BERT学习笔记：run-pretraining.py

BERT 源码初探之 run_pretraining.py

本文源码来源于 Github上的BERT 项目中的 run_pretraining.py 文件。阅读本文需要对Attention Is All You Need以及BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding两篇论文有所了解，以及部分关于深度学习、自然语言处理和Tensorflow的储备知识。

0 前言

• 关于Tensorflow：本文基于谷歌官方在GitHub上公布的BERT预训练模型，基于Tensorflow 1.13.1 运行。有关Tensorflow的部分建议参照官方网站。
• 关于Transformer：Transformer是Google提出的一种完全基于注意力机制的模型，想要对齐进行了解请参照官方论文Attention Is All You Need或者我的另一篇博客Transformer 学习笔记。
• 关于BERT：BERT也是Google提出的一个基于Transformer的预训练网络模型，更多和该模型有关的内容请参照官方论文BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding、官方代码实现Github上的BERT以及我的另一篇博客BERT 学习笔记。

1 简介

略。

2 源码解释

2.1 参数设置

2.1.1 必须参数

flags.DEFINE_string(
    "bert_config_file", None,
    "The config json file corresponding to the pre-trained BERT model. "
    "This specifies the model architecture.")

flags.DEFINE_string(
    "input_file", None,
    "Input TF example files (can be a glob or comma separated).")

flags.DEFINE_string(
    "output_dir", None,
    "The output directory where the model checkpoints will be written.")

• BERT 的 JSON 格式的配置文件的路径
• 输入文件
• 输出目录

2.2.2 其他参数

flags.DEFINE_string(
    "init_checkpoint", None,
    "Initial checkpoint (usually from a pre-trained BERT model).")

flags.DEFINE_integer(
    "max_seq_length", 128,
    "The maximum total input sequence length after WordPiece tokenization. "
    "Sequences longer than this will be truncated, and sequences shorter "
    "than this will be padded. Must match data generation.")

• 用于初始化的检查点
• 最大句子长度

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

每个句子的最大 MLM 预测数，必须和数据匹配。关于 MLM 模型，详情请参照 BERT论文。

flags.DEFINE_bool("do_train", False, "Whether to run training.")

flags.DEFINE_bool("do_eval", False, "Whether to run eval on the dev set.")

flags.DEFINE_integer("train_batch_size", 32, "Total batch size for training.")

flags.DEFINE_integer("eval_batch_size", 8, "Total batch size for eval.")

flags.DEFINE_float("learning_rate", 5e-5, "The initial learning rate for Adam.")

flags.DEFINE_integer("num_train_steps", 100000, "Number of training steps.")

flags.DEFINE_integer("num_warmup_steps", 10000, "Number of warmup steps.")

flags.DEFINE_integer("save_checkpoints_steps", 1000,
                     "How often to save the model checkpoint.")

flags.DEFINE_integer("iterations_per_loop", 1000,
                     "How many steps to make in each estimator call.")

flags.DEFINE_integer("max_eval_steps", 100, "Maximum number of eval steps.")

• 是否进行训练
• 是否在验证集上进行验证
• 训练批大小
• 验证批大小
• 初始化学习率
• 训练步数
• warmup步数
• 保存checkpoint的间隔
• 每隔多少步进行一次估计
• 评估步数的最大值

2.2.3 TPU相关

tf.flags.DEFINE_string(
    "tpu_name", None,
    "The Cloud TPU to use for training. This should be either the name "
    "used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 "
    "url.")

tf.flags.DEFINE_string(
    "tpu_zone", None,
    "[Optional] GCE zone where the Cloud TPU is located in. If not "
    "specified, we will attempt to automatically detect the GCE project from "
    "metadata.")

tf.flags.DEFINE_string(
    "gcp_project", None,
    "[Optional] Project name for the Cloud TPU-enabled project. If not "
    "specified, we will attempt to automatically detect the GCE project from "
    "metadata.")

tf.flags.DEFINE_string("master", None, "[Optional] TensorFlow master URL.")

flags.DEFINE_integer(
    "num_tpu_cores", 8,
    "Only used if `use_tpu` is True. Total number of TPU cores to use.")

这部分参数和TPU配置相关，不在此详细说明，详情请参照上一篇博客或者自行了解和TPU有关的知识。

2.2 建立模型

2.2.1 为TPU估计器自定义一个建立模型的方法(model_fn_builder)

def model_fn_builder(bert_config, init_checkpoint, learning_rate,
                     num_train_steps, num_warmup_steps, use_tpu,
                     use_one_hot_embeddings):

  def model_fn(features, labels, mode, params):
    
  ……

  return model_fn    

我们需要根据自己的设置为TPU的Estimator来自定义一个建立模型的函数。

    tf.logging.info("*** Features ***")
    for name in sorted(features.keys()):
      tf.logging.info("  name = %s, shape = %s" % (name, features[name].shape))

打印特征信息。

    input_ids = features["input_ids"]
    input_mask = features["input_mask"]
    segment_ids = features["segment_ids"]
    masked_lm_positions = features["masked_lm_positions"]
    masked_lm_ids = features["masked_lm_ids"]
    masked_lm_weights = features["masked_lm_weights"]
    next_sentence_labels = features["next_sentence_labels"]

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

获取特征

    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)

根据特征和配置文件建立 BERT 模型

    (masked_lm_loss,
     masked_lm_example_loss, masked_lm_log_probs) = get_masked_lm_output(
         bert_config, model.get_sequence_output(), model.get_embedding_table(),
         masked_lm_positions, masked_lm_ids, masked_lm_weights)

    (next_sentence_loss, next_sentence_example_loss,
     next_sentence_log_probs) = get_next_sentence_output(
         bert_config, model.get_pooled_output(), next_sentence_labels)
    
    total_loss = masked_lm_loss + next_sentence_loss         

获取 MLM 部分的输出和 next sentence 部分的输出，并计算总损失。

    tvars = tf.trainable_variables()

    initialized_variable_names = {}
    scaffold_fn = None
    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)

获取需要训练的变量名称，和检查点中的变量取并集。

    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)

打印所有需要训练的变量名称，是否来源与检查点，和详细信息。

    output_spec = None

准备构建输出

    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)

在训练模式下获取TPUEstimatorSpec

    elif mode == tf.estimator.ModeKeys.EVAL:

如果是在验证模式下，首先需要构建计算损失和准确度的函数

      def metric_fn(masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids,
                    masked_lm_weights, next_sentence_example_loss,
                    next_sentence_log_probs, next_sentence_labels):
        masked_lm_log_probs = tf.reshape(masked_lm_log_probs,
                                         [-1, masked_lm_log_probs.shape[-1]])
        masked_lm_predictions = tf.argmax(
            masked_lm_log_probs, axis=-1, output_type=tf.int32)
        masked_lm_example_loss = tf.reshape(masked_lm_example_loss, [-1])
        masked_lm_ids = tf.reshape(masked_lm_ids, [-1])
        masked_lm_weights = tf.reshape(masked_lm_weights, [-1])
        masked_lm_accuracy = tf.metrics.accuracy(
            labels=masked_lm_ids,
            predictions=masked_lm_predictions,
            weights=masked_lm_weights)
        masked_lm_mean_loss = tf.metrics.mean(
            values=masked_lm_example_loss, weights=masked_lm_weights)

        next_sentence_log_probs = tf.reshape(
            next_sentence_log_probs, [-1, next_sentence_log_probs.shape[-1]])
        next_sentence_predictions = tf.argmax(
            next_sentence_log_probs, axis=-1, output_type=tf.int32)
        next_sentence_labels = tf.reshape(next_sentence_labels, [-1])
        next_sentence_accuracy = tf.metrics.accuracy(
            labels=next_sentence_labels, predictions=next_sentence_predictions)
        next_sentence_mean_loss = tf.metrics.mean(
            values=next_sentence_example_loss)

        return {
            "masked_lm_accuracy": masked_lm_accuracy,
            "masked_lm_loss": masked_lm_mean_loss,
            "next_sentence_accuracy": next_sentence_accuracy,
            "next_sentence_loss": next_sentence_mean_loss,
        }

构建计算损失和准确度的函数如上所示

      eval_metrics = (metric_fn, [
          masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids,
          masked_lm_weights, next_sentence_example_loss,
          next_sentence_log_probs, next_sentence_labels
      ])
      output_spec = tf.contrib.tpu.TPUEstimatorSpec(
          mode=mode,
          loss=total_loss,
          eval_metrics=eval_metrics,
          scaffold_fn=scaffold_fn)

然后构建输出

    else:
      raise ValueError("Only TRAIN and EVAL modes are supported: %s" % (mode))

    return output_spec

如果既不是训练又不是验证那么就返回 ValueError,否则就返回 output_spec

2.2.2 获取 MLM 部分的 loss 和 log probs(get_masked_lm_output)

def get_masked_lm_output(bert_config, input_tensor, output_weights, positions,
                         label_ids, label_weights):
                         
  input_tensor = gather_indexes(input_tensor, positions)

定义方法，获取输入向量

  with tf.variable_scope("cls/predictions"):
      input_tensor = tf.layers.dense(
          input_tensor,
          units=bert_config.hidden_size,
          activation=modeling.get_activation(bert_config.hidden_act),
          kernel_initializer=modeling.create_initializer(
              bert_config.initializer_range))
      input_tensor = modeling.layer_norm(input_tensor)

在输入层上搭建一个在预训练前不被使用的全连接层。

    output_bias = tf.get_variable(
        "output_bias",
        shape=[bert_config.vocab_size],
        initializer=tf.zeros_initializer())
    logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
    logits = tf.nn.bias_add(logits, output_bias)
    log_probs = tf.nn.log_softmax(logits, axis=-1)

输出的权重和输入嵌入相同，但是在输出中有一个对应每个 token 的权重。

    label_ids = tf.reshape(label_ids, [-1])
    label_weights = tf.reshape(label_weights, [-1])

    one_hot_labels = tf.one_hot(
        label_ids, depth=bert_config.vocab_size, dtype=tf.float32)

关于 label 的一些格式处理，处理完之后把 label 转化成 one hot 类型的输出。

    per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels, axis=[-1])
    numerator = tf.reduce_sum(label_weights * per_example_loss)
    denominator = tf.reduce_sum(label_weights) + 1e-5
    loss = numerator / denominator

  return (loss, per_example_loss, log_probs)

计算 loss ，并返回最终的结果。

2.2.3 获取 next sentence prediction（下一句预测） 部分的 loss 以及 log probs (get_next_sentence_output)

def get_next_sentence_output(bert_config, input_tensor, labels):

定义方法头

注意，这是一个简单的二分类问题，0代表是真实的下一句，而1代表的是随机的句子，具体内容请参考BERT论文。

  with tf.variable_scope("cls/seq_relationship"):
    output_weights = tf.get_variable(
        "output_weights",
        shape=[2, bert_config.hidden_size],
        initializer=modeling.create_initializer(bert_config.initializer_range))
    output_bias = tf.get_variable(
        "output_bias", shape=[2], initializer=tf.zeros_initializer())

获取输出的权重(weights)和偏置值(bias)。

    logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
    logits = tf.nn.bias_add(logits, output_bias)
    log_probs = tf.nn.log_softmax(logits, axis=-1)
    labels = tf.reshape(labels, [-1])
    one_hot_labels = tf.one_hot(labels, depth=2, 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, log_probs)

计算所需的返回值并返回。

2.2.4 在一个小批次上收集特定位置的向量(gather_indexes)

def gather_indexes(sequence_tensor, positions):
  """Gathers the vectors at the specific positions over a minibatch."""
  sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3)
  batch_size = sequence_shape[0]
  seq_length = sequence_shape[1]
  width = sequence_shape[2]

  flat_offsets = tf.reshape(
      tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1])
  flat_positions = tf.reshape(positions + flat_offsets, [-1])
  flat_sequence_tensor = tf.reshape(sequence_tensor,
                                    [batch_size * seq_length, width])
  output_tensor = tf.gather(flat_sequence_tensor, flat_positions)
  return output_tensor

此方法比较简单容易理解，目的就是为了获取一个句子张量上特定位置的张量。

2.3 自定义输入函数

def input_fn_builder(input_files,
                     max_seq_length,
                     max_predictions_per_seq,
                     is_training,
                     num_cpu_threads=4):

此方法定义，目的是为了获得一个用于获取输入数据的 input_fn 函数。

  def input_fn(params):
    batch_size = params["batch_size"]

    name_to_features = {
        "input_ids":
            tf.FixedLenFeature([max_seq_length], tf.int64),
        "input_mask":
            tf.FixedLenFeature([max_seq_length], tf.int64),
        "segment_ids":
            tf.FixedLenFeature([max_seq_length], tf.int64),
        "masked_lm_positions":
            tf.FixedLenFeature([max_predictions_per_seq], tf.int64),
        "masked_lm_ids":
            tf.FixedLenFeature([max_predictions_per_seq], tf.int64),
        "masked_lm_weights":
            tf.FixedLenFeature([max_predictions_per_seq], tf.float32),
        "next_sentence_labels":
            tf.FixedLenFeature([1], tf.int64),
    }

开始构建真正的输入函数，首先获取批大小以及根据名称定义的特征

    if is_training:
      d = tf.data.Dataset.from_tensor_slices(tf.constant(input_files))
      d = d.repeat()
      d = d.shuffle(buffer_size=len(input_files))

      cycle_length = min(num_cpu_threads, len(input_files))

      d = d.apply(
          tf.contrib.data.parallel_interleave(
              tf.data.TFRecordDataset,
              sloppy=is_training,
              cycle_length=cycle_length))
      d = d.shuffle(buffer_size=100)

在训练状态下，我们希望尽可能地并行读入文件并且打乱顺序

• cycle_length 代表了能够并行读入文件的数量
• sloppy 模式代表交叉可能会不准确，这增大了训练状态下的随机性

    else:
      d = tf.data.TFRecordDataset(input_files)
      d = d.repeat()

在验证状态下，我们并不希望打乱顺序，同时是否并行也并不关心。同时我们希望用固定的训练步数去训练。

    d = d.apply(
        tf.contrib.data.map_and_batch(
            lambda record: _decode_record(record, name_to_features),
            batch_size=batch_size,
            num_parallel_batches=num_cpu_threads,
            drop_remainder=True))
    return d

• 在训练过程中我们需要 drop 掉残留量因为 TPU 需要固定的尺寸。
• 而在验证中我们假设在 CPU 或 GPU 上进行计算，所以我们不想 drop 多余的数据。

def _decode_record(record, name_to_features):
  example = tf.parse_single_example(record, name_to_features)

  for name in list(example.keys()):
    t = example[name]
    if t.dtype == tf.int64:
      t = tf.to_int32(t)
    example[name] = t

  return example

这个方法把一个 decord 解码成 tensorflow Example ，tf.Example 只支持int64，但是 TPU 只支持 int32 ，因此把所有的 int64 转换成 int32。

2.4 main(_) 函数

2.4.1 初始化部分

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

  if not FLAGS.do_train and not FLAGS.do_eval:
    raise ValueError("At least one of `do_train` or `do_eval` must be True.")

设置日志打印等级，确保 FLAGS.do_train 和 FLAGS.do_eval 至少有一个为 True 。

  bert_config = modeling.BertConfig.from_json_file(FLAGS.bert_config_file)

  tf.gfile.MakeDirs(FLAGS.output_dir)

加载配置文件，创建输出目录。

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

  tf.logging.info("*** Input Files ***")
  for input_file in input_files:
    tf.logging.info("  %s" % input_file)

获取输入文件的目录，并打印出来。

  tpu_cluster_resolver = None
  if FLAGS.use_tpu and FLAGS.tpu_name:
    tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(
        FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)

如果使用TPU那么就创建一个TPU集群分析器。

  is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2
  run_config = tf.contrib.tpu.RunConfig(
      cluster=tpu_cluster_resolver,
      master=FLAGS.master,
      model_dir=FLAGS.output_dir,
      save_checkpoints_steps=FLAGS.save_checkpoints_steps,
      tpu_config=tf.contrib.tpu.TPUConfig(
          iterations_per_loop=FLAGS.iterations_per_loop,
          num_shards=FLAGS.num_tpu_cores,
          per_host_input_for_training=is_per_host))

构建 run config

  model_fn = model_fn_builder(
      bert_config=bert_config,
      init_checkpoint=FLAGS.init_checkpoint,
      learning_rate=FLAGS.learning_rate,
      num_train_steps=FLAGS.num_train_steps,
      num_warmup_steps=FLAGS.num_warmup_steps,
      use_tpu=FLAGS.use_tpu,
      use_one_hot_embeddings=FLAGS.use_tpu)

构建 model_fn 方法

  estimator = tf.contrib.tpu.TPUEstimator(
      use_tpu=FLAGS.use_tpu,
      model_fn=model_fn,
      config=run_config,
      train_batch_size=FLAGS.train_batch_size,
      eval_batch_size=FLAGS.eval_batch_size)

构建估计器

2.4.2 训练部分

  if FLAGS.do_train:
    tf.logging.info("***** Running training *****")
    tf.logging.info("  Batch size = %d", FLAGS.train_batch_size)
    train_input_fn = input_fn_builder(
        input_files=input_files,
        max_seq_length=FLAGS.max_seq_length,
        max_predictions_per_seq=FLAGS.max_predictions_per_seq,
        is_training=True)
    estimator.train(input_fn=train_input_fn, max_steps=FLAGS.num_train_steps)

构建输入方法，进行训练。

2.4.3 验证部分

  if FLAGS.do_eval:
    tf.logging.info("***** Running evaluation *****")
    tf.logging.info("  Batch size = %d", FLAGS.eval_batch_size)

    eval_input_fn = input_fn_builder(
        input_files=input_files,
        max_seq_length=FLAGS.max_seq_length,
        max_predictions_per_seq=FLAGS.max_predictions_per_seq,
        is_training=False)

    result = estimator.evaluate(
        input_fn=eval_input_fn, steps=FLAGS.max_eval_steps)

    output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt")
    with tf.gfile.GFile(output_eval_file, "w") as writer:
      tf.logging.info("***** Eval results *****")
      for key in sorted(result.keys()):
        tf.logging.info("  %s = %s", key, str(result[key]))
        writer.write("%s = %s\n" % (key, str(result[key])))

构建输入函数，进行验证，保存并打印验证结果。

2.5 主程序入口

if __name__ == "__main__":
  flags.mark_flag_as_required("input_file")
  flags.mark_flag_as_required("bert_config_file")
  flags.mark_flag_as_required("output_dir")
  tf.app.run()

定义必须的参数，运行程序。

3 结论

上一篇博文介绍了用BERT模型做分类任务，这篇介绍了如何用自己的数据集对BERT模型进行预训练，接下来一篇的内容应该是

• BERT 模型代码阅读
• BERT 模型其他使用方法
• 关于 Tensorflow 的 预创建Estimator 和 自定义Estimator 的学习

三选一了吧。Google 的代码阅读起来是真的流畅啊。