BERT预训练

BERT

NLP里的迁移学习

• 使用预训练好的模型来抽取词、句子的特征
  • 例如 word2vec 或语言模型
• 不更新预训练好的模型
• 需要构建新的网络来抓取新任务需要的信息
  • Word2vec 忽略了时序信息，语言模型只看了一个方向

BERT 动机

BERT 架构

• 只有编码器的 Transformer
• 两个版本：
  • Base: #blocks = 12, hidden size = 768, #heads = 12, #parameters = 110M
  • Large: #blocks = 24, hidden size = 1024, #heads = 16, #parameter = 340M
• 在大规模数据上训练 > 3B 词

对输入的修改

• 每个样本是一个句子对
• 加入额外的片段嵌入
• 位置编码可学习

预训练任务1：带掩码的语言模型

• Transfomer的编码器是双向，标准语言模型要求单向
• 带掩码的语言模型每次随机（15%概率）将一些词元换成 < mask >
• 因为微调任务中不出现< mask >
  • 80%概率下，将选中的词元变成< mask >
  • 10%概率下换成一个随机词元
  • 10%概率下保持原有的词元

预训练任务2：下一句子预测

• 预测一个句子对中两个句子是不是相邻
• 训练样本中：
  • 50%概率选择相邻句子对：< cls > this movie is great < sep > i like it < sep >
  • 50%概率选择随机句子对：< cls > this movie is great < sep > hello world < sep >
• 将< cls >对应的输出放到一个全连接层来预测

总结

• BERT针对微调设计
• 基于Transformer的编码器做了如下修改
  • 模型更大，训练数据更多
  • 输入句子对，片段嵌入，可学习的位置编码
• 训练时使用两个任务：
  • 带掩码的语言模型
  • 下一个句子预测

BERT 代码

Bidirectional Encoder Representations from Transformers (BERT)
首先导入必要的环境

import torch
from torch import nn
from d2l import torch as d2l

输入表示

在自然语言处理中，有些任务（如情感分析）以单个文本作为输入，而有些任务（如自然语言推断）以一对文本序列作为输入。BERT输入序列明确地表示单个文本和文本对。当输入为单个文本时，BERT输入序列是特殊类别词元“<cls>”、文本序列的标记、以及特殊分隔词元“<sep>”的连结。当输入为文本对时，BERT输入序列是“<cls>”、第一个文本序列的标记、“<sep>”、第二个文本序列标记、以及“<sep>”的连结。下面将始终如一地将术语“BERT输入序列”与其他类型的“序列”区分开来。例如，一个BERT输入序列可以包括一个文本序列或两个文本序列。

为了区分文本对，根据输入序列学到的片段嵌入${\mathbf{e}}_A$和${\mathbf{e}}_B$分别被添加到第一序列和第二序列的词元嵌入中。对于单文本输入，仅使用${\mathbf{e}}_A$。下面将一个句子或两个句子作为输入，然后返回BERT输入序列的标记及其相应的片段索引。

def get_tokens_and_segments(tokens_a, tokens_b=None):
    """获取输入序列的词元及其片段索引"""
    tokens = ['<cls>'] + tokens_a + ['<sep>']
    # 0和1分别标记片段A和B
    segments = [0] * (len(tokens_a) + 2)
    if tokens_b is not None:
        tokens += tokens_b + ['<sep>']
        segments += [1] * (len(tokens_b) + 1)
    return tokens, segments

BERT 编码器类

class BERTEncoder(nn.Module):
    """BERT编码器"""

    def __init__(self, vocab_size, num_hiddens, norm_shape, ffn_num_input,
                 ffn_num_hiddens, num_heads, num_layers, dropout,
                 max_len=1000, key_size=768, query_size=768, value_size=768,
                 **kwargs):
        """
        初始化BERT编码器

        参数：
        vocab_size: 词表大小
        num_hiddens: 隐藏单元数（嵌入维度）
        norm_shape: 层归一化的形状
        ffn_num_input: 前馈网络的输入大小
        ffn_num_hiddens: 前馈网络的隐藏层大小
        num_heads: 多头注意力的头数
        num_layers: 编码器块的层数
        dropout: Dropout概率
        max_len: 最大序列长度
        key_size: 注意力机制中键的维度
        query_size: 注意力机制中查询的维度
        value_size: 注意力机制中值的维度
        """
        super(BERTEncoder, self).__init__(**kwargs)
        
        # 词元嵌入层，将词元索引映射到嵌入向量
        self.token_embedding = nn.Embedding(vocab_size, num_hiddens)
        
        # 片段嵌入层，用于区分句子A和句子B
        self.segment_embedding = nn.Embedding(2, num_hiddens)
        
        # 编码器块的堆叠，使用d2l提供的EncoderBlock实现
        self.blks = nn.Sequential()
        for i in range(num_layers):
            self.blks.add_module(f"{i}", d2l.EncoderBlock(
                key_size, query_size, value_size, num_hiddens, norm_shape,
                ffn_num_input, ffn_num_hiddens, num_heads, dropout, True))
        
        # 可学习的位置嵌入参数，用于表示序列中每个位置的信息
        # 形状为 (1, max_len, num_hiddens)，其中1表示批量维度
        self.pos_embedding = nn.Parameter(torch.randn(1, max_len, num_hiddens))

    def forward(self, tokens, segments, valid_lens):
        """
        前向传播

        参数：
        tokens: 输入的词元索引，形状为 (批量大小, 最大序列长度)
        segments: 输入的片段索引，形状为 (批量大小, 最大序列长度)
        valid_lens: 有效长度，用于掩蔽填充部分，形状为 (批量大小,)

        返回：
        编码后的表示，形状为 (批量大小, 最大序列长度, num_hiddens)
        """
        # 词元嵌入和片段嵌入相加，形状为 (批量大小, 最大序列长度, num_hiddens)
        X = self.token_embedding(tokens) + self.segment_embedding(segments)
        
        # 加上位置嵌入，位置嵌入的形状为 (1, 最大序列长度, num_hiddens)
        # 通过切片选择与输入序列长度匹配的部分
        X = X + self.pos_embedding.data[:, :X.shape[1], :]
        
        # 依次通过每个编码器块
        for blk in self.blks:
            X = blk(X, valid_lens)
        
        # 返回编码后的表示
        return X

假设词表大小为10000，喜爱卖弄演示BERTEncoder的前向推断，创建一个实例并初始化其参数。

vocab_size, num_hiddens, ffn_num_hiddens, num_heads = \
      10000, 768, 1024, 4
norm_shape, ffn_num_input, num_layers, dropout = [768], 768, 2, 0.2
encoder = BERTEncoder(vocab_size, num_hiddens, norm_shape, ffn_num_input,
                      ffn_num_hiddens, num_heads, num_layers, dropout)

tokens = torch.randint(0, vocab_size, (2, 8))
segments = torch.tensor([[0, 0, 0, 0, 1, 1, 1, 1], [0, 0, 0, 1, 1, 1, 1, 1]])
encoded_X = encoder(tokens, segments, None)
encoded_X.shape

# torch.Size([2, 8, 768])

遮蔽语言模型

class MaskLM(nn.Module):
    """BERT的掩蔽语言模型任务"""
    def __init__(self, vocab_size, num_hiddens, num_inputs=768, **kwargs):
        """
        初始化掩蔽语言模型（Masked Language Model）

        参数：
        vocab_size: 词表大小，用于预测的输出维度
        num_hiddens: 隐藏单元数，用于MLP的隐藏层
        num_inputs: 输入特征的维度，默认为768
        """
        super(MaskLM, self).__init__(**kwargs)
        # 定义一个多层感知机（MLP）用于预测掩蔽词元
        self.mlp = nn.Sequential(
            nn.Linear(num_inputs, num_hiddens),  # 全连接层，将输入映射到隐藏层
            nn.ReLU(),                          # 激活函数，使用ReLU
            nn.LayerNorm(num_hiddens),          # 层归一化，稳定训练
            nn.Linear(num_hiddens, vocab_size)  # 输出层，映射到词表大小的维度
        )

    def forward(self, X, pred_positions):
        """
        前向传播

        参数：
        X: 编码后的BERT表示，形状为 (batch_size, seq_len, num_inputs)
        pred_positions: 掩蔽词元的位置，形状为 (batch_size, num_pred_positions)

        返回：
        mlm_Y_hat: 掩蔽位置的预测结果，形状为 (batch_size, num_pred_positions, vocab_size)
        """
        # 获取每个序列中掩蔽词元的数量
        num_pred_positions = pred_positions.shape[1]
        
        # 将掩蔽位置展平为一维向量，形状为 (batch_size * num_pred_positions,)
        pred_positions = pred_positions.reshape(-1)
        
        # 获取批量大小
        batch_size = X.shape[0]
        
        # 创建一个重复的批量索引，用于从X中提取掩蔽位置的特征
        # 假设 batch_size=2，num_pred_positions=3
        # 那么 batch_idx 是 [0, 0, 0, 1, 1, 1]
        batch_idx = torch.arange(0, batch_size)
        batch_idx = torch.repeat_interleave(batch_idx, num_pred_positions)
        
        # 根据批量索引和掩蔽位置索引，从X中提取掩蔽位置的特征
        # masked_X 的形状为 (batch_size * num_pred_positions, num_inputs)
        masked_X = X[batch_idx, pred_positions]
        
        # 将提取的特征重新调整为 (batch_size, num_pred_positions, num_inputs)
        masked_X = masked_X.reshape((batch_size, num_pred_positions, -1))
        
        # 使用MLP对掩蔽位置的特征进行预测
        # mlm_Y_hat 的形状为 (batch_size, num_pred_positions, vocab_size)
        mlm_Y_hat = self.mlp(masked_X)
        
        return mlm_Y_hat

The forward inference of MaskLM

Next Sentence Prediction

#@save
class NextSentencePred(nn.Module):
    """BERT的下一句预测任务"""
    def __init__(self, num_inputs, **kwargs):
        super(NextSentencePred, self).__init__(**kwargs)
        self.output = nn.Linear(num_inputs, 2)

    def forward(self, X):
        # X的形状：(batchsize,num_hiddens)
        return self.output(X)

The forward inference of an NextSentencePred

整合代码

class BERTModel(nn.Module):
    """BERT模型"""
    def __init__(self, vocab_size, num_hiddens, norm_shape, ffn_num_input,
                 ffn_num_hiddens, num_heads, num_layers, dropout,
                 max_len=1000, key_size=768, query_size=768, value_size=768,
                 hid_in_features=768, mlm_in_features=768,
                 nsp_in_features=768):
        super(BERTModel, self).__init__()
        self.encoder = BERTEncoder(vocab_size, num_hiddens, norm_shape,
                    ffn_num_input, ffn_num_hiddens, num_heads, num_layers,
                    dropout, max_len=max_len, key_size=key_size,
                    query_size=query_size, value_size=value_size)
        self.hidden = nn.Sequential(nn.Linear(hid_in_features, num_hiddens),
                                    nn.Tanh())
        self.mlm = MaskLM(vocab_size, num_hiddens, mlm_in_features)
        self.nsp = NextSentencePred(nsp_in_features)

    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
        # 用于下一句预测的多层感知机分类器的隐藏层，0是“<cls>”标记的索引
        nsp_Y_hat = self.nsp(self.hidden(encoded_X[:, 0, :]))
        return encoded_X, mlm_Y_hat, nsp_Y_hat

BERT 预训练数据代码

The WikiText-2 dataset

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

def _read_wiki(data_dir):
    file_name = os.path.join(data_dir, 'wiki.train.tokens')
    with open(file_name, 'r') as f:
        lines = f.readlines()
    paragraphs = [line.strip().lower().split(' . ')
                  for line in lines if len(line.split(' . ')) >= 2]
    random.shuffle(paragraphs)
    return paragraphs

Generating the Next Sentence Prediction Task

def _get_next_sentence(sentence, next_sentence, paragraphs):
    if random.random() < 0.5:
        is_next = True
    else:
        # paragraphs是三重列表的嵌套
        next_sentence = random.choice(random.choice(paragraphs))
        is_next = False
    return sentence, next_sentence, is_next

def _get_nsp_data_from_paragraph(paragraph, paragraphs, vocab, max_len):
    nsp_data_from_paragraph = []
    for i in range(len(paragraph) - 1):
        tokens_a, tokens_b, is_next = _get_next_sentence(
            paragraph[i], paragraph[i + 1], paragraphs)
        # 考虑1个'<cls>'词元和2个'<sep>'词元
        if len(tokens_a) + len(tokens_b) + 3 > max_len:
            continue
        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

Generating the Masked Language Modeling Task

def _replace_mlm_tokens(tokens, candidate_pred_positions, num_mlm_preds,
                        vocab):
    # 为遮蔽语言模型的输入创建新的词元副本，其中输入可能包含替换的“<mask>”或随机词元
    mlm_input_tokens = [token for token in tokens]
    pred_positions_and_labels = []
    # 打乱后用于在遮蔽语言模型任务中获取15%的随机词元进行预测
    random.shuffle(candidate_pred_positions)
    for mlm_pred_position in candidate_pred_positions:
        if len(pred_positions_and_labels) >= num_mlm_preds:
            break
        masked_token = None
        # 80%的时间：将词替换为“<mask>”词元
        if random.random() < 0.8:
            masked_token = '<mask>'
        else:
            # 10%的时间：保持词不变
            if random.random() < 0.5:
                masked_token = tokens[mlm_pred_position]
            # 10%的时间：用随机词替换该词
            else:
                masked_token = random.choice(vocab.idx_to_token)
        mlm_input_tokens[mlm_pred_position] = masked_token
        pred_positions_and_labels.append(
            (mlm_pred_position, tokens[mlm_pred_position]))
    return mlm_input_tokens, pred_positions_and_labels

def _get_mlm_data_from_tokens(tokens, vocab):
    candidate_pred_positions = []
    # tokens是一个字符串列表
    for i, token in enumerate(tokens):
        # 在遮蔽语言模型任务中不会预测特殊词元
        if token in ['<cls>', '<sep>']:
            continue
        candidate_pred_positions.append(i)
    # 遮蔽语言模型任务中预测15%的随机词元
    num_mlm_preds = max(1, round(len(tokens) * 0.15))
    mlm_input_tokens, pred_positions_and_labels = _replace_mlm_tokens(
        tokens, candidate_pred_positions, num_mlm_preds, vocab)
    pred_positions_and_labels = sorted(pred_positions_and_labels,
                                       key=lambda x: x[0])
    pred_positions = [v[0] for v in pred_positions_and_labels]
    mlm_pred_labels = [v[1] for v in pred_positions_and_labels]
    return vocab[mlm_input_tokens], pred_positions, vocab[mlm_pred_labels]

Append the special “< mask >” tokens to the inputs

def _pad_bert_inputs(examples, max_len, vocab):
    max_num_mlm_preds = round(max_len * 0.15)
    all_token_ids, all_segments, valid_lens,  = [], [], []
    all_pred_positions, all_mlm_weights, all_mlm_labels = [], [], []
    nsp_labels = []
    for (token_ids, pred_positions, mlm_pred_label_ids, segments,
         is_next) in examples:
        all_token_ids.append(torch.tensor(token_ids + [vocab['<pad>']] * (
            max_len - len(token_ids)), dtype=torch.long))
        all_segments.append(torch.tensor(segments + [0] * (
            max_len - len(segments)), dtype=torch.long))
        # valid_lens不包括'<pad>'的计数
        valid_lens.append(torch.tensor(len(token_ids), dtype=torch.float32))
        all_pred_positions.append(torch.tensor(pred_positions + [0] * (
            max_num_mlm_preds - len(pred_positions)), dtype=torch.long))
        # 填充词元的预测将通过乘以0权重在损失中过滤掉
        all_mlm_weights.append(
            torch.tensor([1.0] * len(mlm_pred_label_ids) + [0.0] * (
                max_num_mlm_preds - len(pred_positions)),
                dtype=torch.float32))
        all_mlm_labels.append(torch.tensor(mlm_pred_label_ids + [0] * (
            max_num_mlm_preds - len(mlm_pred_label_ids)), dtype=torch.long))
        nsp_labels.append(torch.tensor(is_next, dtype=torch.long))
    return (all_token_ids, all_segments, valid_lens, all_pred_positions,
            all_mlm_weights, all_mlm_labels, nsp_labels)

The WikiText-2 dataset for pretraining BERT

#@save
class _WikiTextDataset(torch.utils.data.Dataset):
    def __init__(self, paragraphs, max_len):
        # 输入paragraphs[i]是代表段落的句子字符串列表；
        # 而输出paragraphs[i]是代表段落的句子列表，其中每个句子都是词元列表
        paragraphs = [d2l.tokenize(
            paragraph, token='word') for paragraph in paragraphs]
        sentences = [sentence for paragraph in paragraphs
                     for sentence in paragraph]
        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))
        # 获取遮蔽语言模型任务的数据
        examples = [(_get_mlm_data_from_tokens(tokens, self.vocab)
                      + (segments, is_next))
                     for tokens, segments, is_next in examples]
        # 填充输入
        (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)

Download and WikiText-2 dataset and generate pretraining examples

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')
    paragraphs = _read_wiki(data_dir)
    train_set = _WikiTextDataset(paragraphs, max_len)
    train_iter = torch.utils.data.DataLoader(train_set, batch_size,
                                        shuffle=True, num_workers=num_workers)
    return train_iter, train_set.vocab

batch_size, max_len = 512, 64
train_iter, vocab = load_data_wiki(batch_size, max_len)

for (tokens_X, segments_X, valid_lens_x, pred_positions_X, mlm_weights_X,
     mlm_Y, nsp_y) in train_iter:
    print(tokens_X.shape, segments_X.shape, valid_lens_x.shape,
          pred_positions_X.shape, mlm_weights_X.shape, mlm_Y.shape,
          nsp_y.shape)
    break

output:
Downloading ../data/wikitext-2-v1.zip from https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-v1.zip...
torch.Size([512, 64]) torch.Size([512, 64]) torch.Size([512]) torch.Size([512, 10]) torch.Size([512, 10]) torch.Size([512, 10]) torch.Size([512])

BERT 预训练代码

A small BERT , using 2 layers , 128 hidden units , and 2 self - attention heads

net = d2l.BERTModel(len(vocab), num_hiddens=128, norm_shape=[128],
                    ffn_num_input=128, ffn_num_hiddens=256, num_heads=2,
                    num_layers=2, dropout=0.2, key_size=128, query_size=128,
                    value_size=128, hid_in_features=128, mlm_in_features=128,
                    nsp_in_features=128)
devices = d2l.try_all_gpus()
loss = nn.CrossEntropyLoss()

Computes the loss for both the masked language modeling and next sentence prediction tasks

def _get_batch_loss_bert(net, loss, vocab_size, tokens_X,
                         segments_X, valid_lens_x,
                         pred_positions_X, mlm_weights_X,
                         mlm_Y, nsp_y):
    # 前向传播
    _, mlm_Y_hat, nsp_Y_hat = net(tokens_X, segments_X,
                                  valid_lens_x.reshape(-1),
                                  pred_positions_X)
    # 计算遮蔽语言模型损失
    mlm_l = loss(mlm_Y_hat.reshape(-1, vocab_size), mlm_Y.reshape(-1)) *\
    mlm_weights_X.reshape(-1, 1)
    mlm_l = mlm_l.sum() / (mlm_weights_X.sum() + 1e-8)
    # 计算下一句子预测任务的损失
    nsp_l = loss(nsp_Y_hat, nsp_y)
    l = mlm_l + nsp_l
    return mlm_l, nsp_l, l

Pretrain BERT (net) on the WikiText-2 (train_iter) dataset

def train_bert(train_iter, net, loss, vocab_size, devices, num_steps):
    net = nn.DataParallel(net, device_ids=devices).to(devices[0])
    trainer = torch.optim.Adam(net.parameters(), lr=0.01)
    step, timer = 0, d2l.Timer()
    animator = d2l.Animator(xlabel='step', ylabel='loss',
                            xlim=[1, num_steps], legend=['mlm', 'nsp'])
    # 遮蔽语言模型损失的和，下一句预测任务损失的和，句子对的数量，计数
    metric = d2l.Accumulator(4)
    num_steps_reached = False
    while step < num_steps and not num_steps_reached:
        for tokens_X, segments_X, valid_lens_x, pred_positions_X,\
            mlm_weights_X, mlm_Y, nsp_y in train_iter:
            tokens_X = tokens_X.to(devices[0])
            segments_X = segments_X.to(devices[0])
            valid_lens_x = valid_lens_x.to(devices[0])
            pred_positions_X = pred_positions_X.to(devices[0])
            mlm_weights_X = mlm_weights_X.to(devices[0])
            mlm_Y, nsp_y = mlm_Y.to(devices[0]), nsp_y.to(devices[0])
            trainer.zero_grad()
            timer.start()
            mlm_l, nsp_l, l = _get_batch_loss_bert(
                net, loss, vocab_size, tokens_X, segments_X, valid_lens_x,
                pred_positions_X, mlm_weights_X, mlm_Y, nsp_y)
            l.backward()
            trainer.step()
            metric.add(mlm_l, nsp_l, tokens_X.shape[0], 1)
            timer.stop()
            animator.add(step + 1,
                         (metric[0] / metric[3], metric[1] / metric[3]))
            step += 1
            if step == num_steps:
                num_steps_reached = True
                break

    print(f'MLM loss {metric[0] / metric[3]:.3f}, '
          f'NSP loss {metric[1] / metric[3]:.3f}')
    print(f'{metric[2] / timer.sum():.1f} sentence pairs/sec on '
          f'{str(devices)}')

train

Representing Text with BERT

def get_bert_encoding(net, tokens_a, tokens_b=None):
    tokens, segments = d2l.get_tokens_and_segments(tokens_a, tokens_b)
    token_ids = torch.tensor(vocab[tokens], device=devices[0]).unsqueeze(0)
    segments = torch.tensor(segments, device=devices[0]).unsqueeze(0)
    valid_len = torch.tensor(len(tokens), device=devices[0]).unsqueeze(0)
    encoded_X, _, _ = net(token_ids, segments, valid_len)
    return encoded_X

Consider the sentence “a crane is flying”

Now consider a sentence pair “a crane driver came” and “he just left”

QA

Q1：什么是 Embedding
A1：对于每一个词（token），学习一个长为固定长度的向量，以便于神经网络能处理。