1. 数据预处理

和上一个博客https://www.cnblogs.com/douzujun/p/13511237.html中的数据和预处理都一致。

import numpy as np
import torch
from torch import nn, optim
import torch.nn.functional as F
from torchtext import data
import math
import time
from torch.autograd import Variable
import copy
import random

SEED = 126
BATCH_SIZE = 128
EMBEDDING_DIM = 100       # 词向量维度
LEARNING_RATE = 1e-3      # 学习率

# 设置device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

#为了保证实验结果可以复现，我们经常会把各种random seed固定在某一个值
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)

# 两个Field对象定义字段的处理方法（文本字段、标签字段）
TEXT = data.Field(tokenize=lambda x: x.split(), batch_first=True, lower=True) # 是否Batch_first. 默认值: False.
LABEL = data.LabelField(dtype=torch.float)

# get_dataset: 返回Dataset所需的 text 和 label
def get_dataset(corpus_path, text_field, label_field):
    fields = [('text', text_field), ('label', label_field)]   # torchtext文本配对关系
    examples = []
    
    with open(corpus_path) as f:
        li = []
        while True:
            content = f.readline().replace('
', '')
            if not content:     # 为空行，表示取完一次数据(一次的数据保存在li中)
                if not li:
                    break
                label = li[0][10]
                text = li[1][6:-7]
                examples.append(data.Example.fromlist([text, label], fields=fields))
                li = []
            else:
                li.append(content)
    return examples, fields

# 得到构建Dataset所需的examples 和 fields
train_examples, train_fileds = get_dataset('./corpus/trains.txt', TEXT, LABEL)
dev_examples, dev_fields = get_dataset('./corpus/dev.txt', TEXT, LABEL)
test_examples, test_fields = get_dataset('./corpus/tests.txt', TEXT, LABEL)

# 构建Dataset数据集
train_data = data.Dataset(train_examples, train_fileds)
dev_data = data.Dataset(dev_examples, dev_fields)
test_data = data.Dataset(test_examples, test_fields)
# for t in test_data:
#     print(t.text, t.label)

print('len of train data:', len(train_data))  # 1000
print('len of dev data:', len(dev_data))      # 200
print('len of test data:', len(test_data))    # 300

# 创建vocabulary
TEXT.build_vocab(train_data, max_size=5000, vectors='glove.6B.100d')
LABEL.build_vocab(train_data)
print(len(TEXT.vocab))         # 3287
print(TEXT.vocab.itos[:12])    # ['<unk>', '<pad>', 'the', 'and', 'a', 'to', 'is', 'was', 'i', 'of', 'for', 'in']
print(TEXT.vocab.stoi['love']) # 129
# print(TEXT.vocab.stoi)         # defaultdict {'<unk>': 0, '<pad>': 1, ....}

# 创建iterators, 每个iteration都会返回一个batch的example
train_iterator, dev_iterator, test_iterator = data.BucketIterator.splits(
                                            (train_data, dev_data, test_data), 
                                            batch_size=BATCH_SIZE,
                                            device=device,
                                            sort = False)

len of train data: 1000
len of dev data: 200
len of test data: 300
3287
['<unk>', '<pad>', 'the', 'and', 'a', 'to', 'is', 'was', 'i', 'of', 'for', 'in']
129

2. 定义模型

2.1 Embedding

class InputEmbeddings(nn.Module):
    def __init__(self, vocab_size, embedding_dim):
        super(InputEmbeddings, self).__init__()
        self.embedding_dim = embedding_dim
        self.embed = nn.Embedding(vocab_size, embedding_dim)
        
    def forward(self, x):
        return self.embed(x) * math.sqrt(self.embedding_dim)

2.2 PositionalEncoding

class PositionalEncoding(nn.Module):
    def __init__(self, embedding_dim, dropout, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)
        
        pe = torch.zeros(max_len, embedding_dim)
        
        position = torch.arange(0., max_len).unsqueeze(1)   # [max_len, 1], 位置编码
        div_term = torch.exp(torch.arange(0., embedding_dim, 2) * -(math.log(10000.0) / embedding_dim))
        
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)              # 增加维度
        print(pe.shape)
        
        self.register_buffer('pe', pe)    # 内存中定一个常量，模型保存和加载的时候，可以写入和读出
        
    def forward(self, x):
        x = x + Variable(self.pe[:, :x.size(1)], requires_grad=False) # Embedding + PositionalEncoding
        return self.dropout(x)

2.3 MultiHeadAttention

self-attention-->建立一个全连接的网络结构

def clones(module, N):
    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])

def attention(query, key, value, mask=None, dropout=None): # q,k,v: [batch, h, seq_len, d_k]    
    d_k = query.size(-1)                                                  # query的维度
    scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)  # 打分机制 [batch, h, seq_len, seq_len]
    
    if mask is not None:
        scores = scores.masked_fill(mask == 0, -1e9) # mask==0的内容填充-1e9, 使计算softmax时概率接近0
    p_atten = F.softmax(scores, dim = -1)            # 对最后一个维度归一化得分, [batch, h, seq_len, seq_len]
    
    if dropout is not None:
        p_atten = dropout(p_atten)
        
    return torch.matmul(p_atten, value), p_atten  # [batch, h, seq_len, d_k]
    

# 建立一个全连接的网络结构
class MultiHeadedAttention(nn.Module):
    def __init__(self, h, embedding_dim, dropout=0.1):
        super(MultiHeadedAttention, self).__init__()
        assert embedding_dim % h == 0 
        
        self.d_k = embedding_dim // h   # 将 embedding_dim 分割成 h份 后的维度
        self.h = h                      # h 指的是 head数量
        self.linears = clones(nn.Linear(embedding_dim, embedding_dim), 4)  
        self.dropout = nn.Dropout(p = dropout)
    
    def forward(self, query, key, value, mask = None):  # q,k,v: [batch, seq_len, embedding_dim]
        
        if mask is not None:
            mask = mask.unsqueeze(1)     # [batch, seq_len, 1]
        nbatches = query.size(0)             
        
        # 1. Do all the linear projections(线性预测) in batch from embeddding_dim => h x d_k
        # [batch, seq_len, h, d_k] -> [batch, h, seq_len, d_k]
        query, key, value = [l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2) 
                                 for l, x in zip(self.linears, (query, key, value))]
        
        # 2. Apply attention on all the projected vectors in batch.
        # atten:[batch, h, seq_len, d_k], p_atten: [batch, h, seq_len, seq_len]
        attn, p_atten = attention(query, key, value, mask=mask, dropout=self.dropout)
        
        # 3. "Concat" using a view and apply a final linear.
        # [batch, h, seq_len, d_k]->[batch, seq_len, embedding_dim]
        attn = attn.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k)
        
        return self.linears[-1](attn)

2.4 MyTransformerModel

class MyTransformerModel(nn.Module):
    def __init__(self, vocab_size, embedding_dim, p_drop, h, output_size):
        super(MyTransformerModel, self).__init__()
        self.drop = nn.Dropout(p_drop)
        
        # Embeddings, 
        self.embeddings = InputEmbeddings(vocab_size=vocab_size, embedding_dim=embedding_dim)
        # H: [e_x1 + p_1, e_x2 + p_2, ....]
        self.position = PositionalEncoding(embedding_dim, p_drop)
        # Multi-Head Attention
        self.atten = MultiHeadedAttention(h, embedding_dim)       # self-attention-->建立一个全连接的网络结构
        # 层归一化(LayerNorm)
        self.norm = nn.LayerNorm(embedding_dim)
        # Feed Forward
        self.linear = nn.Linear(embedding_dim, output_size)
        # 初始化参数
        self.init_weights()
        
    def init_weights(self):
        init_range = 0.1
        self.linear.bias.data.zero_()
        self.linear.weight.data.uniform_(-init_range, init_range)
    
    def forward(self, inputs, mask):      # 维度均为: [batch, seq_len]
        
        embeded = self.embeddings(inputs) # 1. InputEmbedding [batch, seq_len, embedding_dim] 
#         print(embeded.shape)              # torch.Size([36, 104, 100])
        
        embeded = self.position(embeded)  # 2. PosionalEncoding [batch, seq_len, embedding_dim]
#         print(embeded.shape)              # torch.Size([36, 104, 100])
        
        mask = mask.unsqueeze(2)          # [batch, seq_len, 1]

        # 3.1 MultiHeadedAttention [batch, seq_len. embedding_dim]
        inp_atten = self.atten(embeded, embeded, embeded, mask)  
        # 3.2 LayerNorm [batch, seq_len, embedding_dim]
        inp_atten = self.norm(inp_atten + embeded)
#         print(inp_atten.shape)            # torch.Size([36, 104, 100])
        
        # 4. Masked, [batch, seq_len, embedding_dim]
        inp_atten = inp_atten * mask        # torch.Size([36, 104, 100])         

#         print(inp_atten.sum(1).shape, mask.sum(1).shape)  # [batch, emb_dim], [batch, 1]
        b_avg = inp_atten.sum(1) / (mask.sum(1) + 1e-5)  # [batch, embedding_dim]
        
        return self.linear(b_avg).squeeze()              # [batch, 1] -> [batch]

使用模型，使用预训练过的embedding来替换随机初始化，定义优化器、损失函数。

model = MyTransformerModel(len(TEXT.vocab), EMBEDDING_DIM, p_drop=0.5, h=2, output_size=1)

pretrained_embedding = TEXT.vocab.vectors
print('pretrained_embedding:', pretrained_embedding.shape)      #torch.Size([3287, 100])
model.embeddings.embed.weight.data.copy_(pretrained_embedding)  #embeddings是MyTransformerModel的参数, embed是InputEmbedding的参数
print('embedding layer inited.')

optimizer = optim.Adam(model.parameters(), lr=1e-3, weight_decay=0.001)
criteon = nn.BCEWithLogitsLoss()

pretrained_embedding: torch.Size([3287, 100])
embedding layer inited.

3. 训练、评估函数

常规套路：计算准确率、训练函数、评估函数、打印模型表现、用保存的模型参数预测测试数据。

#计算准确率
def binary_acc(preds, y):

    preds = torch.round(torch.sigmoid(preds))
    correct = torch.eq(preds, y).float()
    acc = correct.sum() / len(correct)
    return acc


#训练函数
def train(model, iterator, optimizer, criteon):

    avg_loss = []
    avg_acc = []
    model.train()        #表示进入训练模式

    for i, batch in enumerate(iterator):

        mask = 1 - (batch.text == TEXT.vocab.stoi['<pad>']).float()   #[batch, seq_len]增加了这句，其他都一样
        pred = model(batch.text, mask)

        loss = criteon(pred, batch.label)
        acc = binary_acc(pred, batch.label).item()   #计算每个batch的准确率
        avg_loss.append(loss.item())
        avg_acc.append(acc)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    avg_acc = np.array(avg_acc).mean()
    avg_loss = np.array(avg_loss).mean()
    return avg_loss, avg_acc


#评估函数
def evaluate(model, iterator, criteon):

    avg_loss = []
    avg_acc = []
    model.eval()         #表示进入测试模式

    with torch.no_grad():
        for batch in iterator:
            mask = 1 - (batch.text == TEXT.vocab.stoi['<pad>']).float()
            pred = model(batch.text, mask)
            
            loss = criteon(pred, batch.label)
            acc = binary_acc(pred, batch.label).item()
            avg_loss.append(loss.item())
            avg_acc.append(acc)

    avg_loss = np.array(avg_loss).mean()
    avg_acc = np.array(avg_acc).mean()
    return avg_loss, avg_acc


#训练模型，并打印模型的表现
best_valid_acc = float('-inf')

for epoch in range(30):

    start_time = time.time()

    train_loss, train_acc = train(model, train_iterator, optimizer, criteon)
    dev_loss, dev_acc = evaluate(model, dev_iterator, criteon)

    end_time = time.time()

    epoch_mins, epoch_secs = divmod(end_time - start_time, 60)

    if dev_acc > best_valid_acc:          #只要模型效果变好，就保存
        best_valid_acc = dev_acc
        torch.save(model.state_dict(), 'wordavg-model.pt')

    print(f'Epoch: {epoch+1:02} | Epoch Time: {epoch_mins}m {epoch_secs:.2f}s')
    print(f'	Train Loss: {train_loss:.3f} | Train Acc: {train_acc*100:.2f}%')
    print(f'	 Val. Loss: {dev_loss:.3f} |  Val. Acc: {dev_acc*100:.2f}%')


#用保存的模型参数预测数据
model.load_state_dict(torch.load("wordavg-model.pt"))
test_loss, test_acc = evaluate(model, test_iterator, criteon)
print(f'Test. Loss: {test_loss:.3f} |  Test. Acc: {test_acc*100:.2f}%')