TextCNN结构

以下为原论文中的模型结构图：

1. embedding
2. 卷积层
3. MaxPooling
4. Flatten
5. 全连接层

卷积操作

• 在卷积神经网络中仅涉及离散卷积的情形。
• 卷积运算的作用就类似与滤波，因此也称卷积核为filter滤波器。
• 卷积神经网络的核心思想是捕捉局部特征（n-gram）。CNN的优势在于能够自动地对g-gram特征进行组合和筛选，获得不同抽象层次的语义信息。
• 下图为用于文本分类任务的TextCNN结构描述（这里详细解释了TextCNN架构以及词向量矩阵是如何做卷积的）
  

1. 输入层：$n\ast k$的矩阵，n为句子中的单词数，k为embedding_size。（为了使向量长度一致，对原句进行了padding操作）
2. 卷积层：在NLP中输入层是一个由词向量拼成的词矩阵，且卷积核的宽和该词矩阵的宽相同，该宽度即为词向量大小，且卷积核只会在高度方向移动。输入层的矩阵与我们的filter进行convolution，然后经过激活函数得到feature map。filter这里有三种大小（3，4，5）。
3. 池化层：max-pooling
4. softmax输出结果。

简单代码实现

import torch
import numpy as np
import torch.nn as nn
import torch.optim as optim
import torch.utils.data as Data
import torch.nn.functional as F

dtype = torch.FloatTensor
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# 3 words sentences (=sequence_length is 3)
sentences = ["i love you", "he loves me", "she likes baseball", "i hate you", "sorry for that", "this is awful"]
labels = [1, 1, 1, 0, 0, 0]  # 1 is good, 0 is not good.

embedding_size = 2
sequence_length = len(sentences[0])
num_classes = len(set(labels))
batch_size = 3

word_list = " ".join(sentences).split()
vocab = list(set(word_list))
word2idx = {w:i for i,w in enumerate(vocab)}
vocab_size = len(vocab)

def make_data(sentences, labels):
    inputs = []
    for sen in sentences:
        inputs.append([word2idx[n] for n in sen.split()])

    targets = []
    for out in labels:
        targets.append(out)

    return inputs, targets

input_batch, target_batch = make_data(sentences, labels)
input_batch, target_batch = torch.LongTensor(input_batch), torch.LongTensor(target_batch)

dataset = Data.TensorDataset(input_batch,target_batch)
loader = Data.DataLoader(dataset, batch_size, True)

class TextCNN(nn.Module):

    def __init__(self):
        super(TextCNN, self).__init__()
        self.W = nn.Embedding(vocab_size, embedding_size)
        output_channel = 3
        self.conv = nn.Sequential(nn.Conv2d(1, output_channel, (2,embedding_size)), # inpu_channel, output_channel, 卷积核高和宽 n-gram 和 embedding_size
                                nn.ReLU(),
                                nn.MaxPool2d((2,1)))
        self.fc = nn.Linear(output_channel,num_classes)

    def forward(self, X):
      '''
      X: [batch_size, sequence_length]
      '''
      batch_size = X.shape[0]
      embedding_X = self.W(X) # [batch_size, sequence_length, embedding_size]
      embedding_X = embedding_X.unsqueeze(1) # add channel(=1) [batch, channel(=1), sequence_length, embedding_size]
      conved = self.conv(embedding_X) # [batch_size, output_channel,1,1]
      flatten = conved.view(batch_size, -1)# [batch_size, output_channel*1*1]
      output = self.fc(flatten)
      return output

维度变换

1. 输入X：[batch_size, sequence_length]
2. embedding：相当于把单词增加了一个维度。[batch_size, sequence_length, embedding_size]；然后我们对它做了一个unsqueeze（1）操作，原因是卷积操作的需要。[batch_size, channel(=1), sequence_length, embedding_size]
3. conved：我们这了进行了一个二维卷积，input_channel为1，output_channel为3，filter_size为（2，embedding_size），相当于bi-gram。
   [batch_size, output_channel,sequence_len-1, 1]

卷积输出的高和宽的计算公式：
$$heigh{t}_{out}=\frac{heigh{t}_{in}-heigh{t}_{kernel}+2\times padding}{stride}+1$$$$widthout=\frac{widt{h}_{in}-widt{h}_{kernel}+2\times padding}{stride}+1$$

4. Max-pooling：[batch_size, output_channel,1,1]
5. Flatten: [batch_size,ouput_channel]

model = TextCNN().to(device)
criterion = nn.CrossEntropyLoss().to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-3)

# Training
for epoch in range(5000):
  for batch_x, batch_y in loader:
    batch_x, batch_y = batch_x.to(device), batch_y.to(device)
    pred = model(batch_x)
    loss = criterion(pred, batch_y)
    if (epoch + 1) % 1000 == 0:
        print('Epoch:', '%04d' % (epoch + 1), 'loss =', '{:.6f}'.format(loss))

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

# Test
test_text = 'i hate me'
tests = [[word2idx[n] for n in test_text.split()]]
test_batch = torch.LongTensor(tests).to(device)
# Predict
model = model.eval()
predict = model(test_batch).data.max(1, keepdim=True)[1]
if predict[0][0] == 0:
    print(test_text,"is Bad Mean...")
else:
    print(test_text,"is Good Mean!!")

用TextCNN进行IMDB电影评论情感分析

1. 数据预处理

• 设置种子SEED，保证结果可复现
• 利用torchtext构建数据集

import torch
from torchtext.legacy import data
from torchtext.legacy import datasets
import random
import numpy as np

SEED = 1234

random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.backends.cudnn.deterministic = True

TEXT = data.Field(tokenize = 'spacy', 
                  tokenizer_language = 'en_core_web_sm',
                  batch_first = True)
LABEL = data.LabelField(dtype = torch.float)

train_data, test_data = datasets.IMDB.splits(TEXT, LABEL)

train_data, valid_data = train_data.split(random_state = random.seed(SEED))

• 构建vocab，加载预训练词嵌入：

MAX_VOCAB_SIZE = 25_000

TEXT.build_vocab(train_data, 
                 max_size = MAX_VOCAB_SIZE, 
                 vectors = "glove.6B.100d", 
                 unk_init = torch.Tensor.normal_)

LABEL.build_vocab(train_data)

• 创建迭代器：

BATCH_SIZE = 64

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

train_iterator, valid_iterator, test_iterator = data.BucketIterator.splits(
    (train_data, valid_data, test_data), 
    batch_size = BATCH_SIZE, 
    device = device)

2. 构建模型

• nn.Embedding中padding_idx：当Embedding是随机初始化的矩阵时，会对padding_idx所在的行进行填0。保证了padding行为的正确性。

维度变换

1. 输入：text [batct_size, senquence_len]
2. embedding和unsqueeze(1): [batch_size,1,sequence_len,embedding_dim]
3. conved:[batch_size, output_channel,senquence_len-filter_size[n]+1,1]
4. pooled:[batch_size,output,channel,1,1]
5. Flatten: 这里的flatten在第三步和第四步进行了squeeze操作[batch_size,output_channel]
6. concat: [batch_size, output_channel*len(filter_sizes]

class CNN(nn.Module):
    def __init__(self, vocab_size, embedding_dim, n_filters, filter_sizes, output_dim, 
                 dropout, pad_idx):
        
        super().__init__()
                
        self.embedding = nn.Embedding(vocab_size, embedding_dim, padding_idx = pad_idx)
        
        self.convs = nn.ModuleList([
                                    nn.Conv2d(in_channels = 1, 
                                              out_channels = n_filters, 
                                              kernel_size = (fs, embedding_dim)) 
                                    for fs in filter_sizes
                                    ])
        
        self.fc = nn.Linear(len(filter_sizes) * n_filters, output_dim)
        
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, text):
                
        #text = [batch size, sent len]
        
        embedded = self.embedding(text)
                
        #embedded = [batch size, sent len, emb dim]
        
        embedded = embedded.unsqueeze(1)
        
        #embedded = [batch size, 1, sent len, emb dim]
        
        conved = [F.relu(conv(embedded)).squeeze(3) for conv in self.convs]
            
        #conved_n = [batch size, n_filters, sent len - filter_sizes[n] + 1]
                
        pooled = [F.max_pool1d(conv, conv.shape[2]).squeeze(2) for conv in conved]
        
        #pooled_n = [batch size, n_filters]
        
        cat = self.dropout(torch.cat(pooled, dim = 1))

        #cat = [batch size, n_filters * len(filter_sizes)]
            
        return self.fc(cat)

• 设置超参数，实例化模型

INPUT_DIM = len(TEXT.vocab)
EMBEDDING_DIM = 100
N_FILTERS = 100
FILTER_SIZES = [3,4,5]
OUTPUT_DIM = 1
DROPOUT = 0.5
PAD_IDX = TEXT.vocab.stoi[TEXT.pad_token]

model = CNN(INPUT_DIM, EMBEDDING_DIM, N_FILTERS, FILTER_SIZES, OUTPUT_DIM, DROPOUT, PAD_IDX)

• 查看参数

def count_parameters(model):
    return sum(p.numel() for p in model.parameters() if p.requires_grad)

print(f'The model has {count_parameters(model):,} trainable parameters')

• 加载预训练词嵌入

pretrained_embeddings = TEXT.vocab.vectors

model.embedding.weight.data.copy_(pretrained_embeddings)

• 将unk 和 pad 初始权重归零。

UNK_IDX = TEXT.vocab.stoi[TEXT.unk_token]

model.embedding.weight.data[UNK_IDX] = torch.zeros(EMBEDDING_DIM)
model.embedding.weight.data[PAD_IDX] = torch.zeros(EMBEDDING_DIM)

3. 训练模型

• 初始化优化器、损失函数

import torch.optim as optim

optimizer = optim.Adam(model.parameters())

criterion = nn.BCEWithLogitsLoss()

model = model.to(device)
criterion = criterion.to(device)

• 计算精度的函数

def binary_accuracy(preds, y):
    """
    Returns accuracy per batch, i.e. if you get 8/10 right, this returns 0.8, NOT 8
    """

    #round predictions to the closest integer
    rounded_preds = torch.round(torch.sigmoid(preds))
    correct = (rounded_preds == y).float() #convert into float for division 
    acc = correct.sum() / len(correct)
    return acc

• train

def train(model, iterator, optimizer, criterion):
    
    epoch_loss = 0
    epoch_acc = 0
    
    model.train()
    
    for batch in iterator:
        
        optimizer.zero_grad()
        
        predictions = model(batch.text).squeeze(1)
        
        loss = criterion(predictions, batch.label)
        
        acc = binary_accuracy(predictions, batch.label)
        
        loss.backward()
        
        optimizer.step()
        
        epoch_loss += loss.item()
        epoch_acc += acc.item()
        
    return epoch_loss / len(iterator), epoch_acc / len(iterator)

• eval（注意：同样，由于使用的是dropout，我们必须记住使用model.eval（）来确保在评估时能够关闭 dropout。）

def evaluate(model, iterator, criterion):
    
    epoch_loss = 0
    epoch_acc = 0
    
    model.eval()
    
    with torch.no_grad():
    
        for batch in iterator:

            predictions = model(batch.text).squeeze(1)
            
            loss = criterion(predictions, batch.label)
            
            acc = binary_accuracy(predictions, batch.label)

            epoch_loss += loss.item()
            epoch_acc += acc.item()
        
    return epoch_loss / len(iterator), epoch_acc / len(iterator)

import time

def epoch_time(start_time, end_time):
    elapsed_time = end_time - start_time
    elapsed_mins = int(elapsed_time / 60)
    elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
    return elapsed_mins, elapsed_secs

• training

N_EPOCHS = 5

best_valid_loss = float('inf')

for epoch in range(N_EPOCHS):

    start_time = time.time()
    
    train_loss, train_acc = train(model, train_iterator, optimizer, criterion)
    valid_loss, valid_acc = evaluate(model, valid_iterator, criterion)
    
    end_time = time.time()

    epoch_mins, epoch_secs = epoch_time(start_time, end_time)
    
    if valid_loss < best_valid_loss:
        best_valid_loss = valid_loss
        torch.save(model.state_dict(), 'tut4-model.pt')
    
    print(f'Epoch: {epoch+1:02} | Epoch Time: {epoch_mins}m {epoch_secs}s')
    print(f'\tTrain Loss: {train_loss:.3f} | Train Acc: {train_acc*100:.2f}%')
    print(f'\t Val. Loss: {valid_loss:.3f} |  Val. Acc: {valid_acc*100:.2f}%')

• test

model.load_state_dict(torch.load('tut4-model.pt'))

test_loss, test_acc = evaluate(model, test_iterator, criterion)

print(f'Test Loss: {test_loss:.3f} | Test Acc: {test_acc*100:.2f}%')

4. 模型验证

import spacy
nlp = spacy.load('en_core_web_sm')

def predict_sentiment(model, sentence, min_len = 5):
    model.eval()
    tokenized = [tok.text for tok in nlp.tokenizer(sentence)]
    if len(tokenized) < min_len:
        tokenized += ['<pad>'] * (min_len - len(tokenized))
    indexed = [TEXT.vocab.stoi[t] for t in tokenized]
    tensor = torch.LongTensor(indexed).to(device)
    tensor = tensor.unsqueeze(0)
    prediction = torch.sigmoid(model(tensor))
    return prediction.item()

• 输入需要评测的句子

predict_sentiment(model, "This film is terrible")
predict_sentiment(model, "This film is great")

优化

• 在TextCNN的实践中，有很多地方可以优化：

  Filter尺寸：这个参数决定了抽取n-gram特征的长度，这个参数主要跟数据有关，平均长度在50以内的话，用10以下就可以了，否则可以长一些。在调参时可以先用一个尺寸grid search，找到一个最优尺寸，然后尝试最优尺寸和附近尺寸的组合

  Filter个数：这个参数会影响最终特征的维度，维度太大的话训练速度就会变慢。这里在100-600之间调参即可
  CNN的激活函数：可以尝试Identity、ReLU、tanh

  正则化：指对CNN参数的正则化，可以使用dropout或L2，但能起的作用很小，可以试下小的dropout率(<0.5)，L2限制大一点

  Pooling方法：根据情况选择mean、max、k-max pooling，大部分时候max表现就很好，因为分类任务对细粒度语义的要求不高，只抓住最大特征就好了

  Embedding表：中文可以选择char或word级别的输入，也可以两种都用，会提升些效果。如果训练数据充足（10w+），也可以从头训练
  蒸馏BERT的logits，利用领域内无监督数据

  加深全连接：原论文只使用了一层全连接，而加到3、4层左右效果会更好

TextCNN是很适合中短文本场景的强baseline，但不太适合长文本，因为卷积核尺寸通常不会设很大，无法捕获长距离特征。同时max-pooling也存在局限，会丢掉一些有用特征。另外再仔细想的话，TextCNN和传统的n-gram词袋模型本质是一样的，它的好效果很大部分来自于词向量的引入，解决了词袋模型的稀疏性问题。

参考文献
TextCNN 的 PyTorch 实现
深入TextCNN（一）详述CNN及TextCNN原理
深度学习文本分类模型综述+代码+技巧