Bert的MLM任务loss原理

         bert预训练有MLM和NSP两个任务，其中MLM是类似于“完形填空”的方式，对一个句子里的15%的词进行mask，通过双向transformer+feedforward+rediual_add+layer_norm完成对每个词的embedding编码，然后对mask的这个词进行预测，预测过程相当于做多分类，类别的个数是词汇的总个数，将mask的词的emb经过MLP变换生成在每个类别词汇上的logits 概率，label是mask位置上真实词在整个词汇上的one-hot编码，将logits和label计算交叉熵，又做了加权平均，即可得出MLM的loss，过程如下：

         源码中的get_masked_lm_output()方法过程解析：

1、输入input_tensor:[batch,maskednums, embed_size]

2、经过线性变换+layernorm:[batch,maskednums, 768]

3、logits：将embedding table[3万,768]作为变换矩阵，计算logits：[batch,maskednums, 3万]，相当于得出每个被盖住词在3万个词上的概率，其实就是3万个类别多分类

4、labels：one-hot编码[maskednums,3万]

5、计算交叉熵：[bactch, maskednums]

6、loss：加权平均得出一个实数

def get_masked_lm_output(bert_config, input_tensor, output_weights, positions,
                         label_ids, label_weights):
  """Get loss and log probs for the masked LM."""
  input_tensor = gather_indexes(input_tensor, positions)

  with tf.variable_scope("cls/predictions"):
    # We apply one more non-linear transformation before the output layer.
    # This matrix is not used after pre-training.
    with tf.variable_scope("transform"):
      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)

    # The output weights are the same as the input embeddings, but there is
    # an output-only bias for each token.
    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)

    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)

    # The `positions` tensor might be zero-padded (if the sequence is too
    # short to have the maximum number of predictions). The `label_weights`
    # tensor has a value of 1.0 for every real prediction and 0.0 for the
    # padding predictions.
    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)