BERT数据处理，模型，预训练

代码来自李沐老师《动手学pytorch》
在数据处理时，首先执行以下代码
def load_data_wiki(batch_size, max_len):
    """加载WikiText-2数据集"""
    num_workers = d2l.get_dataloader_workers()
    data_dir = d2l.download_extract('wikitext-2', 'wikitext-2')
    以上两句代码，不再说明
    paragraphs = _read_wiki(data_dir)
    train_set = _WikiTextDataset(paragraphs, max_len)
    train_iter = torch.utils.data.DataLoader(train_set, batch_size,
                                         shuffle=True)
    return train_iter, train_set.vocab

d2l.DATA_HUB['wikitext-2'] = (
    'https://s3.amazonaws.com/research.metamind.io/wikitext/'
    'wikitext-2-v1.zip', '3c914d17d80b1459be871a5039ac23e752a53cbe')

#@save
def _read_wiki(data_dir):
    file_name = os.path.join(data_dir, 'wiki.train.tokens')
    with open(file_name, 'r',encoding='utf-8') as f:
        lines = f.readlines()
    # 大写字母转换为小写字母 ，每行文本中包含两个句子，才进行处理，否则舍去文本
    paragraphs = [line.strip().lower().split(' . ')
                  for line in lines if len(line.split(' . ')) >= 2]
    random.shuffle(paragraphs)
    return paragraphs

首先读取文本，每个文本必须包含两个以上句子（为了第二个预训练任务：判断两个句子，是否连续）。paragraphs 其中一部分结果如下所示

文本中包含了三个句子，每个’‘里面，代表一个句子
['common starlings are trapped for food in some mediterranean countries'
, 'the meat is tough and of low quality , so it is <unk> or made into <unk>'
, 'one recipe said it should be <unk> " until tender , however long that may be "'
, 'even when correctly prepared , it may still be seen as an acquired taste .']

class _WikiTextDataset(torch.utils.data.Dataset):
    def __init__(self, paragraphs, max_len):
        '''
        每一个paragraph就是上面的包含多个句子的列表，将其进行分词处理。下面是一个分词的例子
        [['common', 'starlings', 'are', 'trapped', 'for', 'food', 'in', 'some', 'mediterranean', 'countries']
        , ['the', 'meat', 'is', 'tough', 'and', 'of', 'low', 'quality', ',', 'so', 'it', 'is', '<unk>', 'or', 'made', 'into', '<unk>'], ['one', 'recipe', 'said', 'it', 'should', 'be', '<unk>', '"', 'until', 'tender', ',', 'however', 'long', 'that', 'may', 'be', '"']
        , ['even', 'when', 'correctly', 'prepared', ',', 'it', 'may', 'still', 'be', 'seen', 'as', 'an', 'acquired', 'taste', '.']]
        '''
        paragraphs = [d2l.tokenize(
            paragraph, token='word') for paragraph in paragraphs]
        #将词提取处理，保存
        sentences = [sentence for paragraph in paragraphs
                     for sentence in paragraph]
        #形成一个词典，min_freq为词最少出现的次数，少于5次，则不保存进词典中
        self.vocab = d2l.Vocab(sentences, min_freq=5, reserved_tokens=[
            '<pad>', '<mask>', '<cls>', '<sep>'])
        # 获取下一句子预测任务的数据
        examples = []
        for paragraph in paragraphs:
            examples.extend(_get_nsp_data_from_paragraph(
                paragraph, paragraphs, self.vocab, max_len))
            '''
def _get_nsp_data_from_paragraph(paragraph,paragraphs,vocab,max_len):
    nsp_data_from_paragraph=[]
    for i in range(len(paragraph)-1):
    
_get_next_sentence函数传入的是相邻的句子a,b。函数中b会有一定概率替换为其他的句子

        tokens_a, tokens_b, is_next = _get_next_sentence(
            paragraph[i], paragraph[i + 1], paragraphs)
            
句子长度大于bert限制的长度，则舍去。
        if len(tokens_a)+len(tokens_b)+3>max_len:
            continue
            
        #加上<cls>和<sep>,segments用于区token在哪个句子中
        
        tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
        nsp_data_from_paragraph.append((tokens, segments, is_next))
    return nsp_data_from_paragraph
  
  token和segments的例子： True表示两个句子相邻，False表示b被随机替换，a,b不相邻。
            (['<cls>', 'mushrooms', 'grow', '<unk>', 'or', 'in', '"', '<unk>', 'groups', '"', 'in', 'late', 'summer', 'and', 'throughout', 
            'autumn', ',', 'though', 'it', 'is', 'not', 'commonly', 'encountered', 'species', '<sep>', 'it',
             'can', 'be', 'found', 'in', 'europe', ',', 'asia', 'and', 'north', 'america', '.', '<sep>'], 
             [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
              1, 1, 1, 1, 1], True),
            '''


        # 获取遮蔽语言模型任务的数据
        '''
     在这里我们会将句子中单词，替换为在词典中的索引。13意思为，句子的第13个词，进行了处理，可能不变，可能替换为其他词，可能替换为mask。在这里这个词没有替换。0与1区分两个句子，False代表两个句子不相邻。
        examples中的结果；
        ([3, 2510, 31, 337, 9, 0, 6, 6891, 8, 11621, 6, 21, 11, 60, 3405, 14, 1542, 9546, 4, 2524,
         21, 185, 4421, 649, 38, 277, 2872, 13233, 4], [13], [60], 
         [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
         False)
        '''
        examples = [(_get_mlm_data_from_tokens(tokens, self.vocab)
                      + (segments, is_next))
                     for tokens, segments, is_next in examples]
                     
        #_pad_bert_inputs对数据进行填充,all_mlm_weights中1为需要预测，0为填充
   #    all_mlm_weights= tensor([1., 0., 0., 0., 0., 0., 0., 0., 0., 0.]
        (self.all_token_ids, self.all_segments, self.valid_lens,
         self.all_pred_positions, self.all_mlm_weights,
         self.all_mlm_labels, self.nsp_labels) = _pad_bert_inputs(
            examples, max_len, self.vocab)

    def __getitem__(self, idx):
        return (self.all_token_ids[idx], self.all_segments[idx],
                self.valid_lens[idx], self.all_pred_positions[idx],
                self.all_mlm_weights[idx], self.all_mlm_labels[idx],
                self.nsp_labels[idx])

    def __len__(self):
        return len(self.all_token_ids)

上述已经将数据处理完，最后看一下处理后的例子：

将原来的句子列表填充1，一直到到大小为64
tensor([[    3,     5,     0, 18306,    23,    11,  2659,   156,  5779,   382,
          1296,   110,   158,    22,     5,  1771,   496,     0,  3398,     2,
             5,  3496,   110,  5038,   179,     4,    16,    11, 19837,     6,
            58,    13,     5,   685,     7,    66,   156,     0,  3063,    77,
          3842,    19,     4,     1,     1,     1,     1,     1,     1,     1,
             1,     1,     1,     1,     1,     1,     1,     1,     1,     1,
             1,     1,     1,     1]])
segments用于区分两个句子，0为第一个句子中的词，1为第二个句子中的词，后面的0为填充
tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]])
valid_lens表示句子列表的有效长度
tensor([43.])
pred_positions需要预测的位置，0为填充
tensor([[19,  0,  0,  0,  0,  0,  0,  0,  0,  0]])
mlm_weights需要预测多少个词，0为填充
tensor([[1., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])
预测位置的真实标签，0为填充
tensor([[22,  0,  0,  0,  0,  0,  0,  0,  0,  0]])
两句话是否相邻
tensor([0])

随后就是把处理好的数据，送入bert中。在 BERTEncoder 中，执行如下代码：

 def forward(self, tokens, segments, valid_lens):
        # Shape of `X` remains unchanged in the following code snippet:
        # (batch size, max sequence length, `num_hiddens`)
      #  将token和segment分别进行embedding，
        X = self.token_embedding(tokens) + self.segment_embedding(segments)
      #加入位置编码
        X = X + self.pos_embedding.data[:, :X.shape[1], :]
        for blk in self.blks:
            X = blk(X, valid_lens)
        return X

将编码完后的数据，进行多头注意力和残差化

    def forward(self, X, valid_lens):
        Y = self.addnorm1(X, self.attention(X, X, X, valid_lens))
        return self.addnorm2(Y, self.ffn(Y))

将结果返回到如下代码中：其中encoded_X .shape=torch.Size([1, 64, 128]),1代表批次大小为1，我们设置的每个批次只有行文本，每行文本由64个词组成，bert提取128维的向量来表示每个词。随后进行两个任务，一个是预测被掩盖的单词，另一个为判断两个句子是否为相邻。

    def forward(self, tokens, segments, valid_lens=None, pred_positions=None):
        encoded_X = self.encoder(tokens, segments, valid_lens)
        if pred_positions is not None:
            mlm_Y_hat = self.mlm(encoded_X, pred_positions)
        else:
            mlm_Y_hat = None
        # The hidden layer of the MLP classifier for next sentence prediction.
        # 0 is the index of the '<cls>' token
        nsp_Y_hat = self.nsp(self.hidden(encoded_X[:, 0, :]))
        return encoded_X, mlm_Y_hat, nsp_Y_hat

第一个任务为预测被mask的单词：

'''
例如：batch为1，X为1*64*128,其中num_pred_positions =10，batch_idx 会重复为[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]，pred_positions为[ 3,  6, 10, 12, 15, 20,  0,  0,  0,  0]，X[batch_idx, pred_positions]会将需要预测的向量取出。然后reshape为1*10*128的矩阵。最后连接一个mlp，经过规范化后接nn.Linear(num_hiddens, vocab_size))，会生成再vocab上的预测

'''
 def forward(self, X, pred_positions):
        num_pred_positions = pred_positions.shape[1]
        pred_positions = pred_positions.reshape(-1)
        batch_size = X.shape[0]
        batch_idx = torch.arange(0, batch_size)
        # Suppose that `batch_size` = 2, `num_pred_positions` = 3, then
        # `batch_idx` is `torch.tensor([0, 0, 0, 1, 1, 1])`
        batch_idx = torch.repeat_interleave(batch_idx, num_pred_positions)
        masked_X = X[batch_idx, pred_positions]
        masked_X = masked_X.reshape((batch_size, num_pred_positions, -1))
        mlm_Y_hat = self.mlp(masked_X)
        return mlm_Y_hat

结束后，会返回到上层的代码中：

def forward(self, tokens, segments, valid_lens=None, pred_positions=None):
        encoded_X = self.encoder(tokens, segments, valid_lens)
        if pred_positions is not None:
            mlm_Y_hat = self.mlm(encoded_X, pred_positions)
        else:
            mlm_Y_hat = None
        # The hidden layer of the MLP classifier for next sentence prediction.
        # 0 is the index of the '<cls>' token
        判断句子是否连续，将<cls>的向量，放入mlp中，接一个nn.Linear(num_inputs, 2)，最后变成一个二分类问题。
        nsp_Y_hat = self.nsp(self.hidden(encoded_X[:, 0, :]))
        return encoded_X, mlm_Y_hat, nsp_Y_hat

后面就是计算损失：

将mlm_Y_hat进行reshap，与mlm_Y求loss，最后需要乘mlm_weights_X，将填充的无用数据进行去除。
 mlm_l = loss(mlm_Y_hat.reshape(-1, vocab_size), mlm_Y.reshape(-1)) * mlm_weights_X.reshape(-1, 1)
 取平均loss
 mlm_l = mlm_l.sum() / (mlm_weights_X.sum() + 1e-8)
 nsp_l = loss(nsp_Y_hat, nsp_y)
 l = mlm_l + nsp_l