BERT全称Bidirectional Enoceder Representations from Transformers,即双向的Transformers的Encoder。
BERT结构
Bert的内部结构有两种Size,其中L表示的是transformer的层数,H表示输出的维度,A表示mutil-head attention的个数。
BERT的输入
bert的输入可以是单一的一个句子或者是句子对,实际的输入值是segment embedding与position embedding相加。
BERT的输入词向量是三个向量之和:
Token Embeddings:词向量,第一个单词是CLS标志,可以用于之后的分类任务
Segment Embeddings:表明这个词属于哪个句子(NSP需要两个句子)
Position Embedding:学习出来的embedding向量。这与Transformer不同,Transformer中是预先设定好的值。
BERT的预训练
BERT的预训练阶段包括两个任务,一个是Masked Language Model (Masked LM),还有一个是下句预测 Next Sentence Prediction(NSP)。
#1 Masked Language Model
Masked LM 可以形象地称为完形填空问题,随机掩盖掉每一个句子中15%的词,用其上下文来去判断被盖住的词原本应该是什么。
举例来说,有这样一个未标注句子 my dog is hairy ,我们可能随机选择了hairy进行遮掩,就变成 my dog is [mask] ,训练模型去预测 [mask] 位置的词,使预测出 hairy的可能性最大。
但是因为是mask15%的词,其数量已经很高了,这样就会导致某些词在fine-tuning阶段从未见过,为了解决这个问题,作者做了如下的处理:
80%的时间是采用[mask],my dog is hairy → my dog is [MASK]
10%的时间是随机取一个词来代替mask的词,my dog is hairy -> my dog is apple
10%的时间保持不变,my dog is hairy -> my dog is hairy
为什么以一定的概率使用随机词呢?
因为transformer要保持对每个输入token分布式的表征,否则Transformer很可能会记住这个[MASK]就是"hairy",从而导致若训练样本和微调的样本mask不一致的情况下,模型预测出现很大的偏差。
#2 Next Sentence Prediction
MLM任务倾向于抽取token层次的表征,因此不能直接获取句子层次的表征。为了使模型能够有能力理解句子间的关系,BERT使用了NSP任务来预训练,预测两个句子是否连在一起。
对于每一个训练样例,我们在语料库中挑选出句子A和句子B来组成,50%的时候句子B就是句子A的下一句(标注为IsNext),剩下50%的时候句子B是语料库中的随机句子(标注为NotNext)。
其输入形式是,开头是一个特殊符号[CLS],然后两个句子之间用[SEP]隔断:
训练样例如下:
Input = [CLS] the man went to [MASK] store [SEP]he bought a gallon [MASK] milk [SEP]
Label = IsNext
Input = [CLS] the man [MASK] to the store [SEP]penguin [MASK] are flight ##less birds[SEP]
Label = NotNext
BERT的输出
BERT最终输出的就是句子中每个token的768维的向量,第一个位置是[CLS],它的向量表示蕴含了这个句子整体的信息,用于做文本分类等句子级任务;对于序列标注等token级任务,就需要使用到每一个token的向量表示。
代码实现
"""
Task: BERT模型的实现
Date: 2023/12/5
Reference: https://github.com/graykode/nlp-tutorial/blob/master/5-2.BERT/BERT.py
"""
import math
import re
from random import *
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
def make_batch():
batch = []
positive = negative = 0 # 为了记录NSP任务中的正样本和负样本的个数,比例最好是在一个batch中接近1:1
while positive != batch_size / 2 or negative != batch_size / 2:
# 从整个样本中随机抽取对应的样本的索引;比如tokens_a_index=3,tokens_b_index=1;randrange() 方法返回指定递增基数集合中的一个随机数,基数默认值为1
tokens_a_index, tokens_b_index = randrange(len(sentences)), randrange(len(sentences))
# 根据索引获取对应样本:tokens_a=[5, 23, 26, 20, 9, 13, 18] tokens_b=[27, 11, 23, 8, 17, 28, 12, 22, 16, 25]
tokens_a, tokens_b = token_list[tokens_a_index], token_list[tokens_b_index]
# 加上特殊符号,CLS符号是1,sep符号是2:[1, 5, 23, 26, 20, 9, 13, 18, 2, 27, 11, 23, 8, 17, 28, 12, 22, 16, 25, 2]
input_ids = [word_dict['[CLS]']] + tokens_a + [word_dict['[SEP]']] + tokens_b + [word_dict['[SEP]']]
# 分割句子符号:[0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]。0表示第一个句子,1表示第二个句子
segment_ids = [0] * (1 + len(tokens_a) + 1) + [1] * (len(tokens_b) + 1)
'''
MASK LM:有很多种做法,这里只是其中一种,一些做法是没有这个max_pred的。
n_pred=3:整个句子的15%的字符可以被mask掉,这里取和max_pred中的最小值,确保每次计算损失的时候没有那么多字符以及信息充足
'''
n_pred = min(max_pred, max(1, int(round(len(input_ids) * 0.15))))
# 整个句子input_ids中可以被mask的符号必须是非cls和sep符号的
cand_maked_pos = [i for i, token in enumerate(input_ids)
if token != word_dict['[CLS]'] and token != word_dict['[SEP]']] # cand_maked_pos=[1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18];
shuffle(cand_maked_pos) # 打乱顺序
masked_tokens, masked_pos = [], []
# 取其中的三个;masked_pos=[6, 5, 17] 这里对应的是position信息;
# masked_tokens=[13, 9, 16]:真实标签,这里是被mask的元素之前对应的原始单字数字;
for pos in cand_maked_pos[:n_pred]:
masked_pos.append(pos)
masked_tokens.append(input_ids[pos])
if random() < 0.8: # 80%的概率是将它替换成这个特殊的掩码符号
input_ids[pos] = word_dict['[MASK]'] # make mask
elif random() < 0.5: # 10%的概率将它替换成一个随机的词元
index = randint(0, vocab_size - 1) # random index in vocabulary
input_ids[pos] = word_dict[number_dict[index]] # replace
# 最后有10 % 的概率不改变,用来做预测
# Zero Paddings
n_pad = maxlen - len(input_ids) # maxlen=30;n_pad=10
input_ids.extend([0] * n_pad) # 在input_ids后面补零
segment_ids.extend([0] * n_pad) # 在segment_ids 后面补零
# Zero Padding (100% - 15%) tokens 是为了计算一个batch中句子的mlm损失的时候可以组成一个有效矩阵放进去
if max_pred > n_pred:
n_pad = max_pred - n_pred
masked_tokens.extend(
[0] * n_pad) # masked_tokens= [13, 9, 16, 0, 0] masked_tokens 对应的是被mask的元素的原始真实标签groundtruth
masked_pos.extend([0] * n_pad) # masked_pos= [6, 5, 17,0,0] 记录哪些位置被mask了
# a句子和b句子是相邻句子,那么就是正样例。
if tokens_a_index + 1 == tokens_b_index and positive < batch_size / 2:
batch.append([input_ids, segment_ids, masked_tokens, masked_pos, True]) # IsNext
positive += 1
elif tokens_a_index + 1 != tokens_b_index and negative < batch_size / 2:
batch.append([input_ids, segment_ids, masked_tokens, masked_pos, False]) # NotNext
negative += 1
return batch
# Proprecessing Finished
def get_attn_pad_mask(seq_q, seq_k):
batch_size, len_q = seq_q.size()
batch_size, len_k = seq_k.size()
# eq(zero) is PAD token
pad_attn_mask = seq_k.data.eq(0).unsqueeze(1) # batch_size x 1 x len_k(=len_q), one is masking
return pad_attn_mask.expand(batch_size, len_q, len_k) # batch_size x len_q x len_k
def gelu(x):
"Implementation of the gelu activation function by Hugging Face"
return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
class Embedding(nn.Module):
def __init__(self):
super(Embedding, self).__init__()
self.tok_embed = nn.Embedding(vocab_size, d_model) # token embedding
self.pos_embed = nn.Embedding(maxlen, d_model) # position embedding
self.seg_embed = nn.Embedding(n_segments, d_model) # segment(token type) embedding
self.norm = nn.LayerNorm(d_model)
def forward(self, x, seg):
seq_len = x.size(1)
pos = torch.arange(seq_len, dtype=torch.long)
pos = pos.unsqueeze(0).expand_as(x) # (seq_len,) -> (batch_size, seq_len)
embedding = self.tok_embed(x) + self.pos_embed(pos) + self.seg_embed(seg)
return self.norm(embedding)
"""点积自注意力"""
class ScaledDotProductAttention(nn.Module):
def __init__(self):
super(ScaledDotProductAttention, self).__init__()
def forward(self, Q, K, V, attn_mask):
scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(d_k) # scores : [batch_size x n_heads x len_q(=len_k) x len_k(=len_q)]
scores.masked_fill_(attn_mask, -1e9) # Fills elements of self tensor with value where mask is one.
attn = nn.Softmax(dim=-1)(scores)
context = torch.matmul(attn, V)
return context, attn
"""多头注意力机制,类似于Transformer"""
class MultiHeadAttention(nn.Module):
def __init__(self):
super(MultiHeadAttention, self).__init__()
self.W_Q = nn.Linear(d_model, d_k * n_heads)
self.W_K = nn.Linear(d_model, d_k * n_heads)
self.W_V = nn.Linear(d_model, d_v * n_heads)
def forward(self, Q, K, V, attn_mask):
# q: [batch_size x len_q x d_model], k: [batch_size x len_k x d_model], v: [batch_size x len_k x d_model]
residual, batch_size = Q, Q.size(0)
# (B, S, D) -proj-> (B, S, D) -split-> (B, S, H, W) -trans-> (B, H, S, W)
q_s = self.W_Q(Q).view(batch_size, -1, n_heads, d_k).transpose(1,2) # q_s: [batch_size x n_heads x len_q x d_k]
k_s = self.W_K(K).view(batch_size, -1, n_heads, d_k).transpose(1,2) # k_s: [batch_size x n_heads x len_k x d_k]
v_s = self.W_V(V).view(batch_size, -1, n_heads, d_v).transpose(1,2) # v_s: [batch_size x n_heads x len_k x d_v]
attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1, 1) # attn_mask : [batch_size x n_heads x len_q x len_k]
# context: [batch_size x n_heads x len_q x d_v], attn: [batch_size x n_heads x len_q(=len_k) x len_k(=len_q)]
context, attn = ScaledDotProductAttention()(q_s, k_s, v_s, attn_mask)
context = context.transpose(1, 2).contiguous().view(batch_size, -1, n_heads * d_v) # context: [batch_size x len_q x n_heads * d_v]
output = nn.Linear(n_heads * d_v, d_model)(context)
return nn.LayerNorm(d_model)(output + residual), attn # output: [batch_size x len_q x d_model]
"""前馈神经网络"""
class PoswiseFeedForwardNet(nn.Module):
def __init__(self):
super(PoswiseFeedForwardNet, self).__init__()
self.fc1 = nn.Linear(d_model, d_ff)
self.fc2 = nn.Linear(d_ff, d_model)
def forward(self, x):
# (batch_size, len_seq, d_model) -> (batch_size, len_seq, d_ff) -> (batch_size, len_seq, d_model)
return self.fc2(gelu(self.fc1(x)))
"""编码层,包括多头自注意力和前馈神经网络"""
class EncoderLayer(nn.Module):
def __init__(self):
super(EncoderLayer, self).__init__()
self.enc_self_attn = MultiHeadAttention()
self.pos_ffn = PoswiseFeedForwardNet()
def forward(self, enc_inputs, enc_self_attn_mask):
enc_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs, enc_self_attn_mask) # enc_inputs to same Q,K,V
enc_outputs = self.pos_ffn(enc_outputs) # enc_outputs: [batch_size x len_q x d_model]
return enc_outputs, attn
"""BERT模型整体架构"""
class BERT(nn.Module):
def __init__(self):
super(BERT, self).__init__()
self.embedding = Embedding() # 词向量层,构建词表矩阵
# 把N个Encoder堆叠起来
self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
# 前馈神经网络
self.fc = nn.Linear(d_model, d_model)
self.activ1 = nn.Tanh()
# mlm
self.linear = nn.Linear(d_model, d_model)
self.activ2 = gelu
self.norm = nn.LayerNorm(d_model) # 层归一化
# cls 这是一个分类层,维度是从d_model到2
self.classifier = nn.Linear(d_model, 2)
# decoder is shared with embedding layer
embed_weight = self.embedding.tok_embed.weight
n_vocab, n_dim = embed_weight.size()
self.decoder = nn.Linear(n_dim, n_vocab, bias=False)
self.decoder.weight = embed_weight
self.decoder_bias = nn.Parameter(torch.zeros(n_vocab))
def forward(self, input_ids, segment_ids, masked_pos):
# 生成input_ids对应的embdding;和segment_ids对应的embedding
output = self.embedding(input_ids, segment_ids)
# 和Transformer的get_attn_pad_mask一样
enc_self_attn_mask = get_attn_pad_mask(input_ids, input_ids)
for layer in self.layers:
output, enc_self_attn = layer(output, enc_self_attn_mask)
# output : [batch_size, len, d_model], attn : [batch_size, n_heads, d_model, d_model]
# it will be decided by first token(CLS)
h_pooled = self.activ1(self.fc(output[:, 0])) # [batch_size, d_model]
logits_clsf = self.classifier(h_pooled) # [batch_size, 2],cls分类的任务
# [batch_size, max_pred, d_model] 其中一个 masked_pos= [6, 5, 17,0,0]
masked_pos = masked_pos[:, :, None].expand(-1, -1, output.size(-1))
# get masked position from final output of transformer.
h_masked = torch.gather(output, 1, masked_pos) # masking position [batch_size, max_pred, d_model]
h_masked = self.norm(self.activ2(self.linear(h_masked)))
logits_lm = self.decoder(h_masked) + self.decoder_bias # [batch_size, max_pred, n_vocab]
return logits_lm, logits_clsf
if __name__ == '__main__':
# BERT Parameters
maxlen = 30 # 句子的最大长度
batch_size = 6 # 每一组有多少个句子
max_pred = 5 # max tokens of prediction
n_layers = 6 # number of Encoder of Encoder Layer
n_heads = 12 # number of heads in Multi-Head Attention
d_model = 768 # Embedding Size
d_ff = 3072 # 4*d_model, FeedForward dimension
d_k = d_v = 64 # dimension of K(=Q), V
n_segments = 2
text = (
'Hello, how are you? I am Romeo.\n'
'Hello, Romeo My name is Juliet. Nice to meet you.\n'
'Nice meet you too. How are you today?\n'
'Great. My baseball team won the competition.\n'
'Oh Congratulations, Juliet\n'
'Thanks you Romeo'
)
# 过滤掉'.', ',', '?', '!',以\n为分隔符。
sentences = re.sub("[.,!?\\-]", '', text.lower()).split('\n')
# 句子之间使用分隔符(空格)连接。然后使用空格为分隔符进行分割
word_list = list(set(" ".join(sentences).split()))
# 特殊字符对应的数字
word_dict = {'[PAD]': 0, '[CLS]': 1, '[SEP]': 2, '[MASK]': 3}
for i, w in enumerate(word_list):
word_dict[w] = i + 4 # 0,1,2,3为特殊字符,所以从4开始赋值给句子单词
# 转换为【索引-单词】词典
number_dict = {i: w for i, w in enumerate(word_dict)}
# 词汇大小
vocab_size = len(word_dict)
token_list = list()
# 通过词典,把单词转换为数字,存储在token_list中
for sentence in sentences:
arr = [word_dict[s] for s in sentence.split()]
token_list.append(arr)
# 获取NSP任务中的正样本和负样本
batch = make_batch()
input_ids, segment_ids, masked_tokens, masked_pos, isNext = map(torch.LongTensor, zip(*batch))
model = BERT()
# mask的位置才计算损失,不被mask的不计算。
criterion = nn.CrossEntropyLoss(ignore_index=0) # 注意这里的参数ignore_index=0,表示忽略真实标签为0的样本。
optimizer = optim.Adam(model.parameters(), lr=0.001)
for epoch in range(100):
optimizer.zero_grad()
logits_lm, logits_clsf = model(input_ids, segment_ids,
masked_pos) # logits_lm 【6,5,29】 bs*max_pred*voca logits_clsf:[6*2]
loss_lm = criterion(logits_lm.transpose(1, 2), masked_tokens) # for masked LM ;masked_tokens [6,5]
loss_lm = (loss_lm.float()).mean()
loss_clsf = criterion(logits_clsf, isNext) # for sentence classification
loss = loss_lm + loss_clsf
if (epoch + 1) % 10 == 0:
print('Epoch:', '%04d' % (epoch + 1), 'cost =', '{:.6f}'.format(loss))
loss.backward()
optimizer.step()
# Predict mask tokens ans isNext
input_ids, segment_ids, masked_tokens, masked_pos, isNext = map(torch.LongTensor, zip(batch[0]))
print(text)
print([number_dict[w.item()] for w in input_ids[0] if number_dict[w.item()] != '[PAD]'])
logits_lm, logits_clsf = model(input_ids, segment_ids, masked_pos)
logits_lm = logits_lm.data.max(2)[1][0].data.numpy()
print('masked tokens list : ', [pos.item() for pos in masked_tokens[0] if pos.item() != 0])
print('predict masked tokens list : ', [pos for pos in logits_lm if pos != 0])
logits_clsf = logits_clsf.data.max(1)[1].data.numpy()[0]
print('isNext : ', True if isNext else False)
print('predict isNext : ', True if logits_clsf else False)
参考
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
nlp-tutorial/5-2.BERT/BERT.py at master · graykode/nlp-tutorial · GitHub