bert-pytorch版源码详细解读

前言

bert作为当下最火的NLP模型（或者说该类型的模型，包括AlBert，XLNet等）。对于志在NLP的同学，有必要对其原理和代码都进行比较深入的了解。废话不多说，进入正题。
PS：1.这里的代码有些参数传入是阉割过的，而且代码版本也是比较老版的，但更容易理解，更详细的还是参考：https://huggingface.co/transformers/
2.关键的注解都在代码的注释里。

主要代码

1.主函数入口

class BertModel(nn.Module):
    def __init__(self, config: BertConfig):
        super(BertModel, self).__init__()
        self.embeddings = BERTEmbeddings(config)
        self.encoder = BERTEncoder(config)
        self.pooler = BERTPooler(config)

    def forward(self, input_ids, token_type_ids=None, attention_mask=None):
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids)
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)
            
        # attention_mask的维度应保持和多头的hidden_states一致
        #!!!个人感觉这里extended_attention_mask 还应该扩展一下，感觉这个维度不太对!
        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
        extended_attention_mask = extended_attention_mask.float()
        # mask部分token的权重直接给-10000，使其在self-att的时候基本不起作用。
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
        
        #根据input_ids, token_type_ids以及position_ids来确定初始embeddings
        embedding_output = self.embeddings(input_ids, token_type_ids)
        #核心层，由以多层self_attention为主的神经网络构成
        all_encoder_layers = self.encoder(embedding_output, extended_attention_mask)
        #最后一层隐藏层
        sequence_output = all_encoder_layers[-1]
        #取出最后一层隐藏层的[cls]的表征，经过网络层(self.pooler)后得到pooled_output
        pooled_output = self.pooler(sequence_output)
        return all_encoder_layers, pooled_output

大致讲一下吧：
一般必传的三个参数input_idx，token_type_ids，attention_mask。
维度均为（batch_size, max_sent_length）

• input_idx就是每个token对应的idx，对应关系在预训练模型文件集的vocab.txt里
• token_type_ids有两种取值(0对应sentenceA,1对应sentenceB）该tensor会在self.embeddings的时候和input_iput生成的embedding相加生成初始的embeddings。
• attention_mask有两种取值（1代表非mask词，0代表mask掉的词）一般来说在finetune阶段，我们会把padding部分都设成mask掉的词。

其他基本也都注释了。

2.BertEmbedding层

class BERTEmbeddings(nn.Module):
    def __init__(self, config):
        super(BERTEmbeddings, self).__init__()
        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
        self.LayerNorm = BERTLayerNorm(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, input_ids, token_type_ids=None):
    	#根据每个token的位置生成position_ids，很直观
        seq_length = input_ids.size(1)
        position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device)
        position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
        
        if token_type_ids is None:
            token_type_ids = torch.zeros_like(input_ids)
    
		#这三个embeddings相信大家可以参见下图就一目了然了
        words_embeddings = self.word_embeddings(input_ids)
        position_embeddings = self.position_embeddings(position_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

        embeddings = words_embeddings + position_embeddings + token_type_embeddings
        #最后过一个layerNorm和dropout层
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings

3.BertEnocder层

class BERTEncoder(nn.Module):
    def __init__(self, config):
        super(BERTEncoder, self).__init__()
        layer = BERTLayer(config)
        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.num_hidden_layers)])    

    def forward(self, hidden_states, attention_mask):
        all_encoder_layers = []
        for layer_module in self.layer:
            hidden_states = layer_module(hidden_states, attention_mask)
            all_encoder_layers.append(hidden_states)
        return all_encoder_layers
        
class BERTLayer(nn.Module):
    def __init__(self, config):
        super(BERTLayer, self).__init__()
        self.attention = BERTAttention(config)
        self.intermediate = BERTIntermediate(config)
        self.output = BERTOutput(config)

    def forward(self, hidden_states, attention_mask):
        attention_output = self.attention(hidden_states, attention_mask)
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        return layer_output

BertEncoder层实质上就是由多个（num_hidden_layers）BertLayer层堆叠而成。
而BertLayer又由attention，intermediate和output三部分组成，下面分别来看。

3.1BertTAttention

重头戏开始！详见注释，看完你会发现很简单。

class BERTAttention(nn.Module):
    def __init__(self, config):
        super(BERTAttention, self).__init__()
        self.self = BERTSelfAttention(config)
        self.output = BERTSelfOutput(config)

    def forward(self, input_tensor, attention_mask):
        self_output = self.self(input_tensor, attention_mask)
        attention_output = self.output(self_output, input_tensor)
        return attention_output
        
class BERTSelfAttention(nn.Module):
    def __init__(self, config):
        super(BERTSelfAttention, self).__init__()
        if config.hidden_size % config.num_attention_heads != 0:
            raise ValueError(
                "The hidden size (%d) is not a multiple of the number of attention"
                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
        self.num_attention_heads = config.num_attention_heads#多头self_attention
        self.attention_head_size = int(config.hidden_size /config.num_attention_heads)#每个头的维度，一般是768/12=64
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.value = nn.Linear(config.hidden_size, self.all_head_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(self, hidden_states, attention_mask):
    	#经典生成QKV
    	#(batch_size, max_sen_length, hidden_size)->(batch_size, max_sen_length, hidden_size)
    	#(8, 512, 768)->(8, 512, 768)
        mixed_query_layer = self.query(hidden_states)
        mixed_key_layer = self.key(hidden_states)
        mixed_value_layer = self.value(hidden_states)
		#改变维度，形成多头，记住是在生成QKV之后才干的事
		#(batch_size, max_sen_length, hidden_size)->(batch_size, num_attention_heads, max_sen_length, attention_head_size)
		#(8, 512, 768)->(8, 12, 512, 64)
        query_layer = self.transpose_for_scores(mixed_query_layer)
        key_layer = self.transpose_for_scores(mixed_key_layer)
        value_layer = self.transpose_for_scores(mixed_value_layer)

        #QK tensor相乘，只对最后两维做矩阵乘法
        #(batch_size, num_attention_heads, max_sen_length, attention_head_size)*(batch_size, num_attention_heads, attention_head_size, max_sen_length)->(batch_size, num_attention_heads, max_sen_length, max_sen_length)
        #(8, 12, 512, 64)*(8, 12, 64, 512)->(8, 12, 512, 512)
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
        #除以维度的开方，这是为了使QV的结果方差变为1，使得sortmax后不会发生梯度消失。
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        #之前传的attention_mask在此刻发挥它的作用了！把mask掉的词的“权重”变成-10000，softmax后就基本等于0。
        attention_scores = attention_scores + attention_mask

        # softmax加一个dropout, 这也没啥好说的
        attention_probs = nn.Softmax(dim=-1)(attention_scores)
        attention_probs = self.dropout(attention_probs)
		# 最后再和V相乘,至此就完成了经典的softmax(QK/sqrt(dk))*V的操作！
		#(8, 12, 512, 512)*(8, 12, 512, 64)->(8, 12, 512, 64)
        context_layer = torch.matmul(attention_probs, value_layer)
        #之后就是把维度进行还原
        #(8, 12, 512, 64)->(8, 512,12 ,64)->(8, 512, 768)
        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)
        return context_layer

class BERTSelfOutput(nn.Module):
    def __init__(self, config):
        super(BERTSelfOutput, self).__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = BERTLayerNorm(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
    	#很平淡的全连接层加上dropout和LayerNorm
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states

3.2 BertIntermediate&& BertOutput

class BERTIntermediate(nn.Module):
    def __init__(self, config):
        super(BERTIntermediate, self).__init__()
        #之前一直不清楚这个intermediate_size是干嘛的，原来是self_attention后还跟了BERTIntermediate和BERTOutput2个全连接层。
        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
        self.intermediate_act_fn = gelu

    def forward(self, hidden_states):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        return hidden_states
        
class BERTOutput(nn.Module):
    def __init__(self, config):
        super(BERTOutput, self).__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.LayerNorm = BERTLayerNorm(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states

！！！这个和我之前看的transformers的残差连接层差别还挺大的，所以并不完全和transformers的encoder部分结构一致。
这之后就是主函数里的几步骤收尾工作了，这里也不再赘述。

4.补充

下面补充一下中途涉及到的相关类（LayerNorm）的代码

4.1 BertLayerNorm

class BERTLayerNorm(nn.Module):
    def __init__(self, config, variance_epsilon=1e-12):
        """Construct a layernorm module in the TF style (epsilon inside the square root).
        """
        super(BERTLayerNorm, self).__init__()
        self.gamma = nn.Parameter(torch.ones(config.hidden_size))
        self.beta = nn.Parameter(torch.zeros(config.hidden_size))
        self.variance_epsilon = variance_epsilon

    def forward(self, x):
        u = x.mean(-1, keepdim=True)
        s = (x - u).pow(2).mean(-1, keepdim=True)
        x = (x - u) / torch.sqrt(s + self.variance_epsilon)
        return self.gamma * x + self.beta

1.batchNorm是对多个样本进行标准化，而layerNorm是对单样本标准化。
2.BertLayerNorm除了标准化以外还加上了gamma和beta的变化。

4.2 BertPooler

class BERTPooler(nn.Module):
    def __init__(self, config):
        super(BERTPooler, self).__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.activation = nn.Tanh()
        
    def forward(self, hidden_states):
    	#取出[cls]后过一个全连接层和激活函数。
        first_token_tensor = hidden_states[:, 0]
        pooled_output = self.dense(first_token_tensor)
        pooled_output = self.activation(pooled_output)
        return pooled_output

上文也提到了，BertPooler就是专门为[cls]设计的

4.3 gelu

def gelu(x):
    """Implementation of the gelu activation function.
    """
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))

4.4 transpose_for_scores

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

总结

到此基本就结束了，整体流程看下来其实很快，关键是理清里面每一步的维度的变换和几个核心的类就行。希望能对大家有所帮助。
代码参考来自于：https://github.com/DA-southampton/Read_Bert_Code