文本分类之 residual-connection+selfAttention的词向量平均模型

这是一个文本分类的系列专题，将采用不同的方法有简单到复杂实现文本分类。
使用Stanford sentiment treebank 电影评论数据集 (Socher et al. 2013). 数据集可以从这里下载
链接：数据集
提取码：yeqw
代码请参考：文本分类 和博客code一致

文本分类 之 self attention 机制

在前面 word average model 和 word average with attention model的基础上，我们做个扩展，加上self attention.

我们再定义一种基于self-attention 的句子模型。

$${\alpha }_t=emb(x_t{)}^{T}emb(x_s)$$
α t ∝ e x p { ∑ t α t s } \alpha_t \propto exp\{\sum_t\alpha_{ts}\} αt​∝exp{t∑​αts​}
$$h_{self}=\sum _t{\alpha }_{t}emb(x_t)$$

句子的正面情感的概率为
$$\sigma (W^T{h}_{self})$$
单词的权重是该单词的embedding和所有其他单词的embedding的dot product的和，然后做softmax归一化。这个模型和 word average with attention 的区别是没有额外引入模型参数u.

另一个变种是把词向量的平均向量也加入self-attention向量，相当于一种residual connection 的方法。
$$\sigma (W^T(h_{self}+h_{avg}))$$

本文我们将实现 加residual connection,self-attention的 词向量平均模型。

import random
from collections import Counter
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import math
USE_CUDA = torch.cuda.is_available()
device = torch.device('cuda' if USE_CUDA else 'cpu')

读数据

with open('senti.train.tsv','r') as rf:
    lines = rf.readlines()
print(lines[:10])

['hide new secretions from the parental units\t0\n', 'contains no wit , only labored gags\t0\n', 'that loves its characters and communicates something rather beautiful about human nature\t1\n', 'remains utterly satisfied to remain the same throughout\t0\n', 'on the worst revenge-of-the-nerds clich茅s the filmmakers could dredge up\t0\n', "that 's far too tragic to merit such superficial treatment\t0\n", 'demonstrates that the director of such Hollywood blockbusters as Patriot Games can still turn out a small , personal film with an emotional wallop .\t1\n', 'of saucy\t1\n', "a depressed fifteen-year-old 's suicidal poetry\t0\n", "are more deeply thought through than in most ` right-thinking ' films\t1\n"]

def read_corpus(path):
    sentences = []
    labels = []
    with open(path,'r', encoding='utf-8') as f:
        for line in f:
            sentence, label = line.split('\t')
            sentences.append(sentence.lower().split())
            labels.append(label[0])
    return sentences, labels

train_path,dev_path,test_path = 'senti.train.tsv','senti.dev.tsv','senti.test.tsv'

train_sentences, train_labels = read_corpus(train_path)
dev_sentences, dev_labels = read_corpus(dev_path)
test_sentences, test_labels = read_corpus(test_path)

train_sentences[1], train_labels[1]

(['contains', 'no', 'wit', ',', 'only', 'labored', 'gags'], '0')

构造词典

def build_vocab(sentences, word_size=20000):
    c = Counter()
    for sent in sentences:
        for word in sent:
            c[word] += 1
    print('文本总单词量为：',len(c))
    words_most_common = c.most_common(word_size)
    ## adding unk, pad
    idx2word = ['<pad>','<unk>'] + [item[0] for item in words_most_common]
    word2dix = {w:i for i, w in enumerate(idx2word)}
    return idx2word, word2dix

WORD_SIZE=20000
idx2word, word2dix = build_vocab(train_sentences, word_size=WORD_SIZE)

文本总单词量为： 14828

idx2word[:10]

['<pad>', '<unk>', 'the', ',', 'a', 'and', 'of', '.', 'to', "'s"]

构造batch

def numeralization(sentences, labels, word2idx):
    '把word list表示的句子转成 index 表示的列表'
    numeral_sent = [[word2dix.get(w, word2dix['<unk>']) for w in s] for s in sentences]
    numeral_label =[int(label) for label in labels]
    return list(zip(numeral_sent, numeral_label))

num_train_data = numeralization(train_sentences, train_labels, word2dix)
num_test_data = numeralization(test_sentences, test_labels, word2dix)
num_dev_data = numeralization(dev_sentences, dev_labels, word2dix)

def convert2tensor(batch_sentences):
    '将batch数据转成tensor,这里主要是为了padding'
    lengths = [len(s) for s in batch_sentences]
    max_len = max(lengths)
    batch_size = len(batch_sentences)
    batch = torch.zeros(batch_size, max_len, dtype=torch.long)
    for i, l in enumerate(lengths):
        batch[i, :l] = torch.tensor(batch_sentences[i])
    return batch

def generate_batch(numeral_sentences_labels, batch_size=32):
    '''将list index 数据 分成batch '''
    batches = []
    num_sample = len(numeral_sentences_labels)
    random.shuffle(numeral_sentences_labels)
    numeral_sent = [n[0] for n in numeral_sentences_labels]
    numeral_label = [n[1] for n in numeral_sentences_labels]
    for start in range(0, num_sample, batch_size):
        end = start + batch_size
        if end > num_sample:
            batch_sentences = numeral_sent[start : num_sample]
            batch_labels = numeral_label[start : num_sample]
            batch_sent_tensor = convert2tensor(batch_sentences)
            batch_label_tensor = torch.tensor(batch_labels, dtype=torch.float)
        else:
            batch_sentences = numeral_sent[start : end]
            batch_labels = numeral_label[start : end]
            batch_sent_tensor = convert2tensor(batch_sentences)
            batch_label_tensor = torch.tensor(batch_labels, dtype=torch.float)
        batches.append((batch_sent_tensor.cuda(), batch_label_tensor.cuda()))
    return batches

train_data = generate_batch(num_train_data)

构建模型

class AVGSelfAttnModel(nn.Module):
    def __init__(self, vocab_size, embed_dim, output_size, pad_idx):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, embed_dim, padding_idx=pad_idx)
        initrange = 0.1
        self.embedding.weight.data.uniform_(-initrange, initrange)
        self.qkv = nn.Linear(embed_dim, embed_dim, bias=False)
        self.fc = nn.Linear(embed_dim, output_size,bias=False)
        
    def forward(self, text):
        ## [batch_size, seq_len]->[batch_size, seq_len, embed_dim]
        embed = self.embedding(text)
        ##[batch_size, seq_len, embed_dim]->[batch_size, seq_len, embed_dim]
        x = self.qkv(embed) 
        ## 计算句子attention
        h_attn = self.attention(x)
        ## 添加 residual connection
        h_attn += embed
        ## 添加 layer norm (可以分别看一下添加和不添加的效果)
#         h_attn = self.layer_norma(h_attn)
        ## 计算平 整个句子 attention 之后的embedding 句子相加得到句子的表示
        h_attn = torch.sum(h_attn, dim=1).squeeze()
        out = self.fc(h_attn)
        return out
    
    def attention(self, x):
        d_k = x.size(-1)
        ##[batch_size, seq_len, embed_dim] * [batch_size, embed_dim, seq_len] ->[batch_size, seq_len, seq_len]
        score = torch.matmul(x, x.transpose(-2, -1))/math.sqrt(d_k)
        ## 计算权重 attn:[batch_size, seq_len, seq_len]
        attn = F.softmax(score, dim=-1) 
        ## 计算context 值 attn_x: [batch_size, seq_len, embed_dim]
        attn_x = torch.matmul(attn, x)
        return attn_x
    
    def layer_norm(self, x):
        mean = x.mean(-1, keep_dim=True)
        std = x.std(-1, keep_dim=True)
        x_lm = (x-mean)/std
        return x_lm

    def get_embed_weigth(self):
        return self.embedding.weight.data

VOCAB_SIZE = len(word2dix)
EMBEDDING_DIM = 100
OUTPUT_SIZE = 1
PAD_IDX = word2dix['<pad>']

model = AVGSelfAttnModel(vocab_size=VOCAB_SIZE,
                 embed_dim=EMBEDDING_DIM,
                 output_size=OUTPUT_SIZE, 
                 pad_idx=PAD_IDX)
model.to(device)

AVGSelfAttnModel(
  (embedding): Embedding(14830, 100, padding_idx=0)
  (qkv): Linear(in_features=100, out_features=100, bias=False)
  (fc): Linear(in_features=100, out_features=1, bias=False)
)

定义损失函数 和优化函数

criterion = nn.BCEWithLogitsLoss()
criterion = criterion.to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-4)

训练模型

def get_accuracy(output, label):
    ## output: batch_size 
    y_hat = torch.round(torch.sigmoid(output)) ## 将output 转成0和1
    correct = (y_hat == label).float()
    acc = correct.sum()/len(correct)
    return acc

def evaluate(batch_data, model, criterion, get_accuracy):
    model.eval()
    num_epoch = epoch_loss = epoch_acc = 0
    with torch.no_grad():
        for text, label in batch_data:
            out = model(text).squeeze(1)
            loss = criterion(out, label)
            acc = get_accuracy(out, label)
            num_epoch +=1 
            epoch_loss += loss.item()
            epoch_acc += acc.item()
    
    return epoch_loss/num_epoch, epoch_acc/num_epoch          

def train(batch_data, model, criterion, optimizer, get_accuracy):
    model.train()
    num_epoch = epoch_loss = epoch_acc = 0
    for text, label in batch_data:
        model.zero_grad()
        out = model(text).squeeze(1)
        loss = criterion(out, label)
        acc = get_accuracy(out, label)
        loss.backward()
        optimizer.step()
        num_epoch +=1 
        epoch_loss += loss.item()
        epoch_acc += acc.item()
    
    return epoch_loss/num_epoch, epoch_acc/num_epoch
        

NUM_EPOCH = 30
best_valid_acc = -1

dev_data = generate_batch(num_dev_data)
for epoch in range(NUM_EPOCH):
    train_data = generate_batch(num_train_data)
    train_loss, train_acc = train(train_data, model, criterion, optimizer, get_accuracy)
    valid_loss, valid_acc = evaluate(dev_data, model, criterion, get_accuracy)
    if valid_acc > best_valid_acc:
        best_valid_acc = valid_acc
        torch.save(model.state_dict(),'self-attn-model.pt')
    
    print(f'Epoch: {epoch+1:02} :')
    print(f'\t Train Loss: {train_loss:.4f} | Train Acc: {train_acc*100:.2f}%')
    print(f'\t Valid Loss: {valid_loss:.4f} | Valid Acc: {valid_acc*100:.2f}%')
    

Epoch: 01 :
	 Train Loss: 0.5429 | Train Acc: 72.38%
	 Valid Loss: 0.4695 | Valid Acc: 78.12%
Epoch: 02 :
	 Train Loss: 0.2947 | Train Acc: 88.60%
	 Valid Loss: 0.5573 | Valid Acc: 79.02%
Epoch: 03 :
	 Train Loss: 0.2277 | Train Acc: 91.26%
	 Valid Loss: 0.6375 | Valid Acc: 79.80%
Epoch: 04 :
	 Train Loss: 0.1964 | Train Acc: 92.50%
	 Valid Loss: 0.7260 | Valid Acc: 80.25%
Epoch: 05 :
	 Train Loss: 0.1759 | Train Acc: 93.27%
	 Valid Loss: 0.7696 | Valid Acc: 82.25%
Epoch: 06 :
	 Train Loss: 0.1642 | Train Acc: 93.81%
	 Valid Loss: 0.8865 | Valid Acc: 80.58%
Epoch: 07 :
	 Train Loss: 0.1538 | Train Acc: 94.13%
	 Valid Loss: 0.9686 | Valid Acc: 79.35%
Epoch: 08 :
	 Train Loss: 0.1461 | Train Acc: 94.53%
	 Valid Loss: 0.9697 | Valid Acc: 81.81%
Epoch: 09 :
	 Train Loss: 0.1409 | Train Acc: 94.63%
	 Valid Loss: 1.1235 | Valid Acc: 79.46%
Epoch: 10 :
	 Train Loss: 0.1356 | Train Acc: 94.89%
	 Valid Loss: 1.1045 | Valid Acc: 81.14%
Epoch: 11 :
	 Train Loss: 0.1326 | Train Acc: 95.05%
	 Valid Loss: 1.2394 | Valid Acc: 80.13%
Epoch: 12 :
	 Train Loss: 0.1296 | Train Acc: 95.11%
	 Valid Loss: 1.3044 | Valid Acc: 79.35%
Epoch: 13 :
	 Train Loss: 0.1265 | Train Acc: 95.18%
	 Valid Loss: 1.4154 | Valid Acc: 79.02%
Epoch: 14 :
	 Train Loss: 0.1242 | Train Acc: 95.28%
	 Valid Loss: 1.4540 | Valid Acc: 79.35%
Epoch: 15 :
	 Train Loss: 0.1219 | Train Acc: 95.36%
	 Valid Loss: 1.5596 | Valid Acc: 78.91%
Epoch: 16 :
	 Train Loss: 0.1208 | Train Acc: 95.40%
	 Valid Loss: 1.5866 | Valid Acc: 78.68%
Epoch: 17 :
	 Train Loss: 0.1190 | Train Acc: 95.48%
	 Valid Loss: 1.6453 | Valid Acc: 78.35%
Epoch: 18 :
	 Train Loss: 0.1175 | Train Acc: 95.51%
	 Valid Loss: 1.6904 | Valid Acc: 79.35%
Epoch: 19 :
	 Train Loss: 0.1170 | Train Acc: 95.59%
	 Valid Loss: 1.7406 | Valid Acc: 79.24%
Epoch: 20 :
	 Train Loss: 0.1160 | Train Acc: 95.57%
	 Valid Loss: 1.8767 | Valid Acc: 77.01%
Epoch: 21 :
	 Train Loss: 0.1149 | Train Acc: 95.67%
	 Valid Loss: 1.8612 | Valid Acc: 78.68%
Epoch: 22 :
	 Train Loss: 0.1142 | Train Acc: 95.62%
	 Valid Loss: 1.9032 | Valid Acc: 78.46%
Epoch: 23 :
	 Train Loss: 0.1126 | Train Acc: 95.68%
	 Valid Loss: 1.9864 | Valid Acc: 77.90%
Epoch: 24 :
	 Train Loss: 0.1118 | Train Acc: 95.78%
	 Valid Loss: 2.0475 | Valid Acc: 76.67%
Epoch: 25 :
	 Train Loss: 0.1113 | Train Acc: 95.76%
	 Valid Loss: 2.0904 | Valid Acc: 77.79%
Epoch: 26 :
	 Train Loss: 0.1100 | Train Acc: 95.85%
	 Valid Loss: 2.1268 | Valid Acc: 77.01%
Epoch: 27 :
	 Train Loss: 0.1105 | Train Acc: 95.75%
	 Valid Loss: 2.1717 | Valid Acc: 77.90%
Epoch: 28 :
	 Train Loss: 0.1092 | Train Acc: 95.88%
	 Valid Loss: 2.2729 | Valid Acc: 77.46%
Epoch: 29 :
	 Train Loss: 0.1091 | Train Acc: 95.79%
	 Valid Loss: 2.3031 | Valid Acc: 78.01%
Epoch: 30 :
	 Train Loss: 0.1082 | Train Acc: 95.95%
	 Valid Loss: 2.3582 | Valid Acc: 77.34%

model.load_state_dict(torch.load('self-attn-model.pt'))

<All keys matched successfully>

test_data = generate_batch(num_test_data)
test_loss, test_acc = evaluate(test_data, model, criterion, get_accuracy)
print(f'Test Loss: {test_loss:.4f} |  Test Acc: {test_acc*100:.2f}%')

Test Loss: 0.6522 |  Test Acc: 81.61%