汽车评论情感数据集

链接：https://pan.baidu.com/s/1K5TWrXbXBRXkCUpMbZq2XA
提取码：9mt9

代码

加载库与参数设置

首先先把一些基础的库进行加载

import random
import torch
from torch.utils.data import DataLoader
from transformers import AdamW, BertTokenizerFast, AutoModelForSequenceClassification
from sklearn.metrics import classification_report, accuracy_score, recall_score, f1_score
from tqdm import tqdm
# Set seed for reproducibility
import pandas as pd
import os
from sklearn.model_selection import train_test_split
import numpy as np

做实验时需要固定随机种子，方便实验的可重复性

# 设置种子
seed = 42random.seed(seed)
torch.manual_seed(seed)
os.environ["CUDA_VISIBLE_DEVICES"] = '0'  # 设置GPU型号# 设置训练装置为GPU或CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

在这里设置GPU或CPU，如果你的机器存在多个GPU，则可以修改以下代码

os.environ["CUDA_VISIBLE_DEVICES"] = '0'修改为os.environ["CUDA_VISIBLE_DEVICES"] = '0,1,2,3'

在情感分析中，一般情况存在三种标签：消极、中性、积极。如果是更细粒度的标签，比如电商中对于评论分析有一星到五星的标签，则根据标签修改字典即可。

text_lst = []
label_lst = []
num = 0
label2id = {"消极": 0, "中性": 1, "积极": 2}

数据集的读取

读取数据集，并确保文本与标签数量相同。读者可根据自己实际情况进行修改该部分。

# 读取"消极"类别数据集
with open('negative.txt', 'r', encoding='utf-8') as f:lines = f.readlines()for idx, ele in enumerate(lines):ele = ele.strip('\n')text_lst.append(ele)label_lst.append(label2id['消极'])# 读取"中性"类别数据集
with open('neutral.txt', 'r', encoding='utf-8') as f:lines = f.readlines()for idx, ele in enumerate(lines):ele = ele.strip('\n')text_lst.append(ele)label_lst.append(label2id['中性'])# 读取"积极"类别数据集
with open('positive.txt', 'r', encoding='utf-8') as f:lines = f.readlines()for idx, ele in enumerate(lines):ele = ele.strip('\n')text_lst.append(ele)label_lst.append(label2id['积极'])# 打印文本和标签的数量，确认两者相同
print(len(text_lst))
print(len(label_lst))
assert len(text_lst) == len(label_lst)

超参数设置与数据集的构建

当准备好数据集后，开始初始化参数，然后进行训练前的基本工作。一般情况下，读者需要修改的参数为batch_size 和num_epochs ，其他参数无需修改。根据albert原始论文，学习率一般在e-5级别，因此可以保留默认参数（其实修改了也不会有很大优化）。

在本实验中，我们对数据集进行了4:1的划分，构建训练集和验证集。我们采用了albert_base_chinese模型，具体模型可以在huggingface下载

# 设置超参数
num_labels = len(set(label_lst))  # 根据数据集中的类别数量调整此参数
max_seq_length = 256
batch_size = 16
learning_rate = 2e-5
num_epochs = 10
accumulation_steps = 4# 加载分词器和模型
# model_name = "allenai/led-base-16384"
model_name = './albert_base_chinese'
tokenizer = BertTokenizerFast.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=num_labels).to(device)
model.resize_token_embeddings(len(tokenizer))# 分割文本和标签
texts = text_lst
labels = label_lst# 切分训练集和测试集，80%的训练集和20%的验证集
X_train, X_test, y_train, y_test = train_test_split(texts, labels, test_size=0.2, random_state=42, shuffle=True)# 对训练集和测试集文本进行编码
train_encodings = tokenizer(X_train, truncation=True, padding=True, max_length=max_seq_length, return_tensors="pt")
eval_encodings = tokenizer(X_test, truncation=True, padding=True, max_length=max_seq_length, return_tensors="pt")
train_labels = y_train
eval_labels = y_test# 创建训练和验证数据集
class ClassificationDataset(torch.utils.data.Dataset):def __init__(self, encodings, labels):self.encodings = encodingsself.labels = labelsdef __getitem__(self, idx):item = {key: val[idx].to(device) for key, val in self.encodings.items()}item["labels"] = torch.tensor(self.labels[idx]).to(device)return itemdef __len__(self):return len(self.labels)train_dataset = ClassificationDataset(train_encodings, y_train)
eval_dataset = ClassificationDataset(eval_encodings, y_test)

模型的训练与验证

# 创建用于训练和验证的数据加载器
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
eval_loader = DataLoader(eval_dataset, batch_size=batch_size, shuffle=False)# 初始化变量来存储预测和真实标签以进行评估
all_eval_predictions = []
all_eval_true_labels = []# 准备优化器和损失函数
optimizer = AdamW(model.parameters(), lr=learning_rate)
loss_fn = torch.nn.CrossEntropyLoss()# 训练循环
model.train()
optimizer.zero_grad()
total_steps = len(train_loader) * num_epochs  # 计算总步数
current_step = 0  # 跟踪当前步数
best_f1 = 0
for epoch in range(num_epochs):# 重置下一个epoch的变量all_eval_predictions = []all_eval_true_labels = []model.train()train_loss = 0.0progress_bar = tqdm(train_loader, desc="Epoch {}/{}".format(epoch + 1, num_epochs))for i, batch in enumerate(progress_bar):input_ids = batch["input_ids"]attention_mask = batch["attention_mask"]labels = batch["labels"]# 将输入数据放入GPU或CPU中input_ids = input_ids.to(device)attention_mask = attention_mask.to(device)labels = labels.to(device)outputs = model(input_ids=input_ids, attention_mask=attention_mask, labels=labels)loss = outputs.loss# 进行梯度累加loss = loss / accumulation_stepsloss.backward()if (i + 1) % accumulation_steps == 0:optimizer.step()optimizer.zero_grad()train_loss += loss.item()progress_bar.set_postfix({"Training Loss": train_loss / (i + 1)})# 更新当前步数current_step += 1progress = current_step / total_stepsprogress_bar.set_postfix({"Training Loss": train_loss / (i + 1), "Progress": "{:.1%}".format(progress)})# 验证循环model.eval()for eval_batch in tqdm(eval_loader, desc="Evaluating"):# 将输入数据放入GPU或CPU中eval_input_ids = eval_batch["input_ids"].to(device)eval_attention_mask = eval_batch["attention_mask"].to(device)eval_labels = eval_batch["labels"].to(device)with torch.no_grad():eval_outputs = model(input_ids=eval_input_ids, attention_mask=eval_attention_mask)eval_logits = eval_outputs.logitseval_predictions = torch.argmax(eval_logits, dim=1).tolist()all_eval_predictions.extend(eval_predictions)all_eval_true_labels.extend(eval_labels.tolist())# 计算验证指标，准确率、召回率和F1值eval_accuracy = accuracy_score(all_eval_true_labels, all_eval_predictions)eval_recall = recall_score(all_eval_true_labels, all_eval_predictions, average='weighted')eval_f1 = f1_score(all_eval_true_labels, all_eval_predictions, average='weighted')eval_accuracy = eval_accuracy * 100eval_recall = eval_recall * 100eval_f1 = eval_f1 * 100if eval_f1 > best_f1:best_f1 = eval_f1torch.save(model.state_dict(), './output/best_model.bin')# 打印验证指标print("Validation Accuracy: {:.2f}%".format(eval_accuracy))print("Validation Recall: {:.2f}%".format(eval_recall))print("Validation F1: {:.2f}%".format(eval_f1))print("classification_report:\n", classification_report(all_eval_true_labels, all_eval_predictions))