pytorch版bert modeling_bert代码解析

huggingface源码地址 https://github.com/huggingface/transformers

由于源码一直在更新迭代，我的版本可能不太新了，不过大致意思差不多

modeling_bert.py

预训练模型的下载地址，如果加载时 参数设置没用下好的模型地址，则会自动从这些地址上下载

BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-pytorch_model.bin",
    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-pytorch_model.bin",
    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-pytorch_model.bin",
    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-pytorch_model.bin",
    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-pytorch_model.bin",
    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-pytorch_model.bin",
    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-pytorch_model.bin",
    'bert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-cased-pytorch_model.bin",
    'bert-large-uncased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-pytorch_model.bin",
    'bert-large-cased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-pytorch_model.bin",
    'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-pytorch_model.bin",
    'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-pytorch_model.bin",
    'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
}

json格式存储的参数键值

BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-config.json",
    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-config.json",
    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-config.json",
    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-config.json",
    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-config.json",
    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-config.json",
    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-config.json",
    'bert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-cased-config.json",
    'bert-large-uncased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-config.json",
    'bert-large-cased-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-config.json",
    'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-config.json",
    'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json",
    'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
}

加载tf权重的函数：（主要是从checkpoints中获取weight和bias）

def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
    """ Load tf checkpoints in a pytorch model.
    """
    try:
        import re
        import numpy as np
        import tensorflow as tf
    except ImportError: #import tensorflow失败时
        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
            "https://www.tensorflow.org/install/ for installation instructions.")
        raise
    tf_path = os.path.abspath(tf_checkpoint_path) #取绝对路径
    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path) #Returns list of all variables in the checkpoint.
    names = []
    arrays = []
    for name, shape in init_vars:
        logger.info("Loading TF weight {} with shape {}".format(name, shape))
        array = tf.train.load_variable(tf_path, name)
        names.append(name)
        arrays.append(array)
 
    for name, array in zip(names, arrays):#一一对应的元祖
        name = name.split('/')
        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
        # which are not required for using pretrained model
        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
            logger.info("Skipping {}".format("/".join(name)))
            continue
        pointer = model
        for m_name in name:
            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
                l = re.split(r'_(\d+)', m_name)
            else:
                l = [m_name]
            if l[0] == 'kernel' or l[0] == 'gamma':
                pointer = getattr(pointer, 'weight')
            elif l[0] == 'output_bias' or l[0] == 'beta':
                pointer = getattr(pointer, 'bias')
            elif l[0] == 'output_weights':
                pointer = getattr(pointer, 'weight')
            elif l[0] == 'squad':
                pointer = getattr(pointer, 'classifier')
            else:
                try:
                    pointer = getattr(pointer, l[0])
                except AttributeError:
                    logger.info("Skipping {}".format("/".join(name)))
                    continue
            if len(l) >= 2:
                num = int(l[1])
                pointer = pointer[num]
        if m_name[-11:] == '_embeddings':
            pointer = getattr(pointer, 'weight')
        elif m_name == 'kernel':
            array = np.transpose(array)
        try:
            assert pointer.shape == array.shape
        except AssertionError as e:
            e.args += (pointer.shape, array.shape)
            raise
        logger.info("Initialize PyTorch weight {}".format(name))
        pointer.data = torch.from_numpy(array)
    return model

gelu激活函数  gelu(高斯误差线性单元)

激活函数的作用：给网络模型加非线性因子，让wx+b这样的线性变换加非线性

def gelu(x):
    """Implementation of the gelu activation function.
        For information: OpenAI GPT's gelu is slightly different (and gives slightly different results):
        0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
        Also see https://arxiv.org/abs/1606.08415
    """
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))

swish 激活函数 （不太清楚啥时候用这个激活函数比较好）

def swish(x):
    return x * torch.sigmoid(x)

BertConfig bert参数类(存BERT模型配置参数的类)

参数 ：vocab_size_or_config_json_file=30522: `BertModel`的`inputs_ids`的词典大小(如果是个路径str就直接读里面的参数)
            hidden_size=768: 嵌入层和池化层的大小
            num_hidden_layers=12: Transformer encoder.的隐藏层数
            num_attention_heads=12: 每个注意力层的注意力头数
            intermediate_size=3072: The size of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
            hidden_act='gelu': 非线性激活函数in the encoder and pooler. 支持 "gelu", "relu" and "swish" 
            hidden_dropout_prob=0.1: 全连接层的dropout率 in the embeddings, encoder, and pooler.
            attention_probs_dropout_prob=0.1: The dropout ratio for the attention probabilities.
            max_position_embeddings=512: 最大嵌入长度(e.g., 512 or 1024 or 2048).
            type_vocab_size=2: `token_type_ids` 的词典类型大小（非0即1）
            initializer_range=0.02: truncated_normal_initializer的标准差用于初始化所有的权重矩阵
            layer_norm_eps=1e-12: 层归一化（layerNorm）的epsilon值

class BertConfig(PretrainedConfig):
    r"""
        :class:`~pytorch_transformers.BertConfig` is the configuration class to store the configuration of a `BertModel`.
        Arguments:
            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `BertModel`.
            hidden_size: Size of the encoder layers and the pooler layer.
            num_hidden_layers: Number of hidden layers in the Transformer encoder.
            num_attention_heads: Number of attention heads for each attention layer in
                the Transformer encoder.
            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
                layer in the Transformer encoder.
            hidden_act: The non-linear activation function (function or string) in the
                encoder and pooler. If string, "gelu", "relu" and "swish" are supported.
            hidden_dropout_prob: The dropout probabilitiy for all fully connected
                layers in the embeddings, encoder, and pooler.
            attention_probs_dropout_prob: The dropout ratio for the attention
                probabilities.
            max_position_embeddings: The maximum sequence length that this model might
                ever be used with. Typically set this to something large just in case
                (e.g., 512 or 1024 or 2048).
            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
                `BertModel`.
            initializer_range: The sttdev of the truncated_normal_initializer for
                initializing all weight matrices.
            layer_norm_eps: The epsilon used by LayerNorm.
    """
    pretrained_config_archive_map = BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
    def __init__(self,
                 vocab_size_or_config_json_file=30522,
                 hidden_size=768,
                 num_hidden_layers=12,
                 num_attention_heads=12,
                 intermediate_size=3072,
                 hidden_act="gelu",
                 hidden_dropout_prob=0.1,
                 attention_probs_dropout_prob=0.1,
                 max_position_embeddings=512,
                 type_vocab_size=2,
                 initializer_range=0.02,
                 layer_norm_eps=1e-12,
                 **kwargs):
        super(BertConfig, self).__init__(**kwargs)
        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
                        and isinstance(vocab_size_or_config_json_file, unicode)):
            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
                json_config = json.loads(reader.read())
            for key, value in json_config.items():
                self.__dict__[key] = value
        elif isinstance(vocab_size_or_config_json_file, int):
            self.vocab_size = vocab_size_or_config_json_file
            self.hidden_size = hidden_size
            self.num_hidden_layers = num_hidden_layers
            self.num_attention_heads = num_attention_heads
            self.hidden_act = hidden_act
            self.intermediate_size = intermediate_size
            self.hidden_dropout_prob = hidden_dropout_prob
            self.attention_probs_dropout_prob = attention_probs_dropout_prob
            self.max_position_embeddings = max_position_embeddings
            self.type_vocab_size = type_vocab_size
            self.initializer_range = initializer_range
            self.layer_norm_eps = layer_norm_eps
        else:
            raise ValueError("First argument must be either a vocabulary size (int)"
                             " or the path to a pretrained model config file (str)")

apex貌似是用来加速的

try:
    from apex.normalization.fused_layer_norm import FusedLayerNorm as BertLayerNorm
except (ImportError, AttributeError) as e:
    logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")
    BertLayerNorm = torch.nn.LayerNorm

做embedding嵌入的处理 

nn.Embedding 做lookup操作把对应标号的向量表示揪出来，分三个处理 ，

1. word_embedding相当于input_id就是字向量表示

2. position_embeddings位置嵌入，如果没处理就用0,1,2,3....这样的表明字所对应的位置

3.token_type_embeddings NSP操作 ，用0或1区分是上一句还是下一句 类似1111100000这样，现在很多论文说这个没用

三个处理完相加得到embedding并做层归一化和dropout

class BertEmbeddings(nn.Module):
    """Construct the embeddings from word, position and token_type embeddings.
    """
    def __init__(self, config):
        super(BertEmbeddings, self).__init__()
        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=0)
        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 is not snake-cased to stick with TensorFlow model variable name and be able to load
        # any TensorFlow checkpoint file
        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
    def forward(self, input_ids, token_type_ids=None, position_ids=None):
        seq_length = input_ids.size(1) #第一维是max_len
        if position_ids is None:
            position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device) #0到511
            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)#扩成input_ids的shape
        if token_type_ids is None: #如果这个参数是None，则不做NSP下个句子预测处理，句子类型全用0表示
            token_type_ids = torch.zeros_like(input_ids)
 
        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
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings

BertSelfAttention自注意力机制

embedding完维度依然 [batch,max_len,hidden_size]

linear层后依然不变  transpose时变成[batch,max_len,num_heads,head_size] permute后[batch,num_heads,max_len,head_size]

matmul 后 [batch,num_heads,max_len,max_len] 除以根号64也就是8

然后用attention_scores + attention_mask（论文中的MLM操作）

 torch.matmul(attention_probs, value_layer) 每个字V乘上对应的0-1的分数  维度[batch,num_heads,max_len,max_len] *[batch,num_heads,max_len,head_size]变成维度[batch,num_heads,max_len,head_size]

permute后 [batch,max_len,num_heads,head_size]

context_layer.size()[:-2] + (self.all_head_size,)后[batch,max_len,num_heads,all_head_size]

class BertSelfAttention(nn.Module):
    def __init__(self, config):
        super(BertSelfAttention, self).__init__()
        if config.hidden_size % config.num_attention_heads != 0:#先判断注意力头数能否被隐藏层整除
            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.output_attentions = config.output_attentions #用来控制输出加不加attention_probs
        self.num_attention_heads = config.num_attention_heads #注意力头数
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)#每个头的维度大小 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, head_mask=None):
        mixed_query_layer = self.query(hidden_states)
        mixed_key_layer = self.key(hidden_states)
        mixed_value_layer = self.value(hidden_states)
 
        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)
 
        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
        attention_scores = attention_scores + attention_mask
 
        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores) #在最外面一维加起来和为1
 
        # 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 = self.dropout(attention_probs)
 
        # Mask heads if we want to
        if head_mask is not None:
            attention_probs = attention_probs * head_mask
 
        context_layer = torch.matmul(attention_probs, value_layer)
 
        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)
 
        outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
        return outputs