NLP之基于BERT的预测掩码标记和句间关系判断

BERT

程序步骤

1. 设置基本变量值,数据预处理
2. 构建输入样本
   1. 在样本集中随机选取a和b两个句子
   2. 把ab两个句子合并为1个模型输入句,在句首加入分类符CLS,在ab中间和句末加入分隔符SEP
   3. 在模型输入句中随机选取15%单词准备用于mask, 再在这个些选中的单词中,按照论文策略进行mask
   4. 把所有存储单词的变量都填充至最大长度(有利于统一处理)
   5. 判断句间关系(ab是否相邻)
3. 构建BERT模型
   1. 按照论文图2构建输入encoder的嵌入矩阵, Embedding: input = word + position + segment
   2. 根据transformer的方法对输入句子进行padding mask, 在注意力机制中屏蔽掉填充字符
   3. 根据transformer模型,输入编码器,得到输出
   4. Task1: 根据记录的mask单词位置,从输出中提取相应位置模型预测的各单词概率,经过激活全连接归一化等处理,得到最终的语言模型(即一个嵌入矩阵)
   5. Task2: 从输出中提取第一列的元素值(即输入向量CLS对应的输出元素),经过激活全连接归一化等处理,得到最终分类输出
4. 训练: 求任务1的损失值loss_lm和任务2的损失值loss_clsf的和,然后进行梯度下降,直到连续三次的loss都小于阈值0.01
5. 测试: 遍历每个样本句子对,将其输入模型,从返回的logits_lm和logits_clsf中选择每行最大值,即为预测结果.

"""
Task: 基于BERT的完形填空和句间关系判断
Author: ChengJunkai @github.com/Cheng0829
Email: [email protected]
Date: 2022/09/21
Reference: Tae Hwan Jung(Jeff Jung) @graykode
"""

import math, re, random, torch, time, os, sys
import numpy as np
import torch.nn as nn
import torch.optim as optim

def get_attn_pad_mask(seq_q, seq_k): 
    '''mask大小和(len_q,len_k)一致,
    是为了在点积注意力中,与torch.matmul(Q,K)的大小一致'''
    # (seq_q, seq_k): (dec_inputs, enc_inputs) 
    # dec_inputs:[batch_size, tgt_len] # [1,5]
    # enc_inputs:[batch_size, src_len] # [1,6]
    batch_size, len_q = seq_q.size() # 1,5
    batch_size, len_k = seq_k.size() # 1,6
    """Tensor.data.eq(element)
    eq即equal,对Tensor中所有元素进行判断,和element相等即为True,否则为False,返回二值矩阵
    Examples:
        >>> tensor = torch.FloatTensor([[1, 2, 3], [4, 5, 6]])
        >>> tensor.data.eq(1) 
        tensor([[ True, False, False],
                [False, False, False]])
    """
    # eq(zero) is PAD token
    pad_attn_mask = seq_k.data.eq(0).unsqueeze(1) # 升维 enc: [1,6] -> [1,1,6]
    # 矩阵扩充: enc: pad_attn_mask: [1,1,6] -> [1,5,6]
    return pad_attn_mask.expand(batch_size, len_q, len_k) # batch_size, len_q, len_k

'''Attention = Softmax(Q * K^T) * V '''
def Scaled_Dot_Product_Attention(Q, K, V, attn_mask): 
    # Q_s: [batch_size, n_heads, len_q, d_k] # [1,8,5,64]
    # K_s: [batch_size, n_heads, len_k, d_k] # [1,8,6,64]
    # attn_mask: [batch_size, n_heads, len_q, len_k] # [1,8,5,6]

    """torch.matmul(Q, K)
        torch.matmul是tensor的乘法,输入可以是高维的.
        当输入是都是二维时,就是普通的矩阵乘法.
        当输入有多维时,把多出的一维作为batch提出来,其他部分做矩阵乘法.
        Exeamples:
            >>> a = torch.ones(3,4)
            >>> b = torch.ones(4,2)
            >>> torch.matmul(a,b).shape
            torch.Size([3,2])   

            >>> a = torch.ones(5,3,4)
            >>> b = torch.ones(4,2)
            >>> torch.matmul(a,b).shape
            torch.Size([5,3,2])

            >>> a = torch.ones(2,5,3)
            >>> b = torch.ones(1,3,4)
            >>> torch.matmul(a,b).shape
            torch.Size([2,5,4])
        """
    # [1,8,5,64] * [1,8,64,6] -> [1,8,5,6]
    # scores : [batch_size, n_heads, len_q, len_k]
    scores = torch.matmul(Q, K.transpose(2,3)) / np.sqrt(d_k) # divided by scale

    """scores.masked_fill_(attn_mask, -1e9) 
    由于scores和attn_mask维度相同,根据attn_mask中的元素值,把和attn_mask中值为True的元素的
    位置相同的scores元素的值赋为-1e9
    """
    scores.masked_fill_(attn_mask, -1e9)

    # 'P'的scores元素值为-1e9, softmax值即为0
    softmax = nn.Softmax(dim=-1) # 求行的softmax
    attn = softmax(scores) # [1,8,6,6]
    # [1,8,6,6] * [1,8,6,64] -> [1,8,6,64]
    context = torch.matmul(attn, V) # [1,8,6,64]
    return context, attn

class MultiHeadAttention(nn.Module):
    # dec_enc_attn(dec_outputs, enc_outputs, enc_outputs, dec_enc_attn_mask)
    def __init__(self):
        super().__init__()
        self.W_Q = nn.Linear(d_model, d_k*n_heads) # (512, 64*8) # d_q必等于d_k
        self.W_K = nn.Linear(d_model, d_k*n_heads) # (512, 64*8) # 保持维度不变
        self.W_V = nn.Linear(d_model, d_v*n_heads) # (512, 64*8)
        self.linear = nn.Linear(n_heads*d_v, d_model)
        self.layer_norm = nn.LayerNorm(d_model)

    def forward(self, Q, K, V, attn_mask):
        # dec_outputs: [batch_size, tgt_len, d_model] # [1,5,512]
        # enc_outputs: [batch_size, src_len, d_model] # [1,6,512]
        # dec_enc_attn_mask: [batch_size, tgt_len, src_len] # [1,5,6]
        # q/k/v: [batch_size, len_q/k/v, d_model]
        residual, batch_size = Q, len(Q)
        '''用n_heads=8把512拆成64*8,在不改变计算成本的前提下,让各注意力头相互独立,更有利于学习到不同的特征'''
        # Q_s: [batch_size, len_q, n_heads, d_q] # [1,5,8,64]
        # new_Q_s: [batch_size, n_heads, len_q, d_q] # [1,8,5,64]
        Q_s = self.W_Q(Q).view(batch_size, -1, n_heads, d_k).transpose(1,2)  
        # K_s: [batch_size, n_heads, len_k, d_k] # [1,8,6,64]
        K_s = self.W_K(K).view(batch_size, -1, n_heads, d_k).transpose(1,2)  
        # V_s: [batch_size, n_heads, len_k, d_v] # [1,8,6,64]
        V_s = self.W_V(V).view(batch_size, -1, n_heads, d_v).transpose(1,2)  

        # attn_mask : [1,5,6] -> [1,1,5,6] -> [1,8,5,6]
        # attn_mask : [batch_size, n_heads, len_q, len_k]
        attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1, 1) 

        # context: [batch_size, n_heads, len_q, d_v]
        # attn: [batch_size, n_heads, len_q(=len_k), len_k(=len_q)]
        # context: [1,8,5,64] attn: [1,8,5,6]
        context, attn = Scaled_Dot_Product_Attention(Q_s, K_s, V_s, attn_mask)
        """contiguous() 连续的
        contiguous: view只能用在连续(contiguous)的变量上.
        如果在view之前用了transpose, permute等,
        需要用contiguous()来返回一个contiguous copy
        """
        # context: [1,8,5,64] -> [1,5,512]
        context = context.transpose(1, 2).contiguous().view(batch_size, -1, n_heads * d_v)
        # context: [1,5,512] -> [1,5,512]
        output = self.linear(context)
        """nn.LayerNorm(output) 样本归一化
        和对所有样本的某一特征进行归一化的BatchNorm不同,
        LayerNorm是对每个样本进行归一化,而不是一个特征

        Tips:
            归一化Normalization和Standardization标准化区别:
            Normalization(X[i]) = (X[i] - np.min(X)) / (np.max(X) - np.min(X))
            Standardization(X[i]) = (X[i] - np.mean(X)) / np.var(X)
        """
        output = self.layer_norm(output + residual)
        return output, attn 

class Position_wise_Feed_Forward_Networks(nn.Module):
    def __init__(self):
        super().__init__()
        '''输出层相当于1*1卷积层,也就是全连接层'''
        """nn.Conv1d
        in_channels应该理解为嵌入向量维度,out_channels才是卷积核的个数(厚度)
        """
        # 512 -> 2048
        self.conv1 = nn.Conv1d(in_channels=d_model, out_channels=d_ff, kernel_size=1)
        # 2048 -> 512
        self.conv2 = nn.Conv1d(in_channels=d_ff, out_channels=d_model, kernel_size=1)
        self.layer_norm = nn.LayerNorm(d_model)

    def forward(self, inputs):
        # enc_outputs: [batch_size, source_len, d_model] # [1,6,512]
        residual = inputs 
        relu = nn.ReLU()
        # output: 512 -> 2048 [1,2048,6]
        output = relu(self.conv1(inputs.transpose(1, 2)))
        # output: 2048 -> 512 [1,6,512]
        output = self.conv2(output).transpose(1, 2)
        return self.layer_norm(output + residual)

class EncoderLayer(nn.Module):
    def __init__(self):
        super(EncoderLayer, self).__init__()
        self.enc_attn = MultiHeadAttention()
        self.pos_ffn = Position_wise_Feed_Forward_Networks()

    def forward(self, enc_outputs, enc_attn_mask):
        # enc_attn_mask: [1,6,6]
        # enc_outputs to same Q,K,V
        # enc_outputs: [batch_size, source_len, d_model] # [1, 6, 512]
        enc_outputs, attn = self.enc_attn(enc_outputs, \
            enc_outputs, enc_outputs, enc_attn_mask) 
        # enc_outputs: [batch_size , len_q , d_model]
        enc_outputs = self.pos_ffn(enc_outputs) 
        return enc_outputs, attn

"""*********以上为Transformer架构代码*******************"""

'''1.数据预处理'''
def pre_process(text):
    # sentences 共有6个样本句子 
    # 去除字符
    sentences = re.sub("[.,!?\\-]", '', text.lower()).split('\n')  # filter '.', ',', '?', '!'
    word_sequence = " ".join(sentences).split()
    word_list = []
    '''
    如果用list(set(word_sequence))来去重,得到的将是一个随机顺序的列表(因为set无序),
    这样得到的字典不同,保存的上一次训练的模型很有可能在这一次不能用
    (比如上一次的模型预测碰见i:0,love:1,就输出you:2,但这次模型you在字典3号位置,也就无法输出正确结果)
    '''
    for word in word_sequence:
        if word not in word_list:
            word_list.append(word)
    word_dict = {'[PAD]': 0, '[CLS]': 1, '[SEP]': 2, '[MASK]': 3}
    # word_dict前4个均为特殊字符
    for i, w in enumerate(word_list):
        word_dict[w] = i + 4
    number_dict = {i:w for i,w in enumerate(word_dict)}
    vocab_size = len(word_dict)

    # 二维列表
    sentences_list = []
    for sentence in sentences:
        arr = [word_dict[s] for s in sentence.split()]
        sentences_list.append(arr) # 用序号token来替代单词
    print(word_dict)
    return sentences, word_list, word_dict, number_dict, vocab_size, sentences_list

'''论文所用激活函数,在BERT中优于RELU'''
def gelu(x):
    "Implementation of the gelu activate_ation function by Hugging Face"
    # erf(x):计算x的误差函数,
    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))

'''根据句子数据,构建词元的输入向量'''
def make_batch():
    batch = []
    positive = negative = 0
    a_b = []
    '''由于随机选取句子,所以可能依照pt文件中已训练的模型也误差较大'''
    while positive != batch_size/2 or negative != batch_size/2:
        """random.randrange(start, stop=None, step=1) 
        在(start, stop)范围内随即返回一个数
        Args:
            start: 指定范围内的开始值,包含在范围内
            stop: 指定范围内的结束值,不包含在范围内,若无参数stop则默认范围为0~start
            step: 指定递增基数,默认为1
        Examples:
            >>> random.randrange(1, 100, 2) 
            67 # 以2递增,只能返回奇数
        """
        # 在样本集中随机选取a和b两个句子
        sentence_a_index, sentence_b_index = random.randrange(len(sentences)), random.randrange(len(sentences)) 
        # sentence_ab是包含一个句子中所含单词的序号列表
        sentence_a, sentence_b = sentences_list[sentence_a_index], sentences_list[sentence_b_index]

        if (sentence_a_index, sentence_b_index) not in a_b: # 屏蔽已经选择过的
            a_b.append((sentence_a_index, sentence_b_index))
        elif len(a_b) < batch_size: # 达到要求的样本值
            continue
        else:
            break
        # 加入分类符和分隔符
        input_ids = [word_dict['[CLS]']] + sentence_a + [word_dict['[SEP]']] + sentence_b + [word_dict['[SEP]']] 
        # segment_ids用来标识a和b两个句子
        segment_ids = [0] * (1 + len(sentence_a) + 1) + [1] * (len(sentence_b) + 1)

        # MASK LM 
        # 把句子15%单词准备用于mask 
        # max_words_pred(=5): 最大预测单词数  n_pred: 预测个数
        n_pred = int(0.15 * len(input_ids))
        n_pred = min(max_words_pred, max(1, n_pred)) # 预测个数不能为0,也不能大于最大预测单词数
        # 构建候选词列表(随机打乱后只有前n_pred个会被选择)
        # 要屏蔽cls和seq,所以不能用torch.arange(len(input_ids))
        candidate_mask_tokens = [i for i, token in enumerate(input_ids) if token != word_dict['[CLS]']
                               and token != word_dict['[SEP]']] # [1,2,3......](不包含0([CLS])和[SEP])
        # 随机打乱candidate_mask_tokens
        random.shuffle(candidate_mask_tokens)

        # random.shuffle打乱候选表后,直接选择前n_pred个 
        masked_tokens, masked_pos = [], []
        for pos in candidate_mask_tokens[:n_pred]:
            # 不管是被mask还是被随机还是不变,都会加入masked列表(masked中的不会被替换,替换的是input_ids)
            # 记录原始单词,用于和模型输出比对计算loss
            masked_pos.append(pos) 
            masked_tokens.append(input_ids[pos])
            if random.random() < 0.8:  # 80%  -> MASK
                input_ids[pos] = word_dict['[MASK]'] # mask
            elif random.random() < 0.5:  # 10% -> random
                random_index = random.randint(0, vocab_size - 1) # random index in vocabulary
                input_ids[pos] = word_dict[number_dict[random_index]] # replace
    
        '''把所有存储单词的变量都填充至最大长度,有利于统一处理.'''
        # Zero Paddings
        n_pad = max_len - len(input_ids) # max_len: 允许的句子最大长度
        input_ids.extend([0] * n_pad) # word_dict['Pad'] = 0
        segment_ids.extend([0] * n_pad)

        # Zero Padding (100% - 15%) tokens
        if max_words_pred > n_pred: # max_words_pred(=5): 最大预测长度
            n_pad = max_words_pred - n_pred
            masked_tokens.extend([0] * n_pad) 
            masked_pos.extend([0] * n_pad)

        # 判断句间关系
        if sentence_a_index + 1 == sentence_b_index and positive < batch_size/2: # 论文要求严格保持50% 50%
            batch.append([input_ids, segment_ids, masked_tokens, masked_pos, True]) # IsNext
            positive += 1
        elif sentence_a_index + 1 != sentence_b_index and negative < batch_size/2: # 论文要求严格保持50% 50%
            batch.append([input_ids, segment_ids, masked_tokens, masked_pos, False]) # NotNext
            negative += 1
    # batch: [6,5]
    return batch

'''论文图2所设计的输入encoder的嵌入向量'''
class Embedding(nn.Module):
    def __init__(self):
        super().__init__()
        # vocab_size: len(input_ids)
        # vocab_size: 29  max_len: 30  n_segments: 2
        self.tok_embed = nn.Embedding(vocab_size, d_model) # token embedding
        self.pos_embed = nn.Embedding(max_len, d_model) # position embedding
        self.seg_embed = nn.Embedding(n_segments, d_model) # segment(token type) embedding
        self.norm = nn.LayerNorm(d_model) # 对每个样本的特征归一化
    
    def forward(self, x, seg): # embedding(input_ids, segment_ids) # [6,30] [6,30]
        seq_len = x.size(1) # 30
        # torch.arange(n): 生成0~n-1的一维张量,默认类型为int
        # 加入序列信息,不用transformer的positional encoding
        pos = torch.arange(seq_len, dtype=torch.long).to(device) # [0,1,2,3...seq-1]
        '''
        expand_as/expand拓展后,每一个(1,seq_len)都是同步变化的,不能单独修改某一个
        单独修改某一个只能用repeat(重复倍数)
        '''
        pos = pos.unsqueeze(0).expand_as(x)  # (seq_len,) -> (batch_size, seq_len) [6,30]
        '''
        虽然tok_embed是(29,768),但embedding首参数是指单词类别,x虽然长度=30>29,
        但是单词总类别依然肯定不大于29,所以tok_embed可以作为x的嵌入矩阵
        seg_embed同理
        '''
        # [6,30,768], [6,30,768], [6,30,768], [6,30,768]
        '''论文图2, Embedding: input = word + position + segment'''
        embedding = self.tok_embed(x) + self.pos_embed(pos) + self.seg_embed(seg)
        return self.norm(embedding)

'''2.构建模型'''
class BERT(nn.Module):
    def __init__(self):
        super().__init__()
        self.embedding = Embedding()
        self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])
        self.fc = nn.Linear(d_model, d_model)
        self.norm = nn.LayerNorm(d_model)
        self.classifier = nn.Linear(d_model, 2)
        self.activate_1 = gelu
        self.activate_2 = nn.Tanh()

        # decoder is shared with embedding layer
        token_embed_weight = self.embedding.tok_embed.weight
        vocab_size, n_dim = token_embed_weight.size() # 29, 768
        self.decoder = nn.Linear(n_dim, vocab_size, bias=False)
        self.decoder.weight = token_embed_weight
        self.decoder_bias = nn.Parameter(torch.zeros(vocab_size))

    def forward(self, input_ids, segment_ids, masked_pos):
        
        # 按照论文图2构建输入encoder的嵌入矩阵, Embedding: input = word + position + segment
        output = self.embedding(input_ids, segment_ids).to(device)
        # padding mask, 使其在softmax中为0而不影响注意力机制
        enc_self_attn_mask = get_attn_pad_mask(input_ids, input_ids)

        for layer in self.layers:
            # output : [batch_size, max_len, d_model] [6,30,768]
            # attn : [batch_size, n_heads, d_mode, d_model]
            output, enc_self_attn = layer(output, enc_self_attn_mask)

        '''Task1'''
        # masked_pos ->  [batch_size, max_words_pred, d_model] [6,5,768]
        # [:, :, None]在第三个维度增加一维 masked_pos: [6,5] -> [6,5,768]
        masked_pos = masked_pos[:, :, None].expand(-1, -1, output.size(-1)) # expand(-1)表示不变
        # 从最终输出output中提取mask信息
        """torch.gather() 收集输入的特定维度指定位置的数值
        Args:
            input(tensor): 待操作数.不妨设其维度为(x1, x2, ···, xn)
            dim(int): 待操作的维度
            index(LongTensor): 如何对input进行操作
        """
        # predict_masked: [batch_size, max_words_pred, d_model] # [6,5,768]
        # output:[6,30,768] -> predict_masked:[6,5,768]
        predict_masked = torch.gather(output, 1, masked_pos) 
        predict_masked = self.norm(self.activate_1(self.fc(predict_masked))) # activate_1: gelu
        # (self.decoder:nn.Linear(768,29), self.decoder_bias: nn.Parameter(torch.zeros(29))) 
        # logits_lm: [batch_size, max_words_pred, vocab_size] [6,5,29]
        '''所谓的语言模型就是vocab_size个单词的嵌入矩阵'''
        logits_lm = self.decoder(predict_masked) + self.decoder_bias

        '''Task2'''
        '''只训练第一列? CLS能训练?   input是cls,但输出不是'''
        # 从第一个CLS值得到结果
        # output[:, 0]取output第一列的数据 [6,30,768] -> [6,768]
        # h_pooled: [batch_size, d_model] [6,768]
        h_pooled = self.activate_2(self.fc(output[:, 0])) # activate_2: Tanh
        # 全连接分类 (classifier:768->2) 
        logits_clsf = self.classifier(h_pooled) # [batch_size, 2]
        return logits_lm, logits_clsf

if __name__ == '__main__':
    device = ['cuda:0' if torch.cuda.is_available() else 'cpu'][0]
    # BERT Parameters
    max_len = 30 # 允许的句子最大长度:9+9+1+2=21
    batch_size = 10 # batch中的句子对个数,因为positive数最多为5,所以batch_size最大为10 # 必须是偶数, 否则make_batch陷入死循环
    max_words_pred = 5  # 最大标记预测长度 21*0.15=3.15
    n_layers = 12 # number of Encoder of Encoder Layer
    n_heads = 12 # number of heads in Multi-Head Attention
    d_model = 768 # Embedding Size
    d_ff = 4*d_model   # 4*d_model, FeedForward dimension
    d_k = d_v = 64  # dimension of K(=Q), V
    n_segments = 2 # ab两个句子

    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'
    )
    # print(text)

    '''1.数据预处理'''
    sentences, word_list, word_dict, number_dict, vocab_size, sentences_list = pre_process(text)
    batch = make_batch()

    '''2.构建模型'''
    model = BERT()
    model.to(device)
    criterion = nn.CrossEntropyLoss()
    '''lr如果是0.001,将会很快陷入局部收敛!'''
    optimizer = optim.AdamW(model.parameters(), lr=0.0001) # AdamW: 论文采用的Adam扩展版本

    if os.path.exists('model_param.pt') == True:
        # 加载模型参数到模型结构
        model.load_state_dict(torch.load('model_param.pt', map_location=device))
 
    """map(function, iterable, ...)
    map()会根据提供的函数对指定序列做映射。第一个参数function以参数序列中的每一个元素调用function函数,
    返回包含每次function函数返回值的新列表。
    Args:
        function: 函数
        iterable: 一个或多个序列
    """
    """zip([iterable, ...]) 
    用于将可迭代的对象作为参数,将对象中对应的元素打包成一个个元组,然后返回由这些元组组成的列表.
    Examples
        >>> a = [1,2,3]
        >>> b = [4,5,6]
        >>> list(zip([a,b]))
        [(1, 4), (2, 5), (3, 6)]
        
        >>> c = [a,b] # [[1,2,3],[4,5,6]]
        >>> list(zip(*c))
        [(1, 4), (2, 5), (3, 6)]
        >>> list(zip(c))  
        [([1, 2, 3],), ([4, 5, 6],)]
    """
    # 因为batch二维,所以要用*,不然zip会把batch里的每个样本整体当作一个迭代对象返回一个元组
    # map(torch.LongTensor, zip(*batch))相当于torch.LongTensor(input_ids)...... 
    # batch: [6,5]
    input_ids, segment_ids, masked_tokens, masked_pos, isNext = map(torch.LongTensor, zip(*batch))
    input_ids, segment_ids, masked_tokens, masked_pos, isNext = \
        input_ids.to(device), segment_ids.to(device), masked_tokens.to(device), masked_pos.to(device), isNext.to(device)
    # print(masked_tokens.shape, isNext.shape) [6,5] [6,]

    '''3.训练'''
    print('{}\nTrain\n{}'.format('*'*30, '*'*30))   
    loss_record = []
    for epoch in range(1000):
        optimizer.zero_grad()
        logits_lm, logits_clsf = model(input_ids, segment_ids, masked_pos)
        # Task1: 对比完形填空准确率
        # logits_lm:[6,5,29]->[6,29,5] masked_tokens:[6,5] 
        """nn.CrossEntropyLoss()(input,target)
        如果target是一维,则只要求input.size(0)=target.size
        如果是多维,那么要求input.size(0)=target.size(0)且input.size(-1)=target.size(-1)
        """
        loss_lm = criterion(logits_lm.transpose(1, 2), masked_tokens) 
        # Task2: 对比句间关系分类准确率
        # logits_clsf:[6,2] isNext:[6,]
        loss_clsf = criterion(logits_clsf, isNext) 
        loss = loss_lm + loss_clsf
        loss.backward()
        optimizer.step()

        if loss >= 0.01: # 连续30轮loss小于0.01则提前结束训练
            loss_record = []
        else:
            loss_record.append(loss.item())
            if len(loss_record) == 30:
                torch.save(model.state_dict(), 'model_param.pt')
                break            

        if (epoch + 1) % 10 == 0:
            print('Epoch:', '%04d' % (epoch + 1), 'Loss = {:.6f}'.format(loss))
        if(epoch + 1) % 100 ==0:
            torch.save(model.state_dict(), 'model_param.pt')

    '''4.测试'''
    print('{}\nTest\n{}'.format('*'*30, '*'*30))
    print('text:\n%s'%text)
    for i in range(len(batch)): # 遍历每个句子对
        print('*'*30)
        # Predict mask tokens ans isNext
        input_ids, segment_ids, masked_tokens, masked_pos, isNext = map(torch.LongTensor, zip(batch[i]))
        input_ids, segment_ids, masked_tokens, masked_pos, isNext = \
                input_ids.to(device), segment_ids.to(device), masked_tokens.to(device), masked_pos.to(device), isNext.to(device)

        # 输出input_ids   # w.item() != 0 不是'PAD'
        print('input_ids :', [number_dict[w.item()] for w in input_ids[0] if w.item() != 0]) 

        logits_lm, logits_clsf = model(input_ids, segment_ids, masked_pos)
        logits_lm, logits_clsf = logits_lm.to('cpu'), logits_clsf.to('cpu')

        # Task1
        logits_lm = logits_lm.data.max(2)[1][0].data.numpy()
        print('real masked tokens list :', \
            [(pos.item(), number_dict[pos.item()]) for pos in masked_tokens[0] if pos.item() != 0]) # 不是PAD
        print('predict masked tokens list :', \
            [(pos, number_dict[pos]) for pos in logits_lm if pos != 0]) # 不是PAD

        # Task2
        logits_clsf = logits_clsf.data.max(1)[1].data.numpy()[0]
        print('Real isNext :', bool(isNext))
        print('predict isNext :', bool(logits_clsf))