PyTorch框架实现文本分类的案例

@TOC

本文将介绍3个运用PyTorch结构完成文本分类的事例

事例1：情感剖析

在这个事例中，咱们将运用PyTorch和BERT模型来进行情感剖析。咱们将运用IMDb电影评论数据集，其间包括50,000条正负样本的电影评论。

首要，咱们需要经过以下代码下载必要的Python库和数据集：

!pip install torch transformers pandas scikit-learn
!wget https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz
!tar -xf aclImdb_v1.tar.gz

接下来，咱们将界说预处理函数，它将读取并准备好数据，以便咱们能够将其输入到BERT模型中：

import pandas as pd
from sklearn.model_selection import train_test_split
from transformers import BertTokenizer
def preprocess_data():
    # 读取数据集
    df = pd.read_csv('data.csv')
    # 划分练习集和测验集
    train_data, test_data = train_test_split(df, test_size=0.2)
    # 加载BERT tokenizer
    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
    # 对练习集进行编码
    train_encodings = tokenizer(list(train_data['text']), truncation=True, padding=True)
    train_labels = list(train_data['label'])
    # 对测验集进行编码
    test_encodings = tokenizer(list(test_data['text']), truncation=True, padding=True)
    test_labels = list(test_data['label'])
    return train_encodings, train_labels, test_encodings, test_labels

然后，咱们将界说一个PyTorch模型类，它将运用BERT模型和一个线性层来进行情感剖析：

import torch.nn as nn
from transformers import BertModel
class SentimentClassifier(nn.Module):
    def __init__(self):
        super(SentimentClassifier, self).__init__()
        # 加载BERT模型
        self.bert = BertModel.from_pretrained('bert-base-uncased')
        # 增加线性层
        self.linear = nn.Linear(768, 1)
    def forward(self, input_ids, attention_mask):
        # 将输入传递给BERT模型
        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        # 提取[CLS]向量
        pooled_output = outputs.pooler_output
        # 传递给线性层
        logits = self.linear(pooled_output)
        return logits
最终，咱们能够运用以下代码来练习和测验模型：
import torch
from torch.utils.data import DataLoader, TensorDataset
from transformers import AdamW
# 预处理数据
train_encodings, train_labels, test_encodings, test_labels = preprocess_data()
# 将数据转换为PyTorch张量
train_input_ids = torch.tensor(train_encodings.input_ids)
train_attention_mask = torch.tensor(train_encodings.attention_mask)
train_labels = torch.tensor(train_labels)
test_input_ids = torch.tensor(test_encodings.input_ids)
test_attention_mask = torch.tensor(test_encodings.attention_mask)
test_labels = torch.tensor(test_labels)
# 创立数据集和数据加载器
train_dataset = TensorDataset(train_input_ids, train_attention_mask, train_labels)
train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True)
test_dataset = TensorDataset(test_input_ids, test_attention_mask, test_labels)
test_loader = DataLoader(test_dataset, batch_size=16, shuffle=False)
# 界说模型和优化器
model = SentimentClassifier()
optimizer = AdamW(model.parameters(), lr=2e-5)
# 练习模型
for epoch in range(3):
    for input_ids, attention_mask, labels in train_loader:
        optimizer.zero_grad()
        logits = model(input_ids, attention_mask).squeeze(-1)
        loss = nn.BCEWithLogitsLoss()(logits, labels.float())
        loss.backward()
        optimizer.step()
# 测验模型
with torch.no_grad():
    correct, total = 0, 0
    for input_ids, attention_mask, labels in test_loader:
        logits = model(input_ids, attention_mask).squeeze(-1)
        predictions = (torch.sigmoid(logits) > 0.5).long()
        total += labels.size(0)
        correct += (predictions ==  labels.long()).sum().item()
    accuracy = correct / total
    print(f"Accuracy: {accuracy:.2f}")

事例2：垃圾邮件分类

在这个事例中，咱们将运用PyTorch和卷积神经网络（CNN）来进行垃圾邮件分类。咱们将运用SpamAssassin公共数据集，其间包括4,827封垃圾邮件和6,593封非垃圾邮件。

首要，咱们需要下载必要的Python库和数据集：

!pip install torch pandas scikit-learn
!wget https://spamassassin.apache.org/old/publiccorpus/20050311_spam_2.tar.bz2
!tar -xf 20050311_spam_2.tar.bz2

接下来，咱们将界说预处理函数，它将读取并准备好数据，以便咱们能够将其输入到CNN模型中：

import os
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import CountVectorizer
def preprocess_data():
    # 读取数据集
    spam_filenames = os.listdir('spam')
    ham_filenames = os.listdir('ham')
    spam_emails = []
    ham_emails = []
    for filename in spam_filenames:
        with open(os.path.join('spam', filename), encoding='latin-1') as f:
            spam_emails.append(f.read())
    for filename in ham_filenames:
        with open(os.path.join('ham', filename), encoding='latin-1') as f:
            ham_emails.append(f.read())
    # 将数据集转换为DataFrame
    spam_df = pd.DataFrame({'text': spam_emails, 'label': 1})
    ham_df = pd.DataFrame({'text': ham_emails, 'label': 0})
    df = pd.concat([spam_df, ham_df], ignore_index=True)
    # 划分练习集和测验集
    train_data, test_data = train_test_split(df, test_size=0.2)
    # 运用Bag of Words办法将文本转换为数字特征
    vectorizer = CountVectorizer(stop_words='english', max_features=1000)
    vectorizer.fit(train_data['text'])
    train_features = vectorizer.transform(train_data['text']).toarray()
    train_labels = list(train_data['label'])
    test_features = vectorizer.transform(test_data['text']).toarray()
    test_labels = list(test_data['label'])
    return train_features, train_labels, test_features, test_labels

然后，咱们将界说一个PyTorch模型类，它将运用CNN来进行垃圾邮件分类：

import torch.nn as nn
class SpamClassifier(nn.Module):
    def __init__(self, input_dim, output_dim):
        super(SpamClassifier, self).__init__()
        # 增加卷积层
        self.conv1 = nn.Conv1d(input_dim, 128, kernel_size=5, padding=2)
        self.conv2 = nn.Conv1d(128, 64, kernel_size=5, padding=2)
        # 增加池化层
        self.pool1 = nn.MaxPool1d(kernel_size=4)
        self.pool2 = nn.MaxPool1d(kernel_size=4)
        # 增加线性层
        self.linear1 = nn.Linear(64 * (input_dim // 16), 64)
        self.linear2 = nn.Linear(64, output_dim)
    def forward(self, x):
        # 将输入传递给卷积层和池化层
        x = torch.relu(self.conv1(x))
        x = self.pool1(x)
        x = torch.relu(self.conv2(x))
        x = self.pool2(x)
        # 将输出展平并传递给线性层
        x = x.view(x.size(0), -1)
        x = torch.relu(self.linear1(x))
        logits = self.linear2(x)
        return logits

最终，咱们能够运用以下代码来练习和测验模型：

import torch
from torch.utils.data import DataLoader, TensorDataset
# 预处理数据
train_features, train_labels, test_features, test_labels = preprocess_data()
# 将数据转换为PyTorch张量
train_features = torch.tensor(train_features).float()
train_labels = torch.tensor(train_labels)
test_features = torch.tensor(test_features).float()
test_labels = torch.tensor(test_labels)
# 创立数据集和数据加载器
train_dataset = TensorDataset(train_features.unsqueeze(1), train_labels)
train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True)
test_dataset = TensorDataset(test_features.unsqueeze(1), test_labels)
test_loader = DataLoader(test_dataset, batch_size=16, shuffle=False)
# 界说模型和优化器
model = SpamClassifier(input_dim=train_features.shape[1], output_dim=1)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
# 练习模型
for epoch in range(10):
    for features, labels in train_loader:
        optimizer.zero_grad()
        logits = model(features)
        loss = nn.BCEWithLogitsLoss()(logits.squeeze(-1), labels.float())
        loss.backward()
        optimizer.step()
# 测验模型
with torch.no_grad():
    correct, total = 0, 0
    for features, labels in test_loader:
        logits = model(features)
        predictions = (torch.sigmoid(logits) > 0.5).long()
        total += labels.size(0)
        correct += (predictions == labels).sum().item()
    accuracy = correct / total
    print(f"Accuracy: {accuracy:.2f}")

事例3：多标签分类

在这个事例中，咱们将运用PyTorch和BERT模型来进行多标签分类。咱们将运用Reuters新闻语料库数据集，其间包括8,982个新闻文档，每个新闻文档都标有多个类别。

首要，咱们需要下载必要的Python库和数据集：

!pip install torch transformers pandas scikit-learn
!wget https://archive.ics.uci.edu/ml/machine-learning-databases/reuters21578-mld/reuters21578.tar.gz
!tar -xf reuters21578.tar.gz

接下来，咱们将界说预处理函数，它将读取并准备好数据，以便咱们能够将其输入到BERT模型中：

import os
import xml.etree.ElementTree as ET
import pandas as pd
from sklearn.model_selection import train_test_split
from transformers import BertTokenizer
def preprocess_data():
    # 读取数据集
    topics = ['acq', 'earn', 'money-fx', 'grain', 'crude', 'trade', 'interest', 'ship', 'wheat', 'corn']
    documents = []
    for root, dirs, files in os.walk('reut2-*.sgm'):
        for file in files:
            with open(os.path.join(root, file), encoding='latin-1') as f:
                tree = ET.parse(f)
                for elem in tree.iterfind('.//REUTERS'):
                    if any(topic in elem.find('TOPICS').text for topic in topics):
                        text = elem.findtext('.//TEXT')
                        topics_list = [topic.strip() for topic in elem.find('TOPICS').text.split(',') if topic.strip() in topics]
                        if text and topics_list:
                            documents.append({'text': text, 'topics': topics_list})
    df = pd.DataFrame(documents)
    # 划分练习集和测验集
    train_data, test_data = train_test_split(df, test_size=0.2)
    # 加载BERT tokenizer
    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
    # 对练习集进行编码
    train_encodings = tokenizer(list(train_data['text']), truncation=True, padding=True)
    train_labels = torch.tensor([list(map(lambda t: t in doc_topics, topics)) for doc_topics in train_data['topics']])
    # 对测验集进行编码
    test_encodings = tokenizer(list(test_data['text']), truncation=True, padding=True)
    test_labels = torch.tensor([list(map(lambda t: t in doc_topics, topics)) for doc_topics in test_data['topics']])
    return train_encodings, train_labels, test_encodings, test_labels

然后，咱们将界说一个PyTorch模型类，它将运用BERT模型和一个多标签分类头来进行多标签分类：

import torch.nn as nn
from transformers import BertModel
class MultiLabelClassifier(nn.Module):
    def __init__(self):
        super(MultiLabelClassifier, self).__init__()
        # 加载BERT模型
        self.bert = BertModel.from_pretrained('bert-base-uncased')
        # 增加多标签分类头
        self.classifier = nn.Linear(self.bert.config.hidden_size, 10)
    def forward(self, input_ids, attention_mask):
        # 将输入传递给BERT模型
        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = outputs.pooler_output
        # 将输出传递给多标签分类头
        logits = self.classifier(pooled_output)
        return logits

最终，咱们能够运用以下代码来练习和测验模型：

import torch
from torch.utils.data import DataLoader, TensorDataset
# 预处理数据
train_encodings, train_labels, test_encodings, test_labels = preprocess_data()
# 创立数据集和数据加载器
train_dataset = TensorDataset(torch.tensor(train_encodings['input_ids']), 
                              torch.tensor(train_encodings['attention_mask']),
                              train_labels)
train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True)
test_dataset = TensorDataset(torch.tensor(test_encodings['input_ids']), 
                             torch.tensor(test_encodings['attention_mask']),
                             test_labels)
test_loader = DataLoader(test_dataset, batch_size=16, shuffle=False)
# 界说模型和优化器
model = MultiLabelClassifier()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
# 练习模型
for epoch in range(3):
    for input_ids, attention_mask, labels in train_loader:
        optimizer.zero_grad()
        logits = model(input_ids, attention_mask)
        loss = nn.BCEWithLogitsLoss()(logits, labels.float())
        loss.backward()
        optimizer.step()
# 测验模型
with torch.no_grad():
    all_labels, all_predictions = [], []
    for input_ids, attention_mask, labels in test_loader:
        logits = model(input_ids, attention_mask)
        predictions = (torch.sigmoid(logits) > 0.5).long()
        all_labels.append(labels)
        all_predictions.append(predictions)
    all_labels = torch.cat(all_labels, dim=0)
    all_predictions = torch.cat(all_predictions, dim=0)
    accuracy = (all_predictions == all_labels).float().mean()
    print(f"Accuracy: {accuracy:.2f}")

小结

这三个事例演示了如何运用PyTorch和不同类型的神经网络来处理不同的机器学习问题。PyTorch供给了丰富的工具和函数，使得模型的练习和评价变得非常容易。

本文由mdnice多平台发布