本文链接：https://blog.csdn.net/weixin_39470744/article/details/84401339

         <!--一个博主专栏付费入口结束--><link rel="stylesheet" href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-4a3473df85.css"><div id="content_views" class="markdown_views prism-atom-one-dark"><!-- flowchart 箭头图标 勿删 --><svg xmlns="http://www.w3.org/2000/svg" style="display: none;"><path stroke-linecap="round" d="M5,0 0,2.5 5,5z" id="raphael-marker-block" style="-webkit-tap-highlight-color: rgba(0, 0, 0, 0);"></path></svg><p></p><div class="toc"><h3><a name="t0"></a>目录</h3><ul><li><a href="#_1" rel="nofollow" target="_self">前言</a></li><li><a href="#_6" rel="nofollow" target="_self">源码解析</a></li><ul><li><a href="#_7" rel="nofollow" target="_self">模型配置参数</a></li><li><a href="#BertModel_22" rel="nofollow" target="_self">BertModel</a></li><li><a href="#word_embedding_131" rel="nofollow" target="_self">word embedding</a></li><li><a href="#embedding_postprocessor_181" rel="nofollow" target="_self">embedding_postprocessor</a></li><li><a href="#Transformer_281" rel="nofollow" target="_self">Transformer</a></li><li><a href="#self_attention_423" rel="nofollow" target="_self">self_attention</a></li></ul><li><a href="#_620" rel="nofollow" target="_self">模型应用</a></li></ul></div><p></p>

前言

BERT的模型主要是基于Transformer架构（论文：Attention is all you need）。它抛开了RNN等固有模式，直接用注意力机制处理Seq2Seq问题，体现了大道至简的思想。网上对此模型解析的资料有很多，但大都千篇一律。这里推荐知乎的一篇《Attention is all you need》解读，我觉得这篇把transformer介绍的非常好。
由于模型最闹心的就是维度问题，维度理清了，理解模型就很容易，所以我在源码中会注释每个操作后tensor的维度信息。
下面开始介绍BERT的模型 modeling.py是怎么建立的，我始终认为读代码和注释是理解的最快方法，所以看代码时如果官方注释有的地方看不懂。请善看中文注释和维度信息

源码解析

模型配置参数

" attention_probs_dropout_prob": 0.1,         #乘法attention时，softmax后dropout概率"hidden_act": "gelu",         #激活函数"hidden_dropout_prob": 0.1,        #隐藏层dropout概率"hidden_size": 768,                #隐藏单元数"initializer_range": 0.02,          #初始化范围"intermediate_size": 3072,      #升维维度"max_position_embeddings": 512,     #一个大于seq_length的参数，用于生成position_embedding"num_attention_heads": 12,    #每个隐藏层中的attention head数"num_hidden_layers": 12,       #隐藏层数"type_vocab_size": 2,        #segment_ids类别 [0,1]"vocab_size": 30522       #词典中词数

这里的输入参数：input_ids,input_mask,token_type_ids对应上篇文章中输出的input_ids,input_mask,segment_ids

BertModel

这部分是总流程，整个modling脚本有900多行代码，所以我列个流程图一部一部走。整体流程如下。首先对input_ids和token_type_ids进行embedding操作，将embedding结果送入Transformer训练，最后得到编码结果。

def __init__(self,config,is_training,input_ids,input_mask=None,token_type_ids=None,use_one_hot_embeddings=True,scope=None):"""Constructor for BertModel.Args:config: `BertConfig` instance.is_training: bool. rue for training model, false for eval model. Controlswhether 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 wordembeddings or tf.embedding_lookup() for the word embeddings. On the TPU,it is must faster if this is True, on the CPU or GPU, it is faster ifthis is False.scope: (optional) variable scope. Defaults to "bert".Raises:ValueError: The config is invalid or one of the input tensor shapesis invalid."""config = copy.deepcopy(config)if not is_training:config.hidden_dropout_prob = 0.0config.attention_probs_dropout_prob = 0.0

input_shape <span class="token operator">=</span> <span class="token function">get_shape_list</span><span class="token punctuation">(</span>input_ids<span class="token punctuation">,</span> expected_rank<span class="token operator">=</span><span class="token number">2</span><span class="token punctuation">)</span>
batch_size <span class="token operator">=</span> input_shape<span class="token punctuation">[</span><span class="token number">0</span><span class="token punctuation">]</span>
seq_length <span class="token operator">=</span> input_shape<span class="token punctuation">[</span><span class="token number">1</span><span class="token punctuation">]</span><span class="token keyword">if</span> input_mask is None<span class="token punctuation">:</span>input_mask <span class="token operator">=</span> tf<span class="token punctuation">.</span><span class="token function">ones</span><span class="token punctuation">(</span>shape<span class="token operator">=</span><span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> seq_length<span class="token punctuation">]</span><span class="token punctuation">,</span> dtype<span class="token operator">=</span>tf<span class="token punctuation">.</span>int32<span class="token punctuation">)</span><span class="token keyword">if</span> token_type_ids is None<span class="token punctuation">:</span>token_type_ids <span class="token operator">=</span> tf<span class="token punctuation">.</span><span class="token function">zeros</span><span class="token punctuation">(</span>shape<span class="token operator">=</span><span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> seq_length<span class="token punctuation">]</span><span class="token punctuation">,</span> dtype<span class="token operator">=</span>tf<span class="token punctuation">.</span>int32<span class="token punctuation">)</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span>scope<span class="token punctuation">,</span> default_name<span class="token operator">=</span><span class="token string">"bert"</span><span class="token punctuation">)</span><span class="token punctuation">:</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"embeddings"</span><span class="token punctuation">)</span><span class="token punctuation">:</span># Perform embedding lookup on the word ids<span class="token punctuation">.</span>#<span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span>seq_length<span class="token punctuation">,</span>embedding_size<span class="token punctuation">]</span>    <span class="token punctuation">[</span>vocab_size<span class="token punctuation">,</span>embedding_size<span class="token punctuation">]</span><span class="token punctuation">(</span>self<span class="token punctuation">.</span>embedding_output<span class="token punctuation">,</span> self<span class="token punctuation">.</span>embedding_table<span class="token punctuation">)</span> <span class="token operator">=</span> <span class="token function">embedding_lookup</span><span class="token punctuation">(</span>    #word_embeddinginput_ids<span class="token operator">=</span>input_ids<span class="token punctuation">,</span>           #<span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span>seq_length<span class="token punctuation">]</span>vocab_size<span class="token operator">=</span>config<span class="token punctuation">.</span>vocab_size<span class="token punctuation">,</span>embedding_size<span class="token operator">=</span>config<span class="token punctuation">.</span>hidden_size<span class="token punctuation">,</span>initializer_range<span class="token operator">=</span>config<span class="token punctuation">.</span>initializer_range<span class="token punctuation">,</span>word_embedding_name<span class="token operator">=</span><span class="token string">"word_embeddings"</span><span class="token punctuation">,</span>use_one_hot_embeddings<span class="token operator">=</span>use_one_hot_embeddings<span class="token punctuation">)</span># Add positional embeddings and token type embeddings<span class="token punctuation">,</span> then layer# normalize and perform dropout<span class="token punctuation">.</span>self<span class="token punctuation">.</span>embedding_output <span class="token operator">=</span> <span class="token function">embedding_postprocessor</span><span class="token punctuation">(</span>       #token_embedding和position_embedding        <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span>seq_length<span class="token punctuation">,</span>embedding_size<span class="token punctuation">]</span>input_tensor<span class="token operator">=</span>self<span class="token punctuation">.</span>embedding_output<span class="token punctuation">,</span>use_token_type<span class="token operator">=</span>True<span class="token punctuation">,</span>token_type_ids<span class="token operator">=</span>token_type_ids<span class="token punctuation">,</span>token_type_vocab_size<span class="token operator">=</span>config<span class="token punctuation">.</span>type_vocab_size<span class="token punctuation">,</span>token_type_embedding_name<span class="token operator">=</span><span class="token string">"token_type_embeddings"</span><span class="token punctuation">,</span>use_position_embeddings<span class="token operator">=</span>True<span class="token punctuation">,</span>position_embedding_name<span class="token operator">=</span><span class="token string">"position_embeddings"</span><span class="token punctuation">,</span>initializer_range<span class="token operator">=</span>config<span class="token punctuation">.</span>initializer_range<span class="token punctuation">,</span>max_position_embeddings<span class="token operator">=</span>config<span class="token punctuation">.</span>max_position_embeddings<span class="token punctuation">,</span>dropout_prob<span class="token operator">=</span>config<span class="token punctuation">.</span>hidden_dropout_prob<span class="token punctuation">)</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"encoder"</span><span class="token punctuation">)</span><span class="token punctuation">:</span># This converts a <span class="token number">2</span>D mask <span class="token keyword">of</span> shape <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> seq_length<span class="token punctuation">]</span> to a <span class="token number">3</span>D# mask <span class="token keyword">of</span> shape <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> seq_length<span class="token punctuation">,</span> seq_length<span class="token punctuation">]</span> which is used# <span class="token keyword">for</span> the attention scores<span class="token punctuation">.</span>attention_mask <span class="token operator">=</span> <span class="token function">create_attention_mask_from_input_mask</span><span class="token punctuation">(</span>     input_ids<span class="token punctuation">,</span> input_mask<span class="token punctuation">)</span># Run the stacked transformer<span class="token punctuation">.</span># <span class="token template-string"><span class="token string">`sequence_output`</span></span> shape <span class="token operator">=</span> <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> seq_length<span class="token punctuation">,</span> hidden_size<span class="token punctuation">]</span><span class="token punctuation">.</span>self<span class="token punctuation">.</span>all_encoder_layers <span class="token operator">=</span> <span class="token function">transformer_model</span><span class="token punctuation">(</span>        #transformer_model  <span class="token function">list</span><span class="token punctuation">(</span>#<span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span>seq_length<span class="token punctuation">,</span>embedding_size<span class="token punctuation">]</span><span class="token punctuation">)</span>input_tensor<span class="token operator">=</span>self<span class="token punctuation">.</span>embedding_output<span class="token punctuation">,</span>attention_mask<span class="token operator">=</span>attention_mask<span class="token punctuation">,</span>hidden_size<span class="token operator">=</span>config<span class="token punctuation">.</span>hidden_size<span class="token punctuation">,</span>num_hidden_layers<span class="token operator">=</span>config<span class="token punctuation">.</span>num_hidden_layers<span class="token punctuation">,</span>num_attention_heads<span class="token operator">=</span>config<span class="token punctuation">.</span>num_attention_heads<span class="token punctuation">,</span>intermediate_size<span class="token operator">=</span>config<span class="token punctuation">.</span>intermediate_size<span class="token punctuation">,</span>intermediate_act_fn<span class="token operator">=</span><span class="token function">get_activation</span><span class="token punctuation">(</span>config<span class="token punctuation">.</span>hidden_act<span class="token punctuation">)</span><span class="token punctuation">,</span>hidden_dropout_prob<span class="token operator">=</span>config<span class="token punctuation">.</span>hidden_dropout_prob<span class="token punctuation">,</span>attention_probs_dropout_prob<span class="token operator">=</span>config<span class="token punctuation">.</span>attention_probs_dropout_prob<span class="token punctuation">,</span>initializer_range<span class="token operator">=</span>config<span class="token punctuation">.</span>initializer_range<span class="token punctuation">,</span>do_return_all_layers<span class="token operator">=</span>True<span class="token punctuation">)</span>self<span class="token punctuation">.</span>sequence_output <span class="token operator">=</span> self<span class="token punctuation">.</span>all_encoder_layers<span class="token punctuation">[</span><span class="token operator">-</span><span class="token number">1</span><span class="token punctuation">]</span>     #获取最后一层的输出# The <span class="token string">"pooler"</span> converts the encoded sequence tensor <span class="token keyword">of</span> shape# <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> seq_length<span class="token punctuation">,</span> hidden_size<span class="token punctuation">]</span> to a tensor <span class="token keyword">of</span> shape# <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> hidden_size<span class="token punctuation">]</span><span class="token punctuation">.</span> This is necessary <span class="token keyword">for</span> segment<span class="token operator">-</span>level# <span class="token punctuation">(</span>or segment<span class="token operator">-</span>pair<span class="token operator">-</span>level<span class="token punctuation">)</span> classification tasks where we need a fixed# dimensional representation <span class="token keyword">of</span> the segment<span class="token punctuation">.</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"pooler"</span><span class="token punctuation">)</span><span class="token punctuation">:</span># We <span class="token string">"pool"</span> the model by simply taking the hidden state corresponding# to the first token<span class="token punctuation">.</span> We assume that <span class="token keyword">this</span> has been pre<span class="token operator">-</span>trainedfirst_token_tensor <span class="token operator">=</span> tf<span class="token punctuation">.</span><span class="token function">squeeze</span><span class="token punctuation">(</span>self<span class="token punctuation">.</span>sequence_output<span class="token punctuation">[</span><span class="token punctuation">:</span><span class="token punctuation">,</span> <span class="token number">0</span><span class="token punctuation">:</span><span class="token number">1</span><span class="token punctuation">,</span> <span class="token punctuation">:</span><span class="token punctuation">]</span><span class="token punctuation">,</span> axis<span class="token operator">=</span><span class="token number">1</span><span class="token punctuation">)</span>    #取每个每个训练语料的第一个词的编码结果<span class="token punctuation">[</span><span class="token constant">CLS</span><span class="token punctuation">]</span>，它有整条训练语料的编码信息   <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> hidden_size<span class="token punctuation">]</span>self<span class="token punctuation">.</span>pooled_output <span class="token operator">=</span> tf<span class="token punctuation">.</span>layers<span class="token punctuation">.</span><span class="token function">dense</span><span class="token punctuation">(</span>     #接一个全连接层进行输出 <span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> hidden_size<span class="token punctuation">]</span>first_token_tensor<span class="token punctuation">,</span>config<span class="token punctuation">.</span>hidden_size<span class="token punctuation">,</span>activation<span class="token operator">=</span>tf<span class="token punctuation">.</span>tanh<span class="token punctuation">,</span>kernel_initializer<span class="token operator">=</span><span class="token function">create_initializer</span><span class="token punctuation">(</span>config<span class="token punctuation">.</span>initializer_range<span class="token punctuation">)</span><span class="token punctuation">)</span>

word embedding

首先看word_embedding部分,它传入input_ids，运用one_hot为中介返回embedding结果

def embedding_lookup(input_ids,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 wordids.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 wordembeddings. If False, use `tf.nn.embedding_lookup()`. One hot is betterfor TPUs.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])                 #最低维扩维 [batch_size,seq_length,1]

embedding_table = tf.get_variable(
name=word_embedding_name,
shape=[vocab_size, embedding_size],
initializer=create_initializer(initializer_range))

if use_one_hot_embeddings:
flat_input_ids = tf.reshape(input_ids, [-1]) #[batch_sizeseq_length]
one_hot_input_ids = tf.one_hot(flat_input_ids, depth=vocab_size) #[batch_sizeseq_length,vocab_size]
output = tf.matmul(one_hot_input_ids, embedding_table) #[batch_size*seq_length,embedding_size]
else:
output = tf.nn.embedding_lookup(embedding_table, input_ids)

input_shape = get_shape_list(input_ids)

output = tf.reshape(output,
input_shape[0:-1] + [input_shape[-1] * embedding_size]) #[batch_size,seq_length,embedding_size]
return (output, embedding_table)

embedding_postprocessor

再看embedding_postprocessor 它包括token_type_embedding和position_embedding。也就是图中的Segement Embeddings和Position Embeddings。

但此代码中Position Embeddings部分与之前提出的Transformer不同，此代码中Position Embeddings是训练出来的，而传统的Transformer（如下）是固定值

def embedding_postprocessor(input_tensor,      #[batch_size,seq_length,embedding_size]use_token_type=False,token_type_ids=None,             #[batch_size,seq_length]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 variablefor token type ids.use_position_embeddings: bool. Whether to add position embeddings for theposition of each token in the sequence.position_embedding_name: string. The name of the embedding table variablefor positional embeddings.initializer_range: float. Range of the weight initialization.max_position_embeddings: int. Maximum sequence length that might ever beused with this model. This can be longer than the sequence length ofinput_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]width = input_shape[2]

output = input_tensor

if use_token_type: #Segement Embeddings部分
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.
flat_token_type_ids = tf.reshape(token_type_ids, [-1]) #[batch_sizeseq_length]
one_hot_ids = tf.one_hot(flat_token_type_ids, depth=token_type_vocab_size) #[batch_sizeseq_length,2] token_type只有0，1
token_type_embeddings = tf.matmul(one_hot_ids, token_type_table) #[batch_size*seq_length,embedding_size]
token_type_embeddings = tf.reshape(token_type_embeddings,
[batch_size, seq_length, width]) #[batch_size, seq_length, width=embedding_size]
output += token_type_embeddings #[batch_size, seq_length, embedding_size]

if use_position_embeddings: #Position Embeddings部分
assert_op = tf.assert_less_equal(seq_length, max_position_embeddings) #确保seq_length<max_position_embedding
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))
# Since the position embedding table is a learned variable, we create it
# using a (long) sequence length max_position_embeddings. The actual
# sequence length might be shorter than this, for faster training of
# tasks that do not have long sequences.
#
# So full_position_embeddings is effectively an embedding table
# for position [0, 1, 2, …, max_position_embeddings-1], and the current
# sequence has positions [0, 1, 2, … seq_length-1], so we can just
# perform a slice.
position_embeddings = tf.slice(full_position_embeddings, [0, 0], #[seq_length,embedding_size]
[seq_length, -1])
num_dims = len(output.shape.as_list())

  # Only the last two dimensions are <span class="token function">relevant</span> <span class="token punctuation">(</span><span class="token template-string"><span class="token string">`seq_length`</span></span> and <span class="token template-string"><span class="token string">`width`</span></span><span class="token punctuation">)</span><span class="token punctuation">,</span> so# we broadcast among the first dimensions<span class="token punctuation">,</span> which is typically just# the batch size<span class="token punctuation">.</span>position_broadcast_shape <span class="token operator">=</span> <span class="token punctuation">[</span><span class="token punctuation">]</span><span class="token keyword">for</span> _ <span class="token keyword">in</span> <span class="token function">range</span><span class="token punctuation">(</span>num_dims <span class="token operator">-</span> <span class="token number">2</span><span class="token punctuation">)</span><span class="token punctuation">:</span>position_broadcast_shape<span class="token punctuation">.</span><span class="token function">append</span><span class="token punctuation">(</span><span class="token number">1</span><span class="token punctuation">)</span>position_broadcast_shape<span class="token punctuation">.</span><span class="token function">extend</span><span class="token punctuation">(</span><span class="token punctuation">[</span>seq_length<span class="token punctuation">,</span> width<span class="token punctuation">]</span><span class="token punctuation">)</span>      #<span class="token punctuation">[</span><span class="token number">1</span><span class="token punctuation">,</span>seq_length<span class="token punctuation">,</span>embedding_size<span class="token punctuation">]</span>position_embeddings <span class="token operator">=</span> tf<span class="token punctuation">.</span><span class="token function">reshape</span><span class="token punctuation">(</span>position_embeddings<span class="token punctuation">,</span>     #<span class="token punctuation">[</span><span class="token number">1</span><span class="token punctuation">,</span>seq_length<span class="token punctuation">,</span>embedding_size<span class="token punctuation">]</span>position_broadcast_shape<span class="token punctuation">)</span>output <span class="token operator">+=</span> position_embeddings               #<span class="token punctuation">[</span>batch_size<span class="token punctuation">,</span> seq_length<span class="token punctuation">,</span> embedding_size<span class="token punctuation">]</span> 与#<span class="token punctuation">[</span><span class="token number">1</span><span class="token punctuation">,</span>seq_length<span class="token punctuation">,</span>embedding_size<span class="token punctuation">]</span>相加

#因为每一个batch的同一位置的position_embedding是一样的，所以相当于batch_size个position_embeddings与output相加

output = layer_norm_and_dropout(output, dropout_prob)
return output

Transformer

embedding之后，首先构造一个attention_mask，这个attention_mask表示的含义是将原来的input_mask的[batch_size,seq_length]扩维到[batch_size,from_seq_length,to_seq_length]。保证对于每个from_seq_length都有一个input_mask。之后将他们传入到transformer模型。
transformer整体架构如图所示

下面我们来看transformer_model。首先对embedding进行multi-head attention,对输入进行残差和layer_norm。后传入feed forward，再进行残差和layer_norm。
本块代码中与原论文中不一样的点为：在进行multi-head attention后先链接了一个全连接层，再进行的残差和layer_norm。而原论文中貌似没有那个全连接层。下面是代码，关键部分我已写上注释

def transformer_model(input_tensor,attention_mask=None,   #[batch_size,form_seq_length,to_seq_length]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/1706.03762Also see:https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/models/transformer.pyArgs: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 inpositions 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., feedforward) layer.intermediate_act_fn: function. The non-linear activation function to applyto 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 attentionprobabilities.initializer_range: float. Range of the initializer (stddev of truncatednormal).do_return_all_layers: Whether to also return all layers or just the finallayer.Returns:float Tensor of shape [batch_size, seq_length, hidden_size], the finalhidden 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) #这里官方说为了避免来回升降维，所以直接先变形为2D，最后再恢复成3D [batch_size*seq_length,hidden_size]

all_layer_outputs = []
for layer_idx in range(num_hidden_layers):
with tf.variable_scope(“layer_%d” % layer_idx):
layer_input = prev_output

  <span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"attention"</span><span class="token punctuation">)</span><span class="token punctuation">:</span>attention_heads <span class="token operator">=</span> <span class="token punctuation">[</span><span class="token punctuation">]</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"self"</span><span class="token punctuation">)</span><span class="token punctuation">:</span>attention_head <span class="token operator">=</span> <span class="token function">attention_layer</span><span class="token punctuation">(</span>            #进行self_attention 即multi<span class="token operator">-</span>head attentionfrom_tensor<span class="token operator">=</span>layer_input<span class="token punctuation">,</span>      #<span class="token punctuation">[</span>batch_size<span class="token operator">*</span>seq_length<span class="token punctuation">,</span>hidden_size<span class="token punctuation">]</span>to_tensor<span class="token operator">=</span>layer_input<span class="token punctuation">,</span>        #<span class="token punctuation">[</span>batch_size<span class="token operator">*</span>seq_length<span class="token punctuation">,</span>hidden_size<span class="token punctuation">]</span>attention_mask<span class="token operator">=</span>attention_mask<span class="token punctuation">,</span>num_attention_heads<span class="token operator">=</span>num_attention_heads<span class="token punctuation">,</span>size_per_head<span class="token operator">=</span>attention_head_size<span class="token punctuation">,</span>attention_probs_dropout_prob<span class="token operator">=</span>attention_probs_dropout_prob<span class="token punctuation">,</span>initializer_range<span class="token operator">=</span>initializer_range<span class="token punctuation">,</span>do_return_2d_tensor<span class="token operator">=</span>True<span class="token punctuation">,</span>batch_size<span class="token operator">=</span>batch_size<span class="token punctuation">,</span>from_seq_length<span class="token operator">=</span>seq_length<span class="token punctuation">,</span>to_seq_length<span class="token operator">=</span>seq_length<span class="token punctuation">)</span>attention_heads<span class="token punctuation">.</span><span class="token function">append</span><span class="token punctuation">(</span>attention_head<span class="token punctuation">)</span>attention_output <span class="token operator">=</span> None<span class="token keyword">if</span> <span class="token function">len</span><span class="token punctuation">(</span>attention_heads<span class="token punctuation">)</span> <span class="token operator">==</span> <span class="token number">1</span><span class="token punctuation">:</span>attention_output <span class="token operator">=</span> attention_heads<span class="token punctuation">[</span><span class="token number">0</span><span class="token punctuation">]</span><span class="token keyword">else</span><span class="token punctuation">:</span># In the <span class="token keyword">case</span> where we have other sequences<span class="token punctuation">,</span> we just concatenate# them to the self<span class="token operator">-</span>attention head before the projection<span class="token punctuation">.</span>attention_output <span class="token operator">=</span> tf<span class="token punctuation">.</span><span class="token function">concat</span><span class="token punctuation">(</span>attention_heads<span class="token punctuation">,</span> axis<span class="token operator">=</span><span class="token operator">-</span><span class="token number">1</span><span class="token punctuation">)</span># Run a linear projection <span class="token keyword">of</span> <span class="token template-string"><span class="token string">`hidden_size`</span></span> then add a residual# <span class="token keyword">with</span> <span class="token template-string"><span class="token string">`layer_input`</span></span><span class="token punctuation">.</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"output"</span><span class="token punctuation">)</span><span class="token punctuation">:</span>attention_output <span class="token operator">=</span> tf<span class="token punctuation">.</span>layers<span class="token punctuation">.</span><span class="token function">dense</span><span class="token punctuation">(</span>                  #对attention的输出做一个全连接层attention_output<span class="token punctuation">,</span>hidden_size<span class="token punctuation">,</span>kernel_initializer<span class="token operator">=</span><span class="token function">create_initializer</span><span class="token punctuation">(</span>initializer_range<span class="token punctuation">)</span><span class="token punctuation">)</span>attention_output <span class="token operator">=</span> <span class="token function">dropout</span><span class="token punctuation">(</span>attention_output<span class="token punctuation">,</span> hidden_dropout_prob<span class="token punctuation">)</span>   attention_output <span class="token operator">=</span> <span class="token function">layer_norm</span><span class="token punctuation">(</span>attention_output <span class="token operator">+</span> layer_input<span class="token punctuation">)</span>         #残差和layer_norm#Feed Foward过程，先对输出升维、再进行降维# The activation is only applied to the <span class="token string">"intermediate"</span> hidden layer<span class="token punctuation">.</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"intermediate"</span><span class="token punctuation">)</span><span class="token punctuation">:</span>intermediate_output <span class="token operator">=</span> tf<span class="token punctuation">.</span>layers<span class="token punctuation">.</span><span class="token function">dense</span><span class="token punctuation">(</span>             #升维attention_output<span class="token punctuation">,</span>intermediate_size<span class="token punctuation">,</span>activation<span class="token operator">=</span>intermediate_act_fn<span class="token punctuation">,</span>kernel_initializer<span class="token operator">=</span><span class="token function">create_initializer</span><span class="token punctuation">(</span>initializer_range<span class="token punctuation">)</span><span class="token punctuation">)</span># Down<span class="token operator">-</span>project back to <span class="token template-string"><span class="token string">`hidden_size`</span></span> then add the residual<span class="token punctuation">.</span><span class="token keyword">with</span> tf<span class="token punctuation">.</span><span class="token function">variable_scope</span><span class="token punctuation">(</span><span class="token string">"output"</span><span class="token punctuation">)</span><span class="token punctuation">:</span>                       #降维layer_output <span class="token operator">=</span> tf<span class="token punctuation">.</span>layers<span class="token punctuation">.</span><span class="token function">dense</span><span class="token punctuation">(</span>intermediate_output<span class="token punctuation">,</span>hidden_size<span class="token punctuation">,</span>kernel_initializer<span class="token operator">=</span><span class="token function">create_initializer</span><span class="token punctuation">(</span>initializer_range<span class="token punctuation">)</span><span class="token punctuation">)</span>layer_output <span class="token operator">=</span> <span class="token function">dropout</span><span class="token punctuation">(</span>layer_output<span class="token punctuation">,</span> hidden_dropout_prob<span class="token punctuation">)</span>layer_output <span class="token operator">=</span> <span class="token function">layer_norm</span><span class="token punctuation">(</span>layer_output <span class="token operator">+</span> attention_output<span class="token punctuation">)</span>    #加入残差prev_output <span class="token operator">=</span> layer_output                 #本层输出作为下一层输入all_layer_outputs<span class="token punctuation">.</span><span class="token function">append</span><span class="token punctuation">(</span>layer_output<span class="token punctuation">)</span>       #所有层的输出结果列表

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

self_attention

接下来介绍self_attention机制。他运用乘法注意力，自己和自己做attention，使每个词都全局语义信息。同时运用Multi-head attention。即将hidden_size平分为多个部分(head)。每个head进行self_attention。不同head学习不同子空间语义。

下面是代码，关键部分我已写上注释。首先将输入的key和value，reshape成[batch_size,num_head,seq_length,size_per_head]。在对这些head进行乘法注意力运算。经过softmax后乘以value。最后返回tensor with shape [batch_size*seq_length,hidden_size]

def attention_layer(from_tensor,      #from_tensor和to_tensor都是输入embedding  [batch_size*seq_length,hidden_size]to_tensor,attention_mask=None,    #[batch_size,form_seq_length,to_seq_length]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 "Attentionis all you Need". If `from_tensor` and `to_tensor` are the same, thenthis is self-attention. Each timestep in `from_tensor` attends to thecorresponding 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 listof 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 aresoftmaxed to obtain attention probabilities. The value tensors are theninterpolated by these probabilities, then concatenated back to a singletensor and returned.In practice, the multi-headed attention are done with transposes andreshapes 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. Theattention scores will effectively be set to -infinity for any positions inthe 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 theattention 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, theoutput 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 sizeof the 3D version of the `from_tensor` and `to_tensor`.from_seq_length: (Optional) If the input is 2D, this might be the seq lengthof the 3D version of the `from_tensor`.to_seq_length: (Optional) If the input is 2D, this might be the seq lengthof 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` istrue, 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 <span class="token operator">=</span> tf<span class="token punctuation">.</span><span class="token function">transpose</span><span class="token punctuation">(</span>output_tensor<span class="token punctuation">,</span> <span class="token punctuation">[</span><span class="token number">0</span><span class="token punctuation">,</span> <span class="token number">2</span><span class="token punctuation">,</span> <span class="token number">1</span><span class="token punctuation">,</span> <span class="token number">3</span><span class="token punctuation">]</span><span class="token punctuation">)</span>
<span class="token keyword">return</span> 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) #[batch_sizeseq_length,hidden_size]
to_tensor_2d = reshape_to_matrix(to_tensor) #[batch_sizeseq_length,hidden_size]
#首先将key和value输入进全连接层但是激活函数为None，这里为什么我也不知道。。。

`query_layer` = [BF, NH]

query_layer = tf.layers.dense(
from_tensor_2d,
num_attention_heads size_per_head,
activation=query_act, #None
name=“query”,
kernel_initializer=create_initializer(initializer_range)) # [batch_sizeseq_length,hidden_size] hidden_size即num_attention_heads*size_per_head

`key_layer` = [BT, NH]

key_layer = tf.layers.dense(
to_tensor_2d,
num_attention_heads * size_per_head,
activation=key_act, #None
name=“key”,
kernel_initializer=create_initializer(initializer_range))

`value_layer` = [BT, NH]

value_layer = tf.layers.dense(
to_tensor_2d,
num_attention_heads * size_per_head,
activation=value_act, #None
name=“value”,
kernel_initializer=create_initializer(initializer_range))
#reshape成四位，用于注意力矩阵运算

`query_layer` = [B, N, F, H]

query_layer = transpose_for_scores(query_layer, batch_size, #将num_attention_heads调到第二维。这里表示每个batch有N个head，每个head有F个token，每个token用H表示。不同head学习不同子空间的特征
num_attention_heads, from_seq_length,
size_per_head)

`key_layer` = [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. 乘法注意力

`attention_scores` = [B, N, F, T]

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])

#这部分将每条训练语料的结尾padding的部分都变为一个极小值，其他有实数据的部分都为<span class="token number">0</span>
# Since attention_mask is <span class="token number">1.0</span> <span class="token keyword">for</span> positions we want to attend and <span class="token number">0.0</span> <span class="token keyword">for</span>
# masked positions<span class="token punctuation">,</span> <span class="token keyword">this</span> operation will create a tensor which is <span class="token number">0.0</span> <span class="token keyword">for</span>
# positions we want to attend and <span class="token operator">-</span><span class="token number">10000.0</span> <span class="token keyword">for</span> masked positions<span class="token punctuation">.</span>
adder <span class="token operator">=</span> <span class="token punctuation">(</span><span class="token number">1.0</span> <span class="token operator">-</span> tf<span class="token punctuation">.</span><span class="token function">cast</span><span class="token punctuation">(</span>attention_mask<span class="token punctuation">,</span> tf<span class="token punctuation">.</span>float32<span class="token punctuation">)</span><span class="token punctuation">)</span> <span class="token operator">*</span> <span class="token operator">-</span><span class="token number">10000.0</span># Since we are adding it to the raw scores before the softmax<span class="token punctuation">,</span> <span class="token keyword">this</span> is
# effectively the same <span class="token keyword">as</span> removing these entirely<span class="token punctuation">.</span>
#相加后，有实数据的部分加的，padding部分都是一个极小值
attention_scores <span class="token operator">+=</span> 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]

注意力矩阵乘以value

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:

返回2D结果

# <span class="token template-string"><span class="token string">`context_layer`</span></span> <span class="token operator">=</span> <span class="token punctuation">[</span><span class="token constant">B</span><span class="token operator">*</span><span class="token constant">F</span><span class="token punctuation">,</span> <span class="token constant">N</span><span class="token operator">*</span><span class="token constant">V</span><span class="token punctuation">]</span>
context_layer <span class="token operator">=</span> tf<span class="token punctuation">.</span><span class="token function">reshape</span><span class="token punctuation">(</span>context_layer<span class="token punctuation">,</span><span class="token punctuation">[</span>batch_size <span class="token operator">*</span> from_seq_length<span class="token punctuation">,</span> num_attention_heads <span class="token operator">*</span> size_per_head<span class="token punctuation">]</span><span class="token punctuation">)</span>

else:
# context_layer = [B, F, NV]
context_layer = tf.reshape(
context_layer,
[batch_size, from_seq_length, num_attention_heads size_per_head])

return context_layer

模型应用

模型怎么用呢，在BertModel class中有两个函数。get_pool_output表示获取每个batch第一个词的[CLS]表示结果。BERT认为这个词包含了整条语料的信息；适用于句子级别的分类问题。get_sequence_output表示BERT最终的输出结果,shape为[batch_size,seq_length,hidden_size]。可以直观理解为对每条语料的最终表示，适用于seq2seq问题。

def get_pooled_output(self):return self.pooled_outp          #[batch_size, hidden_size]
def get_sequence_output(self):"""Gets final hidden layer of encoder.Returns:float Tensor of shape [batch_size, seq_length, hidden_size] correspondingto the final hidden of the transformer encoder."""return self.sequence_output

下一篇是训练过程。最近突然有两件事要忙，所以可能要鸽几天了

                                </div><link href="https://csdnimg.cn/release/phoenix/mdeditor/markdown_views-b6c3c6d139.css" rel="stylesheet"></div>

谷歌BERT预训练源码解析（二）：模型构建相关推荐

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Transfo ...
谷歌BERT预训练源码解析（三）：训练过程
目录前言源码解析主函数自定义模型遮蔽词预测下一句预测规范化数据集前言本部分介绍BERT训练过程,BERT模型训练过程是在自己的TPU上进行的,这部分我没做过研究所以不做深入探讨.BE ...
【深度学习模型】智云视图中文车牌识别源码解析(二)
[深度学习模型]智云视图中文车牌识别源码解析(二) 感受 HyperLPR可以识别多种中文车牌包括白牌,新能源车牌,使馆车牌,教练车牌,武警车牌等. 代码不可谓不混乱(别忘了这是职业公司的准产品级代码 ...
AlphaFold2源码解析(9)--模型之损失
AlphaFold2源码解析(9)–模型之损失损失函数和辅助头该网络是端到端训练的,梯度来自主帧对齐点误差 (FAPE) 损失 LFAPEL_{FAPE}LFAPE和许多辅助损失. 每个示例的总 ...
AlphaFold2源码解析(4)--模型架构
AlphaFold2源码解析(4)–模型架构我们将Alphafold的流程分为一下几个部分: 搜索同源序列和模板特征构造特征表示 MSA表示与残基对表示之间互相交换信息残基的抽象表示转换成具体 ...
php 框架源码分析,Laravel框架源码解析之模型Model原理与用法解析
本文实例讲述了Laravel框架源码解析之模型Model原理与用法.分享给大家供大家参考,具体如下: 前言提前预祝猿人们国庆快乐,吃好.喝好.玩好,我会在电视上看着你们. 根据单一责任开发原则来讲, ...
erlang下lists模块sort（排序）方法源码解析(二)
上接erlang下lists模块sort(排序)方法源码解析(一),到目前为止,list列表已经被分割成N个列表,而且每个列表的元素是有序的(从大到小) 下面我们重点来看看mergel和rmergel ...
Kubernetes学习笔记之Calico CNI Plugin源码解析(二)
女主宣言今天小编继续为大家分享Kubernetes Calico CNI Plugin学习笔记,希望能对大家有所帮助. PS:丰富的一线技术.多元化的表现形式,尽在"360云计算" ...
Mobx 源码解析二（autorun)
前言我们在Mobx 源码解析一(observable)已经知道了observable 做的事情了, 但是我们的还是没有讲解明白在我们的Demo中,我们在Button 的Click 事件中只是对ba ...

谷歌BERT预训练源码解析（二）：模型构建

前言

源码解析

模型配置参数

BertModel

word embedding

embedding_postprocessor

Transformer

The Transformer performs sum residuals on all layers so the input needs

to be the same as the 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.

self_attention

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

query_layer = [BF, NH]

key_layer = [BT, NH]

value_layer = [BT, NH]

query_layer = [B, N, F, H]

key_layer = [B, N, T, H]