谷歌BERT预训练源码解析（一）：训练数据生成

目录
预训练源码结构简介
输入输出
源码解析
参数
主函数
创建训练实例
下一句预测&实例生成
随机遮蔽
输出
结果一览
预训练源码结构简介
关于BERT，简单来说，它是一个基于Transformer架构，结合遮蔽词预测和上下句识别的预训练NLP模型。至于效果：在11种不同NLP测试中创出最佳成绩
关于介绍BERT的文章我看了一些，个人感觉介绍的最全面的是机器之心
再放上谷歌官方源码链接：BERT官方源码
在看本博客之前，读者先要了解：
1.Transformer架构
2.BERT模型的创新之处
3.python语言及tensorflow框架
我会在代码中直接指出对应的原理，如果没有了解架构直接刚代码可能会有些吃力
BERT的预训练主要分为三个部分：
1.预训练数据的预处理(create_pretraining_data.py)
2.核心模型的构建(modeling.py)
3.训练过程(run_pretraining.py)
我将分三次分别介绍这三个部分的源码，这次先介绍训练数据的训练数据生成脚本即create_pretraining_data.py。

输入输出
关于输入和输出，我们可以直接从官方提供的训练命令行中窥之一二
 

python create_pretraining_data.py \
  --input_file=./sample_text.txt \
  --output_file=/tmp/tf_examples.tfrecord \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --do_lower_case=True \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --masked_lm_prob=0.15 \
  --random_seed=12345 \
  --dupe_factor=5

可以看到 这里谷歌为我们提供了一个小的训练样本sample_text.txt（输入），将这个训练样本进行处理后输出到**tf_examples.tfrecord（输出）**这个文件。在sample_text.txt中，空行前后是不同的文章，每个文章中的每句话都占一行（也就是说每篇文章的上下两行是一篇文章的上下句）。vocab_file是官方模型中提供的词汇表。
sample_text.txt
 

源码解析
参数
input_file：指定输入文档路径
output_file：指定输出路径
vocab_file：指定词典路径（谷歌已在预训练模型中提供）
do_lower_case：为True则忽略大小写
max_seq_length：每一条训练数据（两句话）相加后的最大长度限制
max_predictions_per_seq：每一条训练数据mask的最大数量
random_seed：一个随机种子
dupe_factor：对文档多次重复随机产生训练集，随机的次数
masked_lm_prob：一条训练数据产生mask的概率，即每条训练数据随机产生max_predictions_per_seq×masked_lm_prob数量的mask
short_seq_prob：为了缩小预训练和微调过程的差距，以此概率产生小于max_seq_length的训练数据
 

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import collections
import random

import tokenization
import tensorflow as tf

flags = tf.flags

FLAGS = flags.FLAGS

flags.DEFINE_string("input_file", None,
                    "Input raw text file (or comma-separated list of files).")

flags.DEFINE_string(
    "output_file", None,
    "Output TF example file (or comma-separated list of files).")

flags.DEFINE_string("vocab_file", None,
                    "The vocabulary file that the BERT model was trained on.")

flags.DEFINE_bool(
    "do_lower_case", True,
    "Whether to lower case the input text. Should be True for uncased "
    "models and False for cased models.")

flags.DEFINE_integer("max_seq_length", 128, "Maximum sequence length.")

flags.DEFINE_integer("max_predictions_per_seq", 20,
                     "Maximum number of masked LM predictions per sequence.")

flags.DEFINE_integer("random_seed", 12345, "Random seed for data generation.")

flags.DEFINE_integer(
    "dupe_factor", 10,
    "Number of times to duplicate the input data (with different masks).")

flags.DEFINE_float("masked_lm_prob", 0.15, "Masked LM probability.")

flags.DEFINE_float(
    "short_seq_prob", 0.1,
    "Probability of creating sequences which are shorter than the "
    "maximum length.")

主函数
首先获取输入文本列表，对输入文本创建训练实例，再进行输出
简要介绍一下FullTokenizer这个类，它以vocab_file为词典，将词转化为该词对应的id，对于某些特殊词，如johanson，会先将johanson按照最大长度拆分，再看拆分的部分是否在vocab_file里。vocab_file里有没有"johanson"这个词，但有"johan"和"##son"这两个词，所以将"johanson"这个词拆分成两个词（##表示非开头匹配）
 

def main(_):
  tf.logging.set_verbosity(tf.logging.INFO)

  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))        #获得输入文件列表

  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)
  instances = create_training_instances(      #创建训练实例
      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)

  write_instance_to_example_files(instances, tokenizer, FLAGS.max_seq_length,    #输出
                                  FLAGS.max_predictions_per_seq, output_files)

创建训练实例

这部分先将文章和每篇文章的每个句子加到二维列表，再将列表传入create_instances_from_document生成训练实例.
返回值：instances 一个列表 里面包含每个样例的TrainingInstance类

 

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 = [[]]

  # Input file format:
  # (1) One sentence per line. These should ideally be actual sentences, not
  # entire paragraphs or arbitrary spans of text. (Because we use the
  # sentence boundaries for the "next sentence prediction" task).
  # (2) Blank lines between documents. Document boundaries are needed so
  # that the "next sentence prediction" task doesn't span between documents.
  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()

        # Empty lines are used as document delimiters
        if not line:
          all_documents.append([])
        tokens = tokenizer.tokenize(line)
        if tokens:
          all_documents[-1].append(tokens)    #二维列表  [文章，句子]

  # Remove empty documents
  all_documents = [x for x in all_documents if x]   #删除空列表
  rng.shuffle(all_documents)   #随机排序

  vocab_words = list(tokenizer.vocab.keys())
  instances = []
  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

 

下一句预测&实例生成
这部分是生成训练数据的具体过程，对每条数据生成TrainingInstance。这里的每条数据其实包含两个句子的信息。TrainingInstance包括tokens,segement_ids,is_random_next,masked_lm_positions,masked_lm_labels。下面给出这些属性的含义
tokens：词
segement_id：句子编码 第一句为0 第二句为1
is_random_next：第二句是随机查找，还是为第一句的下文
masked_lm_positions：tokens中被mask的位置
masked_lm_labels：tokens中被mask的原来的词
本部分含有BERT的创新点之一：下一句预测 类标的生成
返回值：instances
以下在关键代码出进行注释
 

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
 if rng.random() < short_seq_prob:         #产生一个随机数如果小于short_seq_prob 则产生一个较短的训练序列
   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 = []    #产生训练集的候选集
 current_length = 0
 i = 0
 while i < len(document):
   segment = document[i]
   current_chunk.append(segment)
   current_length += len(segment)
   if i == len(document) - 1 or current_length >= target_seq_length:
     if current_chunk:
       # `a_end` is how many segments from `current_chunk` go into the `A`
       # (first) sentence.
       a_end = 1
       if len(current_chunk) >= 2:
         a_end = rng.randint(1, len(current_chunk) - 1)           #从current_chunk中随机选出一个文档作为句子1的截止文档

       tokens_a = []
       for j in range(a_end):
         tokens_a.extend(current_chunk[j])   #将截止文档之前的文档都加入到tokens_a

       tokens_b = []
       # Random next
       is_random_next = False
       if len(current_chunk) == 1 or rng.random() < 0.5:    #候选集只有一句的情况则随机抽取句子作为句子2；或以0.5的概率随机抽取句子作为句子2
         is_random_next = True
         target_b_length = target_seq_length - len(tokens_a)

         # This should rarely go for more than one iteration for large
         # corpora. However, just to be careful, we try to make sure that
         # the random document is not the same as the document
         # we're processing.
         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]        #随机找一个文档作为截止文档
         random_start = rng.randint(0, len(random_document) - 1)                #随机找一个初始文档
         for j in range(random_start, len(random_document)):                      
           tokens_b.extend(random_document[j])                                           #将随机文档加入到token_b
           if len(tokens_b) >= target_b_length:
             break
         # We didn't actually use these segments so we "put them back" so
         # they don't go to waste.
         num_unused_segments = len(current_chunk) - a_end
         i -= num_unused_segments
       # Actual next
       else:
         is_random_next = False                       以第1句的后续作为句子2
         for j in range(a_end, len(current_chunk)):
           tokens_b.extend(current_chunk[j])
       truncate_seq_pair(tokens_a, tokens_b, max_num_tokens, rng)      #对两个句子进行长度剪裁

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

       tokens = []
       segment_ids = []
       tokens.append("[CLS]")
       segment_ids.append(0)
       for token in tokens_a:
         tokens.append(token)
         segment_ids.append(0)

       tokens.append("[SEP]")
       segment_ids.append(0)

       for token in tokens_b:
         tokens.append(token)
         segment_ids.append(1)
       tokens.append("[SEP]")
       segment_ids.append(1)

       (tokens, masked_lm_positions,
        masked_lm_labels) = create_masked_lm_predictions(             #对token创建mask
            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

随机遮蔽
这部分对token进行随机mask。这部分是BERT的创新点之二，随机遮蔽。为了防止双向模型在多层之后“看到自己”。这里对一部分词进行随机遮蔽，并在预训练中进行预测。遮蔽方案：
1.以80%的概率直接变成[MASK]
2.以10%的概率保留原词
3.以10%的概率在词典中随机找一个词替代
返回值：经过随机遮蔽后的（词，遮蔽位置，遮蔽前原词）
 

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
  cand_indexes.append(i)

rng.shuffle(cand_indexes)                 #打乱顺序

output_tokens = list(tokens)

masked_lm = collections.namedtuple("masked_lm", ["index", "label"])  # p定义一个名为masked_lm的元组，里面有两个属性

num_to_predict = min(max_predictions_per_seq,
                     max(1, int(round(len(tokens) * masked_lm_prob))))  #所有要mask的词的数量为定值，取两个定义好参数的最小值

masked_lms = []
covered_indexes = set()
for index in cand_indexes:
  if len(masked_lms) >= num_to_predict:
    break
  if index in covered_indexes:
    continue
  covered_indexes.add(index)              #要被mask的词的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              #用masked_token替换原词

  masked_lms.append(masked_lm(index=index, label=tokens[index]))   

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)          #被mask的index
  masked_lm_labels.append(p.label)                 #被mask的label（即原词）

return (output_tokens, masked_lm_positions, masked_lm_labels)

 

输出

最后是将处理好的数据保存为tfrecord文件。首先将token转为id，增加input_mask用于记录实句长度。最后将不到最大长度的部分用0补齐。

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)      #词转id
   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:            #未到最大长度时后面补0
     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)                      #mask位置记录
   masked_lm_ids = tokenizer.convert_tokens_to_ids(instance.masked_lm_labels)    #mask预测值转id
   masked_lm_weights = [1.0] * len(masked_lm_ids)              #mask位置的权重都为1，用于排除后续的“0”以便loss计算

   while len(masked_lm_positions) < max_predictions_per_seq:      #补0
     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))      #生成训练样例

   writers[writer_index].write(tf_example.SerializeToString())     #输出到文件
   writer_index = (writer_index + 1) % len(writers)   

   total_written += 1

   if inst_index < 20:               对前20个训练样例进行打印
     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)

 

结果一览

最后打印的结果是这酱的

 

 

谷歌对训练数据的处理就介绍这么多，如果有错误欢迎大家批评指正，如果有问题也欢迎大家提问互相探讨。关于模型篇的代码解析我会在下一篇博客中给出。