Local sentiment analysis upload.

2025-08-23 15:55:07 +08:00
parent 4e33224633
commit f448ddd466
65 changed files with 1563359 additions and 11 deletions
@@ -0,0 +1,142 @@
+# 微博情感识别模型-GPT2-LoRA微调
+
+## 项目说明
+这是一个基于GPT2的微博情感二分类模型，采用LoRA（Low-Rank Adaptation）微调技术。通过PEFT库实现的LoRA微调，只需训练极少量参数就可以让模型适应情感分析任务，大幅降低计算资源需求和模型体积。
+
+## 数据集
+使用微博情感数据集(weibo_senti_100k)，包含约10万条带情感标注的微博内容，正负向评论各约5万条。数据集标签：
+- 标签0：负面情感
+- 标签1：正面情感
+
+## 文件结构
+```
+GPT2-Lora/
+├── train.py                  # 训练脚本（基于PEFT库的LoRA实现）
+├── predict.py                # 预测脚本（交互式使用）
+├── requirements.txt          # 依赖包列表
+├── models/                   # 本地存储的预训练模型
+│   └── gpt2-chinese/        # 中文GPT2模型及配置
+├── dataset/                  # 数据集目录
+│   └── weibo_senti_100k.csv # 微博情感数据集
+└── best_weibo_sentiment_lora/ # 训练好的LoRA权重（训练后生成）
+```
+
+## 技术特点
+
+1. **极度参数高效**：相比全参数微调，仅训练约0.1%-1%的参数
+2. **使用PEFT库**：基于Hugging Face官方的参数高效微调库，稳定可靠
+3. **模型性能保持**：在仅训练极少参数的情况下，保持良好的分类性能
+4. **部署友好**：LoRA权重文件小，便于模型部署和分享
+
+## LoRA技术优势
+
+LoRA (Low-Rank Adaptation) 是目前最流行的参数高效微调技术：
+
+1. **超低参数量**：通过低秩分解，将大矩阵分解为两个小矩阵的乘积
+2. **插件式设计**：LoRA权重可以动态加载和卸载，一个基础模型支持多个任务
+3. **训练速度快**：参数少，训练时间短，内存占用小
+4. **无损原模型**：原始预训练模型权重保持不变，避免灾难性遗忘
+
+## 环境依赖
+
+安装所需依赖：
+```bash
+pip install -r requirements.txt
+```
+
+主要依赖包：
+- Python 3.8+
+- PyTorch 1.13+
+- Transformers 4.28+
+- PEFT 0.4+
+- Pandas, NumPy, Scikit-learn
+
+## 使用方法
+
+### 1. 安装依赖
+```bash
+pip install -r requirements.txt
+```
+
+### 2. 训练模型
+```bash
+python train.py
+```
+
+训练过程会自动：
+- 下载并本地保存中文GPT2预训练模型
+- 加载微博情感数据集
+- 使用LoRA技术训练模型
+- 保存最佳LoRA权重到 `./best_weibo_sentiment_lora/`
+
+### 3. 情感分析预测
+```bash
+python predict.py
+```
+
+运行后将进入交互模式：
+- 在控制台输入要分析的微博文本
+- 系统会返回情感分析结果（正面/负面）和置信度
+- 输入'q'退出程序
+
+## 模型配置
+
+- **基础模型**: `uer/gpt2-chinese-cluecorpussmall` 中文预训练模型
+- **模型本地保存路径**: `./models/gpt2-chinese/`
+- **LoRA配置**:
+  - rank (r): 8 - 低秩矩阵的秩
+  - alpha: 32 - 缩放因子
+  - target_modules: ["c_attn", "c_proj"] - 目标线性层
+  - dropout: 0.1 - 防止过拟合
+
+## 性能对比
+
+| 方法 | 可训练参数占比 | 模型文件大小 | 训练时间 | 推理速度 |
+|------|----------------|--------------|----------|----------|
+| 全参数微调 | 100% | ~500MB | 长 | 慢 |
+| Adapter微调 | ~3% | ~50MB | 中等 | 中等 |
+| **LoRA微调** | **~0.5%** | **~2MB** | **短** | **快** |
+
+## 使用示例
+
+```
+使用设备: cuda
+LoRA模型加载成功!
+
+============= 微博情感分析 (LoRA版) =============
+输入微博内容进行分析 (输入 'q' 退出):
+
+请输入微博内容: 这部电影真是太好看了，我非常喜欢！
+预测结果: 正面情感 (置信度: 0.9876)
+
+请输入微博内容: 服务态度差，价格还贵，一点都不推荐
+预测结果: 负面情感 (置信度: 0.9742)
+
+请输入微博内容: q
+```
+
+## 注意事项
+
+1. **首次运行**：首次运行 `train.py` 时会自动下载预训练模型，请确保网络连接
+2. **GPU推荐**：虽然LoRA参数少，但建议使用GPU加速训练
+3. **模型加载**：预测时需要先有训练好的LoRA权重文件
+4. **兼容性**：基于PEFT库实现，与Hugging Face生态系统完全兼容
+
+## 扩展功能
+
+- **多任务支持**：可以为不同任务训练不同的LoRA权重，共享同一个基础模型
+- **权重合并**：可以将多个LoRA权重合并，或将LoRA权重合并到基础模型中
+- **动态切换**：支持运行时动态加载和切换不同的LoRA权重
+
+## 技术原理
+
+LoRA通过在原始线性层旁边添加两个小的矩阵A和B，使得：
+```
+h = W₀x + BAx
+```
+其中：
+- W₀是冻结的预训练权重
+- B ∈ ℝᵈˣʳ, A ∈ ℝʳˣᵏ是可训练的低秩矩阵
+- r << min(d,k)，大大减少了参数量
+
+这种设计既保持了预训练模型的知识，又能高效地适应新任务。
@@ -0,0 +1,280 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# weibo_senti_100k 说明\n",
+    "0. **下载地址：** [百度网盘](https://pan.baidu.com/s/1DoQbki3YwqkuwQUOj64R_g)\n",
+    "1. **数据概览：** 10 万多条，带情感标注 新浪微博，正负向评论约各 5 万条\n",
+    "2. **推荐实验：** 情感/观点/评论 倾向性分析\n",
+    "2. **数据来源：** [新浪微博](https://weibo.com/)\n",
+    "3. **原数据集：** [新浪微博，情感分析标记语料共12万条](https://download.csdn.net/download/weixin_38442818/10214750)，网上搜集，具体作者、来源不详\n",
+    "4. **加工处理：**\n",
+    "    1. 将原来的 2 份文档，整合成 1 份 csv 文件\n",
+    "    2. 编码统一为 UTF-8\n",
+    "    3. 去重"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import pandas as pd"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "path = 'weibo_senti_100k_文件夹_所在_路径'"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# 1. weibo_senti_100k.csv"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 加载数据"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 15,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "评论数目（总体）：119988\n",
+      "评论数目（正向）：59993\n",
+      "评论数目（负向）：59995\n"
+     ]
+    }
+   ],
+   "source": [
+    "pd_all = pd.read_csv(path + 'weibo_senti_100k.csv')\n",
+    "\n",
+    "print('评论数目（总体）：%d' % pd_all.shape[0])\n",
+    "print('评论数目（正向）：%d' % pd_all[pd_all.label==1].shape[0])\n",
+    "print('评论数目（负向）：%d' % pd_all[pd_all.label==0].shape[0])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 字段说明\n",
+    "\n",
+    "| 字段 | 说明 |\n",
+    "| ---- | ---- |\n",
+    "| label | 1 表示正向评论，0 表示负向评论 |\n",
+    "| review | 微博内容 |"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<div>\n",
+       "<style scoped>\n",
+       "    .dataframe tbody tr th:only-of-type {\n",
+       "        vertical-align: middle;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe tbody tr th {\n",
+       "        vertical-align: top;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe thead th {\n",
+       "        text-align: right;\n",
+       "    }\n",
+       "</style>\n",
+       "<table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       "    <tr style=\"text-align: right;\">\n",
+       "      <th></th>\n",
+       "      <th>label</th>\n",
+       "      <th>review</th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <th>62050</th>\n",
+       "      <td>0</td>\n",
+       "      <td>太过分了@Rexzhenghao  //@Janie_Zhang:招行最近负面新闻越来越多呀...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>68263</th>\n",
+       "      <td>0</td>\n",
+       "      <td>希望你?得好?我本＂?肥血?史＂[晕][哈哈]@Pete三姑父</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>81472</th>\n",
+       "      <td>0</td>\n",
+       "      <td>有点想参加????[偷?]想安排下时间再决定[抓狂]//@黑晶晶crystal: @细腿大羽...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>42021</th>\n",
+       "      <td>1</td>\n",
+       "      <td>[给力]感谢所有支持雯婕的芝麻！[爱你]</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>7777</th>\n",
+       "      <td>1</td>\n",
+       "      <td>2013最后一天，在新加坡开心度过，向所有的朋友们问声：新年快乐！2014年，我们会更好[调...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>100399</th>\n",
+       "      <td>0</td>\n",
+       "      <td>大中午出门办事找错路，曝晒中。要多杯具有多杯具。[泪][泪][汗]</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>82398</th>\n",
+       "      <td>0</td>\n",
+       "      <td>马航还会否认吗？到底在隐瞒啥呢？[抓狂]//@头条新闻: 转发微博</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>106423</th>\n",
+       "      <td>0</td>\n",
+       "      <td>克罗地亚球迷很爱放烟火！球又没进，就硝烟四起。[晕]</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>24798</th>\n",
+       "      <td>1</td>\n",
+       "      <td>[抱抱]福芦 TangRoulou 吉祥书 8.8折优惠 &gt;&gt;&gt; http://t.cn/z...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>6598</th>\n",
+       "      <td>1</td>\n",
+       "      <td>回复@钱旭明QXM:[嘻嘻][嘻嘻] //@钱旭明QXM:杨大哥[good][good][g...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>53920</th>\n",
+       "      <td>1</td>\n",
+       "      <td>人家这脸长的!!!!!![哈哈]</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>15587</th>\n",
+       "      <td>1</td>\n",
+       "      <td>这个价不算高，和一天内训相比相差无几。。[哈哈]//@博通传媒v: 6个月！一个月工资1万，...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>101237</th>\n",
+       "      <td>0</td>\n",
+       "      <td>终于收工啦，脚丫子快冻掉了[泪][泪][泪]</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>82449</th>\n",
+       "      <td>0</td>\n",
+       "      <td>我决定从今天开始我想吃什么就去吃什么，一个人吃也无所谓，重点是不要因为别人的意见委屈了自己[...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>32537</th>\n",
+       "      <td>1</td>\n",
+       "      <td>飘雪的北京 需要双份早餐.......//@美食天下: [哈哈]//@王淼Margay: 屁...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>10630</th>\n",
+       "      <td>1</td>\n",
+       "      <td>[耶]，这个太赞了，生活大爆炸第六季马上要出啦[鼓掌] //@-郑瑜-:这个不错 //@经典...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>85130</th>\n",
+       "      <td>0</td>\n",
+       "      <td>刚追完#倾世皇妃#，#千山暮雪#又紧随其后，网速和更新速度都太不给力，尽管我看过原著，还是焦...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>105956</th>\n",
+       "      <td>0</td>\n",
+       "      <td>晚上看金二胖?察前?，推出的火炮基座?糟了，可以PK了[泪] //@艾米粒er: //@wi...</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>72391</th>\n",
+       "      <td>0</td>\n",
+       "      <td>必须把中国足球的伟大，用我的职业演说出来 //@袁腾飞:[泪]</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>10761</th>\n",
+       "      <td>1</td>\n",
+       "      <td>[鼓掌] //@宁波香格里拉大酒店: 小编来答疑，周五晚惊艳全场的树根蛋糕到底有多长？蛋糕全...</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table>\n",
+       "</div>"
+      ],
+      "text/plain": [
+       "        label                                             review\n",
+       "62050       0  太过分了@Rexzhenghao  //@Janie_Zhang:招行最近负面新闻越来越多呀...\n",
+       "68263       0                    希望你?得好?我本＂?肥血?史＂[晕][哈哈]@Pete三姑父\n",
+       "81472       0  有点想参加????[偷?]想安排下时间再决定[抓狂]//@黑晶晶crystal: @细腿大羽...\n",
+       "42021       1                               [给力]感谢所有支持雯婕的芝麻！[爱你]\n",
+       "7777        1  2013最后一天，在新加坡开心度过，向所有的朋友们问声：新年快乐！2014年，我们会更好[调...\n",
+       "100399      0                  大中午出门办事找错路，曝晒中。要多杯具有多杯具。[泪][泪][汗]\n",
+       "82398       0                  马航还会否认吗？到底在隐瞒啥呢？[抓狂]//@头条新闻: 转发微博\n",
+       "106423      0                         克罗地亚球迷很爱放烟火！球又没进，就硝烟四起。[晕]\n",
+       "24798       1  [抱抱]福芦 TangRoulou 吉祥书 8.8折优惠 >>> http://t.cn/z...\n",
+       "6598        1  回复@钱旭明QXM:[嘻嘻][嘻嘻] //@钱旭明QXM:杨大哥[good][good][g...\n",
+       "53920       1                                   人家这脸长的!!!!!![哈哈]\n",
+       "15587       1  这个价不算高，和一天内训相比相差无几。。[哈哈]//@博通传媒v: 6个月！一个月工资1万，...\n",
+       "101237      0                             终于收工啦，脚丫子快冻掉了[泪][泪][泪]\n",
+       "82449       0  我决定从今天开始我想吃什么就去吃什么，一个人吃也无所谓，重点是不要因为别人的意见委屈了自己[...\n",
+       "32537       1  飘雪的北京 需要双份早餐.......//@美食天下: [哈哈]//@王淼Margay: 屁...\n",
+       "10630       1  [耶]，这个太赞了，生活大爆炸第六季马上要出啦[鼓掌] //@-郑瑜-:这个不错 //@经典...\n",
+       "85130       0  刚追完#倾世皇妃#，#千山暮雪#又紧随其后，网速和更新速度都太不给力，尽管我看过原著，还是焦...\n",
+       "105956      0  晚上看金二胖?察前?，推出的火炮基座?糟了，可以PK了[泪] //@艾米粒er: //@wi...\n",
+       "72391       0                    必须把中国足球的伟大，用我的职业演说出来 //@袁腾飞:[泪]\n",
+       "10761       1  [鼓掌] //@宁波香格里拉大酒店: 小编来答疑，周五晚惊艳全场的树根蛋糕到底有多长？蛋糕全..."
+      ]
+     },
+     "execution_count": 16,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "pd_all.sample(20)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.6.4"
+  },
+  "widgets": {
+   "state": {},
+   "version": "1.1.2"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
@@ -0,0 +1,107 @@
+import torch
+from transformers import GPT2ForSequenceClassification, BertTokenizer
+from peft import PeftModel
+import os
+import re
+
+def preprocess_text(text):
+    return text
+
+def main():
+    # 设置设备
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    print(f"使用设备: {device}")
+    
+    # 模型和权重路径
+    base_model_path = './models/gpt2-chinese'
+    lora_model_path = './best_weibo_sentiment_lora'
+    
+    print("加载模型和tokenizer...")
+    
+    # 检查LoRA模型是否存在
+    if not os.path.exists(lora_model_path):
+        print(f"错误: 找不到LoRA模型路径 {lora_model_path}")
+        print("请先运行 train.py 进行训练")
+        return
+    
+    # 加载tokenizer
+    try:
+        tokenizer = BertTokenizer.from_pretrained(base_model_path)
+        if tokenizer.pad_token is None:
+            tokenizer.pad_token = '[PAD]'
+    except Exception as e:
+        print(f"加载tokenizer失败: {e}")
+        print("请确保models/gpt2-chinese目录包含tokenizer文件")
+        return
+    
+    # 加载基础模型
+    try:
+        base_model = GPT2ForSequenceClassification.from_pretrained(
+            base_model_path, 
+            num_labels=2
+        )
+        base_model.config.pad_token_id = tokenizer.pad_token_id
+    except Exception as e:
+        print(f"加载基础模型失败: {e}")
+        print("请确保models/gpt2-chinese目录包含模型文件")
+        return
+    
+    # 加载LoRA权重
+    try:
+        model = PeftModel.from_pretrained(base_model, lora_model_path)
+        model.to(device)
+        model.eval()
+        print("LoRA模型加载成功!")
+    except Exception as e:
+        print(f"加载LoRA权重失败: {e}")
+        print("请确保LoRA权重文件存在且格式正确")
+        return
+    
+    print("\n============= 微博情感分析 (LoRA版) =============")
+    print("输入微博内容进行分析 (输入 'q' 退出):")
+    
+    while True:
+        text = input("\n请输入微博内容: ")
+        if text.lower() == 'q':
+            break
+        
+        if not text.strip():
+            print("输入不能为空，请重新输入")
+            continue
+        
+        try:
+            # 预处理文本
+            processed_text = preprocess_text(text)
+            
+            # 对文本进行编码
+            encoding = tokenizer(
+                processed_text,
+                max_length=128,
+                padding='max_length',
+                truncation=True,
+                return_tensors='pt'
+            )
+            
+            # 转移到设备
+            input_ids = encoding['input_ids'].to(device)
+            attention_mask = encoding['attention_mask'].to(device)
+            
+            # 预测
+            with torch.no_grad():
+                outputs = model(input_ids=input_ids, attention_mask=attention_mask)
+                logits = outputs.logits
+                probabilities = torch.softmax(logits, dim=1)
+                prediction = torch.argmax(probabilities, dim=1).item()
+            
+            # 输出结果
+            confidence = probabilities[0][prediction].item()
+            label = "正面情感" if prediction == 1 else "负面情感"
+            
+            print(f"预测结果: {label} (置信度: {confidence:.4f})")
+            
+        except Exception as e:
+            print(f"预测时发生错误: {e}")
+            continue
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,10 @@
+torch>=1.13.0
+transformers>=4.28.0
+peft>=0.4.0
+pandas>=1.5.0
+numpy>=1.21.0
+scikit-learn>=1.0.0
+tqdm>=4.64.0
+datasets>=2.0.0
+accelerate>=0.20.0
+safetensors>=0.3.0
@@ -0,0 +1,283 @@
+import os
+import random
+import numpy as np
+import pandas as pd
+import torch
+import torch.nn as nn
+from torch.utils.data import Dataset, DataLoader
+from transformers import (
+    GPT2ForSequenceClassification, 
+    BertTokenizer, 
+    get_linear_schedule_with_warmup,
+    TrainingArguments,
+    Trainer
+)
+from torch.optim import AdamW
+from sklearn.model_selection import train_test_split
+from sklearn.metrics import accuracy_score, f1_score
+from tqdm import tqdm
+
+# 导入PEFT库中的LoRA相关组件
+from peft import LoraConfig, TaskType, get_peft_model
+
+# 设置随机种子
+def set_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+set_seed(42)
+
+# 定义微博情感分析数据集
+class WeiboSentimentDataset(Dataset):
+    def __init__(self, reviews, labels, tokenizer, max_length=128):
+        self.reviews = reviews
+        self.labels = labels
+        self.tokenizer = tokenizer
+        self.max_length = max_length
+        
+    def __len__(self):
+        return len(self.reviews)
+    
+    def __getitem__(self, idx):
+        review = str(self.reviews[idx])
+        label = self.labels[idx]
+        
+        encoding = self.tokenizer(
+            review,
+            max_length=self.max_length,
+            padding='max_length',
+            truncation=True,
+            return_tensors='pt'
+        )
+        
+        return {
+            'input_ids': encoding['input_ids'].flatten(),
+            'attention_mask': encoding['attention_mask'].flatten(),
+            'labels': torch.tensor(label, dtype=torch.long)
+        }
+
+# 训练函数
+def train_model(model, train_dataloader, val_dataloader, optimizer, scheduler, device, epochs=3):
+    best_f1 = 0.0
+    
+    for epoch in range(epochs):
+        print(f"======== Epoch {epoch+1} / {epochs} ========")
+        model.train()
+        total_loss = 0
+        
+        # 训练循环
+        progress_bar = tqdm(train_dataloader, desc="Training", position=0, leave=True)
+        for batch in progress_bar:
+            # 将数据移到GPU
+            batch = {k: v.to(device) for k, v in batch.items()}
+            
+            # 清零梯度
+            optimizer.zero_grad()
+            
+            # 前向传播
+            outputs = model(
+                input_ids=batch['input_ids'],
+                attention_mask=batch['attention_mask'],
+                labels=batch['labels']
+            )
+            
+            loss = outputs.loss
+            total_loss += loss.item()
+            
+            # 反向传播
+            loss.backward()
+            
+            # 梯度裁剪，防止梯度爆炸
+            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+            
+            # 参数更新
+            optimizer.step()
+            scheduler.step()
+            
+            # 更新进度条
+            progress_bar.set_postfix({"loss": loss.item()})
+        
+        # 计算平均训练损失
+        avg_train_loss = total_loss / len(train_dataloader)
+        print(f"Average training loss: {avg_train_loss:.4f}")
+        
+        # 评估模型
+        val_metrics = evaluate_model(model, val_dataloader, device)
+        print(f"Validation Loss: {val_metrics['loss']:.4f}")
+        print(f"Validation Accuracy: {val_metrics['accuracy']:.4f}")
+        print(f"Validation F1 Score: {val_metrics['f1']:.4f}")
+        
+        # 保存最佳模型
+        if val_metrics['f1'] > best_f1:
+            best_f1 = val_metrics['f1']
+            # 保存LoRA权重
+            model.save_pretrained("./best_weibo_sentiment_lora")
+            print("Saved best LoRA model!")
+
+# 评估函数
+def evaluate_model(model, dataloader, device):
+    model.eval()
+    total_loss = 0
+    all_preds = []
+    all_labels = []
+    
+    with torch.no_grad():
+        for batch in tqdm(dataloader, desc="Evaluating"):
+            batch = {k: v.to(device) for k, v in batch.items()}
+            
+            outputs = model(
+                input_ids=batch['input_ids'],
+                attention_mask=batch['attention_mask'],
+                labels=batch['labels']
+            )
+            
+            loss = outputs.loss
+            total_loss += loss.item()
+            
+            # 获取预测结果
+            logits = outputs.logits
+            preds = torch.argmax(logits, dim=1).cpu().numpy()
+            labels = batch['labels'].cpu().numpy()
+            
+            all_preds.extend(preds)
+            all_labels.extend(labels)
+    
+    # 计算评估指标
+    accuracy = accuracy_score(all_labels, all_preds)
+    f1 = f1_score(all_labels, all_preds, average='macro')
+    avg_loss = total_loss / len(dataloader)
+    
+    return {
+        'loss': avg_loss,
+        'accuracy': accuracy,
+        'f1': f1
+    }
+
+def main():
+    # 设置模型本地保存路径
+    model_name = 'uer/gpt2-chinese-cluecorpussmall'
+    local_model_path = './models/gpt2-chinese'
+    
+    # 确保目录存在
+    os.makedirs(local_model_path, exist_ok=True)
+    os.makedirs('./best_weibo_sentiment_lora', exist_ok=True)
+    
+    # 加载数据集
+    print("加载微博情感数据集...")
+    df = pd.read_csv('dataset/weibo_senti_100k.csv')
+    
+    # 分割数据集
+    train_df, val_df = train_test_split(df, test_size=0.1, random_state=42, stratify=df['label'])
+    
+    # 加载tokenizer
+    print("加载预训练模型和tokenizer...")
+    
+    # 检查本地是否已有模型
+    if os.path.exists(os.path.join(local_model_path, 'config.json')):
+        print(f"从本地路径加载tokenizer: {local_model_path}")
+        tokenizer = BertTokenizer.from_pretrained(local_model_path)
+    else:
+        print(f"从Hugging Face下载tokenizer到: {local_model_path}")
+        tokenizer = BertTokenizer.from_pretrained(model_name, cache_dir=local_model_path)
+        # 保存tokenizer到本地
+        tokenizer.save_pretrained(local_model_path)
+    
+    # 设置padding token
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = '[PAD]'
+    
+    # 记录pad_token的ID
+    pad_token_id = tokenizer.pad_token_id
+    
+    # 创建数据集
+    train_dataset = WeiboSentimentDataset(
+        train_df['review'].values,
+        train_df['label'].values,
+        tokenizer
+    )
+    
+    val_dataset = WeiboSentimentDataset(
+        val_df['review'].values,
+        val_df['label'].values,
+        tokenizer
+    )
+    
+    # 创建数据加载器
+    train_dataloader = DataLoader(train_dataset, batch_size=16, shuffle=True)
+    val_dataloader = DataLoader(val_dataset, batch_size=16)
+    
+    # 设置设备
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    print(f"使用设备: {device}")
+    
+    # 加载预训练的GPT2模型
+    print("加载GPT2模型...")
+    if (os.path.exists(os.path.join(local_model_path, 'pytorch_model.bin')) or 
+        os.path.exists(os.path.join(local_model_path, 'model.safetensors'))):
+        print(f"从本地路径加载模型权重: {local_model_path}")
+        model = GPT2ForSequenceClassification.from_pretrained(local_model_path, num_labels=2)
+    else:
+        print(f"从Hugging Face下载模型权重到: {local_model_path}")
+        # 直接从Hugging Face下载并保存完整模型
+        model = GPT2ForSequenceClassification.from_pretrained(model_name, num_labels=2)
+        model.save_pretrained(local_model_path)
+    
+    # 确保模型使用与tokenizer相同的pad_token_id
+    model.config.pad_token_id = pad_token_id
+    
+    # 配置LoRA参数
+    print("配置LoRA参数...")
+    lora_config = LoraConfig(
+        task_type=TaskType.SEQ_CLS,  # 序列分类任务
+        target_modules=["c_attn", "c_proj"],  # GPT2的注意力投影层
+        inference_mode=False,  # 训练模式
+        r=8,  # LoRA秩，控制可训练参数数量
+        lora_alpha=32,  # LoRA alpha参数，缩放因子
+        lora_dropout=0.1,  # LoRA Dropout
+    )
+    
+    # 将模型转换为PEFT格式的LoRA模型
+    print("创建LoRA模型...")
+    model = get_peft_model(model, lora_config)
+    model.print_trainable_parameters()  # 打印可训练参数数量和占比
+    
+    model.to(device)
+    
+    # 设置优化器和学习率调度器
+    print("设置优化器...")
+    optimizer = AdamW(
+        model.parameters(),  # PEFT会自动处理参数筛选
+        lr=5e-4,  # LoRA通常使用较高的学习率
+        eps=1e-8
+    )
+    
+    # 设置总训练步数和warmup步数
+    total_steps = len(train_dataloader) * 3  # 3个epoch
+    warmup_steps = int(total_steps * 0.1)  # 10%的warmup
+    
+    scheduler = get_linear_schedule_with_warmup(
+        optimizer,
+        num_warmup_steps=warmup_steps,
+        num_training_steps=total_steps
+    )
+    
+    # 训练模型
+    print("开始训练...")
+    train_model(
+        model=model,
+        train_dataloader=train_dataloader,
+        val_dataloader=val_dataloader,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        device=device,
+        epochs=3
+    )
+    
+    print("训练完成!")
+    print("LoRA权重已保存到: ./best_weibo_sentiment_lora/")
+
+if __name__ == "__main__":
+    main()