import gradio as gr import torch from transformers import AutoTokenizer, AutoModelForSequenceClassification import numpy as np import time import json class ChineseSentimentClassifier: def __init__(self): self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') self.model = None self.tokenizer = None self.labels = ['负面', '中性', '正面'] # 三分类：对应label2id = {"negative": 0, "neutral": 1, "positive": 2} self.loaded = False def load_model(self): """加载模型""" try: # 方法1：尝试正常加载 self.tokenizer = AutoTokenizer.from_pretrained('./chinese_sentiment_model') self.model = AutoModelForSequenceClassification.from_pretrained('./chinese_sentiment_model') self.model.to(self.device) self.model.eval() self.loaded = True return "✅ 模型加载成功！" except Exception as e1: try: # 方法2：尝试使用safe_tensors加载 self.tokenizer = AutoTokenizer.from_pretrained('./chinese_sentiment_model') self.model = AutoModelForSequenceClassification.from_pretrained( './chinese_sentiment_model', use_safetensors=False # 强制不使用safetensors ) self.model.to(self.device) self.model.eval() self.loaded = True return "✅ 模型加载成功（兼容模式）！" except Exception as e2: try: # 方法3：尝试从预训练模型重新初始化 self.tokenizer = AutoTokenizer.from_pretrained('hfl/chinese-roberta-wwm-ext') self.model = AutoModelForSequenceClassification.from_pretrained( 'hfl/chinese-roberta-wwm-ext', num_labels=3 ) self.model.to(self.device) self.model.eval() self.loaded = True return "⚠️ 使用预训练模型（未微调版本），建议重新上传训练好的模型文件" except Exception as e3: return f"❌ 所有加载方法都失败：\n1. {str(e1)}\n2. {str(e2)}\n3. {str(e3)}" def predict_sentiment(self, text): """情感预测""" if not self.loaded: return "❌ 模型未加载，请先加载模型" if not text.strip(): return "❌ 请输入有效文本" try: # 文本预处理 inputs = self.tokenizer( text, return_tensors='pt', max_length=512, truncation=True, padding=True ) inputs = {k: v.to(self.device) for k, v in inputs.items()} # 模型推理 with torch.no_grad(): outputs = self.model(**inputs) predictions = torch.nn.functional.softmax(outputs.logits, dim=-1) predicted_class = torch.argmax(predictions, dim=-1).item() confidence = predictions[0][predicted_class].item() # 获取所有类别的概率 probs = predictions[0].cpu().numpy() result = { 'predicted_label': self.labels[predicted_class], 'confidence': confidence, 'all_scores': {self.labels[i]: float(probs[i]) for i in range(len(self.labels))} } return result except Exception as e: return f"❌ 预测失败：{str(e)}" # 初始化分类器 classifier = ChineseSentimentClassifier() def format_prediction_result(result): """格式化预测结果""" if isinstance(result, str): return result if isinstance(result, dict) and 'predicted_label' in result: # 构建结果显示 pred_label = result['predicted_label'] confidence = result['confidence'] all_scores = result['all_scores'] # 情感emoji emoji_map = {"正面": "😊", "中性": "😐", "负面": "😞"} emoji = emoji_map.get(pred_label, "🤔") output = f""" ## {emoji} 预测结果 **情感类别**: {pred_label} **置信度**: {confidence:.4f} ({confidence*100:.2f}%) ### 📊 详细分数： """ for label, score in all_scores.items(): bar_length = int(score * 20) # 简单的文本进度条 bar = "█" * bar_length + "░" * (20 - bar_length) output += f"- **{label}**: {score:.4f} `{bar}` {score*100:.2f}%\n" # 置信度解释 if confidence > 0.8: confidence_desc = "🔥 高置信度" elif confidence > 0.6: confidence_desc = "✅ 中等置信度" else: confidence_desc = "⚠️ 低置信度" output += f"\n**置信度评估**: {confidence_desc}" return output return str(result) def predict_interface(text): """预测接口""" if not text.strip(): return "请输入要分析的文本内容" start_time = time.time() result = classifier.predict_sentiment(text) end_time = time.time() formatted_result = format_prediction_result(result) # 添加性能信息 if isinstance(result, dict): inference_time = (end_time - start_time) * 1000 formatted_result += f"\n\n⚡ **推理时间**: {inference_time:.2f}ms" return formatted_result def batch_predict_interface(texts): """批量预测接口""" if not texts.strip(): return "请输入要分析的文本内容（每行一个）" lines = [line.strip() for line in texts.split('\n') if line.strip()] if not lines: return "请输入有效的文本内容" results = [] start_time = time.time() for i, text in enumerate(lines[:10]): # 限制最多10条 result = classifier.predict_sentiment(text) if isinstance(result, dict): label = result['predicted_label'] confidence = result['confidence'] emoji_map = {"正面": "😊", "中性": "😐", "负面": "😞"} emoji = emoji_map.get(label, "🤔") results.append(f"{i+1}. {emoji} {text[:50]}... → **{label}** ({confidence:.3f})") else: results.append(f"{i+1}. ❌ {text[:50]}... → 预测失败") end_time = time.time() total_time = (end_time - start_time) * 1000 output = "## 📋 批量预测结果\n\n" + "\n".join(results) output += f"\n\n⚡ **总处理时间**: {total_time:.2f}ms | **平均时间**: {total_time/len(lines):.2f}ms/条" return output def load_model_interface(): """加载模型接口""" return classifier.load_model() # 示例文本 - 对应三分类 EXAMPLE_TEXTS = { "正面评价": "这部电影真的很好看，演员表演很棒，剧情吸引人！", "负面评价": "剧情很差，浪费时间，难看透顶，浪费钱。", "中性评价": "一般般，不太喜欢，还行，没有特别惊喜。", "复杂情感": "虽然有些地方不太满意，但总体来说还可以接受。" } # 创建Gradio界面 with gr.Blocks( title="🎭 中文情感分类系统", theme=gr.themes.Soft(), css=""" .gradio-container {max-width: 1400px; margin: auto;} .example-box {background: #f0f0f0; padding: 10px; border-radius: 5px; margin: 5px 0;} """ ) as demo: gr.Markdown(""" # 🎭 中文情感分类系统演示 **基于BERT的中文情感分析模型** - 展示预训练模型微调和部署能力 🔥 **技术特点**：BERT微调 + 高精度分类 + 实时推理 + 批量处理 """) # 模型加载区域 with gr.Row(): with gr.Column(scale=1): gr.Markdown("### 🚀 模型管理") load_btn = gr.Button("📥 加载情感分类模型", variant="primary", size="lg") load_status = gr.Textbox(label="📊 加载状态", value="模型未加载", interactive=False) with gr.Column(scale=1): gr.Markdown("### 📈 模型信息") gr.Markdown(""" - **基础模型**: BERT-base-chinese - **任务类型**: 二分类情感分析 - **训练数据**: 中文评论数据集 - **输出类别**: 正面 / 负面 """) gr.Markdown("---") # 主要功能区域 with gr.Row(): with gr.Column(scale=1): gr.Markdown("### 🔍 单文本情感分析") text_input = gr.Textbox( label="📝 输入文本", placeholder="输入要分析情感的中文文本...", lines=4, max_lines=10 ) predict_btn = gr.Button("🎯 分析情感", variant="primary") # 示例文本按钮 gr.Markdown("**📚 快速测试样例**：") with gr.Row(): for label, text in EXAMPLE_TEXTS.items(): example_btn = gr.Button(label, variant="secondary", size="sm") example_btn.click( lambda t=text: t, outputs=[text_input] ) with gr.Column(scale=1): gr.Markdown("### 📊 预测结果") result_output = gr.Markdown( label="分析结果", value="等待输入文本进行情感分析..." ) gr.Markdown("---") # 批量处理区域 with gr.Row(): with gr.Column(): gr.Markdown("### 📋 批量情感分析") batch_input = gr.Textbox( label="📝 批量输入（每行一个文本）", placeholder="今天天气真好\n这部电影太无聊了\n产品质量不错\n服务态度很差", lines=6 ) batch_btn = gr.Button("🔄 批量分析", variant="primary") with gr.Column(): gr.Markdown("### 📈 批量结果") batch_output = gr.Markdown( value="等待批量输入..." ) # 技术详情折叠面板 with gr.Accordion("🔧 技术实现详情", open=False): gr.Markdown(""" ### 📋 技术栈详情 **模型架构**： - 基础模型：chinese-roberta-wwm-ext (哈工大版本) - 微调任务：三分类情感分析 (正面/中性/负面) - 优化器：AdamW + 学习率调度 - 损失函数：CrossEntropyLoss **数据处理**： - 文本长度：最大512 tokens - 预处理：分词 + 填充/截断 - 后处理：Softmax概率输出 **性能优化**： - 推理加速：torch.no_grad() - 批量处理：支持批量推理 - 设备适配：自动GPU/CPU切换 ### 🎯 项目亮点 1. **完整的微调流程**：从数据预处理到模型训练部署 2. **工程化实现**：错误处理、性能监控、用户友好界面 3. **可扩展设计**：支持多分类、批量处理、模型热更新 4. **部署优化**：适配Hugging Face Spaces环境限制 """) # 事件绑定 load_btn.click( fn=load_model_interface, outputs=[load_status] ) predict_btn.click( fn=predict_interface, inputs=[text_input], outputs=[result_output] ) text_input.submit( fn=predict_interface, inputs=[text_input], outputs=[result_output] ) batch_btn.click( fn=batch_predict_interface, inputs=[batch_input], outputs=[batch_output] ) # 启动应用 if __name__ == "__main__": demo.launch()