原因是为了模型训练的过程中可以看到下文”。于是，文章中选取的策略是对输入序列中15%的词使用[MASK]标记掩盖掉，然后通过上下文去预测这些被mask的token。但是为了防止模型过拟
合地学习到【MASK】这个标记，对15%mask掉的词进一步优化：

以80%的概率用[MASK]替换：
hello world， i am bert. ----> hello world, i am [MASK].
以10%的概率随机替换：
hello world， i am bert. ----> hello world, i am Elmo.
以10%的概率不进行替换：
hello world， i am bert. ----> hello world, i am bert.

def create_masked_lm_predictions(tokens, masked_lm_prob,
                                 max_predictions_per_seq, vocab_words, rng):
  """Creates the predictions for the masked LM objective."""

  cand_indexes = []
  for (i, token) in enumerate(tokens):
    if token == "[CLS]" or token == "[SEP]":
      continue
    # Whole Word Masking means that if we mask all of the wordpieces
    # corresponding to an original word. When a word has been split into
    # WordPieces, the first token does not have any marker and any subsequence
    # tokens are prefixed with ##. So whenever we see the ## token, we
    # append it to the previous set of word indexes.
    #
    # Note that Whole Word Masking does *not* change the training code
    # at all -- we still predict each WordPiece independently, softmaxed
    # over the entire vocabulary.
    if (FLAGS.do_whole_word_mask and len(cand_indexes) >= 1 and
        token.startswith("##")):
      cand_indexes[-1].append(i)
    else:
      cand_indexes.append([i])

  rng.shuffle(cand_indexes)

  output_tokens = list(tokens)

  num_to_predict = min(max_predictions_per_seq,
                       max(1, int(round(len(tokens) * masked_lm_prob))))

  masked_lms = []
  covered_indexes = set()
  for index_set in cand_indexes:
    if len(masked_lms) >= num_to_predict:
      break
    # If adding a whole-word mask would exceed the maximum number of
    # predictions, then just skip this candidate.
    if len(masked_lms) + len(index_set) > num_to_predict:
      continue
    is_any_index_covered = False
    for index in index_set:
      if index in covered_indexes:
        is_any_index_covered = True
        break
    if is_any_index_covered:
      continue
    for index in index_set:
      covered_indexes.add(index)

      masked_token = None
      # 80% of the time, replace with [MASK]
      if rng.random() < 0.8:
        masked_token = "[MASK]"
      else:
        # 10% of the time, keep original
        if rng.random() < 0.5:
          masked_token = tokens[index]
        # 10% of the time, replace with random word
        else:
          masked_token = vocab_words[rng.randint(0, len(vocab_words) - 1)]

      output_tokens[index] = masked_token

      masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
  assert len(masked_lms) <= num_to_predict
  masked_lms = sorted(masked_lms, key=lambda x: x.index)

  masked_lm_positions = []
  masked_lm_labels = []
  for p in masked_lms:
    masked_lm_positions.append(p.index)
    masked_lm_labels.append(p.label)

  return (output_tokens, masked_lm_positions, masked_lm_labels)

2.2.3write_instance_to_example_files（保存tfrecord数据）

就上述处理的数据保存为tfrecord格式

def write_instance_to_example_files(instances, tokenizer, max_seq_length,
                                    max_predictions_per_seq, output_files):
  """Create TF example files from `TrainingInstance`s."""
  writers = []
  for output_file in output_files:
    writers.append(tf.python_io.TFRecordWriter(output_file))

  writer_index = 0

  total_written = 0
  for (inst_index, instance) in enumerate(instances):
    input_ids = tokenizer.convert_tokens_to_ids(instance.tokens)
    input_mask = [1] * len(input_ids)
    segment_ids = list(instance.segment_ids)
    assert len(input_ids) <= max_seq_length

    while len(input_ids) < max_seq_length:
      input_ids.append(0)
      input_mask.append(0)
      segment_ids.append(0)

    assert len(input_ids) == max_seq_length
    assert len(input_mask) == max_seq_length
    assert len(segment_ids) == max_seq_length

    masked_lm_positions = list(instance.masked_lm_positions)
    masked_lm_ids = tokenizer.convert_tokens_to_ids(instance.masked_lm_labels)
    masked_lm_weights = [1.0] * len(masked_lm_ids)

    while len(masked_lm_positions) < max_predictions_per_seq:
      masked_lm_positions.append(0)
      masked_lm_ids.append(0)
      masked_lm_weights.append(0.0)

    next_sentence_label = 1 if instance.is_random_next else 0

    features = collections.OrderedDict()
    features["input_ids"] = create_int_feature(input_ids)
    features["input_mask"] = create_int_feature(input_mask)
    features["segment_ids"] = create_int_feature(segment_ids)
    features["masked_lm_positions"] = create_int_feature(masked_lm_positions)
    features["masked_lm_ids"] = create_int_feature(masked_lm_ids)
    features["masked_lm_weights"] = create_float_feature(masked_lm_weights)
    features["next_sentence_labels"] = create_int_feature([next_sentence_label])

    tf_example = tf.train.Example(features=tf.train.Features(feature=features))

    writers[writer_index].write(tf_example.SerializeToString())
    writer_index = (writer_index + 1) % len(writers)

    total_written += 1

    if inst_index < 20:
      tf.logging.info("*** Example ***")
      tf.logging.info("tokens: %s" % " ".join(
          [tokenization.printable_text(x) for x in instance.tokens]))

      for feature_name in features.keys():
        feature = features[feature_name]
        values = []
        if feature.int64_list.value:
          values = feature.int64_list.value
        elif feature.float_list.value:
          values = feature.float_list.value
        tf.logging.info(
            "%s: %s" % (feature_name, " ".join([str(x) for x in values])))

  for writer in writers:
    writer.close()

  tf.logging.info("Wrote %d total instances", total_written)

还有Bert的预训练、Finetune由于篇幅过长打算再开一篇文章！！！！！！！！！！

参考资料：

[1] https://github.com/google-research/bert
[2] https://arxiv.org/abs/1706.03762
[3] NLP大杀器BERT模型解读:https://blog.csdn.net/Kaiyuan_sjtu/article/detail
s/83991186