于BERT的中文问答系统12

主要改进点

日志配置：

确保日志文件按日期和时间生成，便于追踪不同运行的记录。
数据处理：

增加了对数据加载过程中错误的捕获和日志记录，确保程序能够跳过无效数据并继续运行。
模型训练：

增加了重新训练模型的功能，用户可以选择重新训练现有模型或从头开始训练。
用户交互：

增加了输入验证，确保用户输入的问题不为空。
增加了模糊匹配功能，支持部分输入问题的匹配。
错误处理：

在关键步骤增加了异常捕获和日志记录，提高了程序的健壮性。

import os
import json
import jsonlines
import torch
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from transformers import BertModel, BertTokenizer
import tkinter as tk
from tkinter import filedialog, messagebox
import logging
from difflib import SequenceMatcher
from datetime import datetime# 配置日志
LOGS_DIR = os.path.join(PROJECT_ROOT, 'logs')
os.makedirs(LOGS_DIR, exist_ok=True)def setup_logging():log_file = os.path.join(LOGS_DIR, datetime.now().strftime('%Y-%m-%d/%H-%M-%S/羲和.txt'))os.makedirs(os.path.dirname(log_file), exist_ok=True)logging.basicConfig(level=logging.INFO,format='%(asctime)s - %(levelname)s - %(message)s',handlers=[logging.FileHandler(log_file),logging.StreamHandler()])# 获取项目根目录
PROJECT_ROOT = os.path.dirname(os.path.abspath(__file__))
setup_logging()# 数据集类
class XihuaDataset(Dataset):def __init__(self, file_path, tokenizer, max_length=128):self.tokenizer = tokenizerself.max_length = max_lengthself.data = self.load_data(file_path)def load_data(self, file_path):data = []if file_path.endswith('.jsonl'):with jsonlines.open(file_path) as reader:for i, item in enumerate(reader):try:data.append(item)except jsonlines.jsonlines.InvalidLineError as e:logging.warning(f"跳过无效行 {i + 1}: {e}")elif file_path.endswith('.json'):with open(file_path, 'r') as f:try:data = json.load(f)except json.JSONDecodeError as e:logging.warning(f"跳过无效文件 {file_path}: {e}")return datadef __len__(self):return len(self.data)def __getitem__(self, idx):item = self.data[idx]question = item['question']human_answer = item['human_answers'][0]chatgpt_answer = item['chatgpt_answers'][0]try:inputs = self.tokenizer(question, return_tensors='pt', padding='max_length', truncation=True, max_length=self.max_length)human_inputs = self.tokenizer(human_answer, return_tensors='pt', padding='max_length', truncation=True, max_length=self.max_length)chatgpt_inputs = self.tokenizer(chatgpt_answer, return_tensors='pt', padding='max_length', truncation=True, max_length=self.max_length)except Exception as e:logging.warning(f"跳过无效项 {idx}: {e}")return self.__getitem__((idx + 1) % len(self.data))return {'input_ids': inputs['input_ids'].squeeze(),'attention_mask': inputs['attention_mask'].squeeze(),'human_input_ids': human_inputs['input_ids'].squeeze(),'human_attention_mask': human_inputs['attention_mask'].squeeze(),'chatgpt_input_ids': chatgpt_inputs['input_ids'].squeeze(),'chatgpt_attention_mask': chatgpt_inputs['attention_mask'].squeeze(),'human_answer': human_answer,'chatgpt_answer': chatgpt_answer}# 获取数据加载器
def get_data_loader(file_path, tokenizer, batch_size=8, max_length=128):dataset = XihuaDataset(file_path, tokenizer, max_length)return DataLoader(dataset, batch_size=batch_size, shuffle=True)# 模型定义
class XihuaModel(torch.nn.Module):def __init__(self, pretrained_model_name='F:/models/bert-base-chinese'):super(XihuaModel, self).__init__()self.bert = BertModel.from_pretrained(pretrained_model_name)self.classifier = torch.nn.Linear(self.bert.config.hidden_size, 1)def forward(self, input_ids, attention_mask):outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)pooled_output = outputs.pooler_outputlogits = self.classifier(pooled_output)return logits# 训练函数
def train(model, data_loader, optimizer, criterion, device):model.train()total_loss = 0.0for batch in data_loader:try:input_ids = batch['input_ids'].to(device)attention_mask = batch['attention_mask'].to(device)human_input_ids = batch['human_input_ids'].to(device)human_attention_mask = batch['human_attention_mask'].to(device)chatgpt_input_ids = batch['chatgpt_input_ids'].to(device)chatgpt_attention_mask = batch['chatgpt_attention_mask'].to(device)optimizer.zero_grad()human_logits = model(human_input_ids, human_attention_mask)chatgpt_logits = model(chatgpt_input_ids, chatgpt_attention_mask)human_labels = torch.ones(human_logits.size(0), 1).to(device)chatgpt_labels = torch.zeros(chatgpt_logits.size(0), 1).to(device)loss = criterion(human_logits, human_labels) + criterion(chatgpt_logits, chatgpt_labels)loss.backward()optimizer.step()total_loss += loss.item()except Exception as e:logging.warning(f"跳过无效批次: {e}")return total_loss / len(data_loader)# 主训练函数
def main_train(retrain=False):device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')logging.info(f'Using device: {device}')tokenizer = BertTokenizer.from_pretrained('F:/models/bert-base-chinese')model = XihuaModel(pretrained_model_name='F:/models/bert-base-chinese').to(device)if retrain:model.load_state_dict(torch.load(os.path.join(PROJECT_ROOT, 'models/xihua_model.pth'), map_location=device, weights_only=True))optimizer = optim.Adam(model.parameters(), lr=1e-5)criterion = torch.nn.BCEWithLogitsLoss()train_data_loader = get_data_loader(os.path.join(PROJECT_ROOT, 'data/train_data.jsonl'), tokenizer, batch_size=8, max_length=128)num_epochs = 5for epoch in range(num_epochs):train_loss = train(model, train_data_loader, optimizer, criterion, device)logging.info(f'Epoch [{epoch+1}/{num_epochs}], Loss: {train_loss:.4f}')torch.save(model.state_dict(), os.path.join(PROJECT_ROOT, 'models/xihua_model.pth'))logging.info("模型训练完成并保存")# GUI界面
class XihuaChatbotGUI:def __init__(self, root):self.root = rootself.root.title("羲和聊天机器人")self.tokenizer = BertTokenizer.from_pretrained('F:/models/bert-base-chinese')self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')self.model = XihuaModel(pretrained_model_name='F:/models/bert-base-chinese').to(self.device)self.load_model()self.model.eval()# 加载训练数据集以便在获取答案时使用self.data = self.load_data(os.path.join(PROJECT_ROOT, 'data/train_data.jsonl'))self.create_widgets()def create_widgets(self):self.question_label = tk.Label(self.root, text="问题:")self.question_label.pack()self.question_entry = tk.Entry(self.root, width=50)self.question_entry.pack()self.answer_button = tk.Button(self.root, text="获取回答", command=self.get_answer)self.answer_button.pack()self.answer_label = tk.Label(self.root, text="回答:")self.answer_label.pack()self.answer_text = tk.Text(self.root, height=10, width=50)self.answer_text.pack()self.train_button = tk.Button(self.root, text="训练模型", command=self.train_model)self.train_button.pack()self.retrain_button = tk.Button(self.root, text="重新训练模型", command=lambda: self.train_model(retrain=True))self.retrain_button.pack()def get_answer(self):question = self.question_entry.get()if not question:messagebox.showwarning("输入错误", "请输入问题")returninputs = self.tokenizer(question, return_tensors='pt', padding='max_length', truncation=True, max_length=128)with torch.no_grad():input_ids = inputs['input_ids'].to(self.device)attention_mask = inputs['attention_mask'].to(self.device)logits = self.model(input_ids, attention_mask)if logits.item() > 0:answer_type = "人类回答"else:answer_type = "ChatGPT回答"specific_answer = self.get_specific_answer(question, answer_type)self.answer_text.delete(1.0, tk.END)self.answer_text.insert(tk.END, f"{answer_type}\n{specific_answer}")def get_specific_answer(self, question, answer_type):# 使用模糊匹配查找最相似的问题best_match = Nonebest_ratio = 0.0for item in self.data:ratio = SequenceMatcher(None, question, item['question']).ratio()if ratio > best_ratio:best_ratio = ratiobest_match = itemif best_match:if answer_type == "人类回答":return best_match['human_answers'][0]else:return best_match['chatgpt_answers'][0]return "未找到具体答案"def load_data(self, file_path):data = []if file_path.endswith('.jsonl'):with jsonlines.open(file_path) as reader:for i, item in enumerate(reader):try:data.append(item)except jsonlines.jsonlines.InvalidLineError as e:logging.warning(f"跳过无效行 {i + 1}: {e}")elif file_path.endswith('.json'):with open(file_path, 'r') as f:try:data = json.load(f)except json.JSONDecodeError as e:logging.warning(f"跳过无效文件 {file_path}: {e}")return datadef load_model(self):model_path = os.path.join(PROJECT_ROOT, 'models/xihua_model.pth')if os.path.exists(model_path):self.model.load_state_dict(torch.load(model_path, map_location=self.device, weights_only=True))logging.info("加载现有模型")else:logging.info("没有找到现有模型，将使用预训练模型")def train_model(self, retrain=False):file_path = filedialog.askopenfilename(filetypes=[("JSONL files", "*.jsonl"), ("JSON files", "*.json")])if not file_path:messagebox.showwarning("文件选择错误", "请选择一个有效的数据文件")returntry:dataset = XihuaDataset(file_path, self.tokenizer)data_loader = DataLoader(dataset, batch_size=8, shuffle=True)# 加载已训练的模型权重if retrain:self.model.load_state_dict(torch.load(os.path.join(PROJECT_ROOT, 'models/xihua_model.pth'), map_location=self.device, weights_only=True))self.model.to(self.device)self.model.train()optimizer = torch.optim.Adam(self.model.parameters(), lr=1e-5)criterion = torch.nn.BCEWithLogitsLoss()num_epochs = 5for epoch in range(num_epochs):train_loss = train(self.model, data_loader, optimizer, criterion, self.device)logging.info(f'Epoch [{epoch+1}/{num_epochs}], Loss: {train_loss:.4f}')torch.save(self.model.state_dict(), os.path.join(PROJECT_ROOT, 'models/xihua_model.pth'))logging.info("模型训练完成并保存")messagebox.showinfo("训练完成", "模型训练完成并保存")except Exception as e:logging.error(f"模型训练失败: {e}")messagebox.showerror("训练失败", f"模型训练失败: {e}")# 主函数
if __name__ == "__main__":# 启动GUIroot = tk.Tk()app = XihuaChatbotGUI(root)root.mainloop()