小白学习pytorch教程十基于大型电影评论数据集训练第一个lstm模型(代码片段)

@Author:Runsen

本博客对原始IMDB数据集进行预处理，建立一个简单的深层神经网络模型，对给定数据进行情感分析。

• 数据集下载 here.
• 原始数据集，没有进行处理here.

import numpy as np

# read data from text files
with open('reviews.txt', 'r') as f:
    reviews = f.read()
with open('labels.txt', 'r') as f:
    labels = f.read()

编码

在将数据输入深度学习模型之前，应该将其转换为数值，文本转换被称为编码，这涉及到每个字符转换成一个整数。在进行编码之前，需要清理数据。
有以下几个预处理步骤：

1. 删除标点符号。
2. 使用\\n作为分隔符拆分文本。
3. 把所有的评论重新组合成一个大串。

from string import punctuation

# remove punctuation
reviews = reviews.lower()
text = ''.join([c for c in reviews if c not in punctuation])
print(punctuation)  # !"#$%&'()*+,-./:;<=>?@[\\]^_`|~


# split by new lines and spaces
reviews_split = text.split('\\n')
text = ' '.join(reviews_split)

# create a list of words
words = text.split()

建立字典并对评论进行编码

创建一个字典，将词汇表中的单词映射为整数。然后通过这个字典，评论可以转换成整数，然后再传送到模型网络。

from collections import Counter

word_counts = Counter(words)
vocab = sorted(word_counts, key = word_counts.get, reverse = True)

vocab2idx = vocab:idx for idx, vocab in enumerate(vocab, 1)
print("Size of Vocabulary: ", len(vocab))

Size of Vocabulary: 74072

encoded_reviews = []
for review in reviews_split:
    encoded_reviews.append([vocab2idx[vocab] for vocab in review.split()])
print("The number of reviews: ", len(encoded_reviews))

The number of reviews: 25001

对标签进行编码

Negative 和Positive应分别标记为0和1（整数）

splitted_labels = labels.split("\\n")
encoded_labels = np.array([
    1 if label == "positive" else 0 for label in splitted_labels
])

删除异常值

应删除长度为0评论，然后，将对剩余的数据进行填充，保证所有数据具有相同的长度。

length_reviews = Counter([len(x) for x in encoded_reviews])
print("Zero-length reviews: ", length_reviews[0])
print("Maximum review length: ", max(length_reviews))

Zero-length reviews: 1
Maximum review length: 2514

# reviews with length 0
non_zero_idx = [i for i, review in enumerate(encoded_reviews) if len(review) != 0]

# Remove 0-length reviews
encoded_reviews = [encoded_reviews[i] for i in non_zero_idx]
encoded_labels = np.array([encoded_labels[i] for i in non_zero_idx])

填充序列

下面要处理很长和很短的评论，需要使用0填充短评论，使其适合特定的长度，

并将长评论剪切为seq_length的单词。这里设置seq_length=200

def text_padding(encoded_reviews, seq_length):
    
    reviews = []
    
    for review in encoded_reviews:
        if len(review) >= seq_length:
            reviews.append(review[:seq_length])
        else:
            reviews.append([0]*(seq_length-len(review)) + review)
        
    return np.array(reviews)

seq_length = 200
padded_reviews = text_padding(encoded_reviews, seq_length)
print(padded_reviews[:12, :12])

数据加载器

将数据按8:1:1的比例拆分为训练集、验证集和测试集，然后使用“TensorDataset”和“DataLoader”函数来处理评论和标签数据。

ratio = 0.8
train_length = int(len(padded_reviews) * ratio)

X_train = padded_reviews[:train_length]
y_train = encoded_labels[:train_length]

remaining_x = padded_reviews[train_length:]
remaining_y = encoded_labels[train_length:]

test_length = int(len(remaining_x)*0.5)

X_val = remaining_x[: test_length]
y_val = remaining_y[: test_length]

X_test = remaining_x[test_length :]
y_test = remaining_y[test_length :]
print("Feature shape of train review set: ", X_train.shape)
print("Feature shape of   val review set: ", X_val.shape)
print("Feature shape of  test review set: ", X_test.shape)

import torch
from torch.utils.data import TensorDataset, DataLoader

batch_size = 50
device = "cuda" if torch.cuda.is_available() else "cpu"
train_dataset = TensorDataset(torch.from_numpy(X_train).to(device), torch.from_numpy(y_train).to(device))
valid_dataset = TensorDataset(torch.from_numpy(X_val).to(device), torch.from_numpy(y_val).to(device))
test_dataset = TensorDataset(torch.from_numpy(X_test).to(device), torch.from_numpy(y_test).to(device))

train_loader = DataLoader(train_dataset, batch_size = batch_size, shuffle = True)
valid_loader = DataLoader(valid_dataset, batch_size = batch_size, shuffle = True)
test_loader = DataLoader(test_dataset, batch_size = batch_size, shuffle = True)
data_iter = iter(train_loader)
X_sample, y_sample = data_iter.next()

RNN模型的实现

到目前为止，包括标记化在内的预处理已经完成。现在建立一个神经网络模型来预测评论的情绪。

• 首先，嵌入层将单词标记转换为特定大小。

• 第二，一个 LSTM层，由hidden_size和num_layers定义。

• 第三，通过完全连接的层从LSTM层的输出映射期望的输出大小。

• 最后，sigmoid激活层以概率0到1的形式返回输出。

import torch.nn as nn
from torch.autograd import Variable

class Model(nn.Module):
    def __init__(self, vocab_size, embedding_dim, hidden_dim, output_dim, num_layers):
        super(Model, self).__init__()
        
        self.hidden_dim = hidden_dim
        self.num_layers = num_layers
        
        # embedding and LSTM
        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        
        self.lstm = nn.LSTM(input_size = embedding_dim, 
                            hidden_size = hidden_dim, 
                            num_layers = num_layers, 
                            batch_first = True, 
                            dropout = 0.5, 
                            bidirectional = False)
        
        # 完连接层
        self.fc = nn.Sequential(
            nn.Dropout(0.5),
            nn.Linear(hidden_dim, output_dim),
            nn.Sigmoid()
        )
        
    def forward(self, token, hidden):
        
        batch_size = token.size(0)
        
        # embedding and lstm output
        out = self.embedding(token.long())
        out, hidden = self.lstm(out, hidden)
        
        # stack up lstm outputs
        out = out.contiguous().view(-1, self.hidden_dim)
        
        # fully connected layer
        out = self.fc(out)
        
        # reshape to be batch_size first
        out = out.view(batch_size, -1)
        
        # get the last batch of labels
        out = out[:, -1]
    
        return out
    
    def init_hidden(self, batch_size):
        return (Variable(torch.zeros(self.num_layers, batch_size, self.hidden_dim).to(device)), 
                 Variable(torch.zeros(self.num_layers, batch_size, self.hidden_dim).to(device)))

• vocab_size : 词汇量
• embedding_dim : 嵌入查找表中的列数
• hidden_dim : LSTM单元隐藏层中的单元数
• output_dim : 期望输出的大小

vocab_size = len(vocab)+1 # +1 for the 0 padding + our word tokens
embedding_dim = 400
hidden_dim = 256
output_dim = 1
num_layers = 2

model = Model(vocab_size, embedding_dim, hidden_dim, output_dim, num_layers).to(device)
model

训练

对于损失函数，BCELoss被用于二分类交叉熵损失，通过给出介于0和1之间的概率进行分类。使用Adam优化器，学习率为0.001

另外，torch.nn.utils.clip_grad_norm_(model.parameters(), clip = 5)，防止了RNN中梯度的爆炸和消失问题clip是要剪裁最大值。

# Loss function and Optimizer
criterion = nn.BCELoss()
optimizer = torch.optim.Adam(model.parameters(), lr = 0.001)

for epoch in range(num_epochs):
    
    model.train()
    hidden = model.init_hidden(batch_size)
    
    for i, (review, label) in enumerate(train_loader):
        review, label = review.to(device), label.to(device)
        
        # Initialize Optimizer 
        optimizer.zero_grad()
        
        hidden = tuple([h.data for h in hidden])
        
        # Feed Forward 
        output = model(review, hidden)
        
        # Calculate the Loss
        loss = criterion(output.squeeze(), label.float())
        
        # Back Propagation 
        loss.backward()
        
        # Prevent Exploding Gradient Problem 
        nn.utils.clip_grad_norm_(model.parameters(), 5)
        
        # Update 
        optimizer.step()
        
        train_losses.append(loss.item())
        
        # Print Statistics 
        if (i+1) % 100 == 0:
            
            ### Evaluation ###
            
            # initialize hidden state
            val_h = model.init_hidden(batch_size)
            val_losses = []

            model.eval()
            
            for review, label in valid_loader:
                review, label = review.to(device), label.to(device)
                val_h = tuple([h.data for h in val_h])
                output = model(review, val_h)
                val_loss = criterion(output.squeeze(), label.float())
                
                val_losses.append(val_loss.item())
                
            print("Epoch: / | Step , Train Loss :.4f, Val Loss :.4f".
                  format(epoch+1, num_epochs, i+1, np.mean(train_losses), np.mean(val_losses)))