nli-MiniLM2-L6-H768实战教程：将零样本分类嵌入Flask后端服务的完整代码示例-深圳市維司達科技有限公司

nli-MiniLM2-L6-H768实战教程：将零样本分类嵌入Flask后端服务的完整代码示例

1. 模型简介

nli-MiniLM2-L6-H768是一个轻量级的自然语言推理(NLI)模型，特别适合处理文本关系判断任务。与生成式模型不同，它的核心能力是分析两段文本之间的逻辑关系，主要判断以下三种关系：

矛盾(contradiction)：两段文本表达的意思相互冲突
蕴含(entailment)：一段文本可以从另一段文本中推导出来
中立(neutral)：两段文本相关但不能直接推导

这个768维的小模型在保持轻量级的同时，能够高效完成以下任务：

文本对语义相似度计算
零样本文本分类(无需训练数据)
搜索结果重排序
问答匹配度评估

2. 环境准备与快速部署

2.1 基础环境要求

在开始之前，请确保你的开发环境满足以下要求：

Python 3.7+
PyTorch 1.8+
Transformers库
Flask框架
CUDA环境(如需GPU加速)

2.2 安装依赖包

pip install torch transformers flask flask-cors

2.3 下载模型

模型可以通过Hugging Face直接加载：

from transformers import AutoModelForSequenceClassification, AutoTokenizer model_name = "cross-encoder/nli-MiniLM2-L6-H768" model = AutoModelForSequenceClassification.from_pretrained(model_name) tokenizer = AutoTokenizer.from_pretrained(model_name)

3. 核心功能实现

3.1 文本对打分功能

这是模型最基础的功能，用于判断两段文本之间的关系：

def score_text_pair(text_a, text_b): features = tokenizer([text_a], [text_b], padding=True, truncation=True, return_tensors="pt") with torch.no_grad(): scores = model(**features).logits # 转换为概率分布 probabilities = torch.softmax(scores, dim=1) return { "contradiction": probabilities[0][0].item(), "entailment": probabilities[0][1].item(), "neutral": probabilities[0][2].item(), "predicted_label": model.config.id2label[torch.argmax(scores).item()] }

3.2 零样本分类实现

基于NLI的零样本分类是模型的亮点功能：

def zero_shot_classification(text, candidate_labels): # 将标签转换为假设语句 hypothesis_template = "This example is about {}." pairs = [(text, hypothesis_template.format(label)) for label in candidate_labels] # 批量编码 features = tokenizer(pairs, padding=True, truncation=True, return_tensors="pt") with torch.no_grad(): scores = model(**features).logits # 只取entailment分数 entailment_scores = scores[:, 1] # 归一化处理 normalized_scores = torch.softmax(entailment_scores, dim=0) results = [] for label, score in zip(candidate_labels, normalized_scores): results.append({"label": label, "score": score.item()}) # 按分数降序排列 results.sort(key=lambda x: x["score"], reverse=True) return { "best_label": results[0]["label"], "scores": results }

4. Flask服务集成

4.1 基础服务框架

创建一个完整的Flask应用来封装这些功能：

from flask import Flask, request, jsonify from flask_cors import CORS app = Flask(__name__) CORS(app) # 允许跨域请求 @app.route('/health', methods=['GET']) def health_check(): return jsonify({"status": "healthy"}) @app.route('/score', methods=['POST']) def score_api(): data = request.json text_a = data.get('text_a', '') text_b = data.get('text_b', '') result = score_text_pair(text_a, text_b) return jsonify(result) @app.route('/zero_shot', methods=['POST']) def zero_shot_api(): data = request.json text = data.get('text', '') labels = data.get('labels', []) result = zero_shot_classification(text, labels) return jsonify(result) if __name__ == '__main__': app.run(host='0.0.0.0', port=5000)

4.2 性能优化技巧

在实际部署中，我们可以通过以下方式优化服务性能：

启用批处理：修改代码支持同时处理多个请求
GPU加速：确保模型在CUDA设备上运行
缓存机制：对常见查询结果进行缓存
输入长度限制：设置合理的最大文本长度

优化后的批处理版本：

def batch_score_text_pairs(text_pairs): # 解压文本对 texts_a, texts_b = zip(*text_pairs) # 批量编码 features = tokenizer(list(texts_a), list(texts_b), padding=True, truncation=True, return_tensors="pt").to(model.device) with torch.no_grad(): scores = model(**features).logits probabilities = torch.softmax(scores, dim=1) results = [] for i in range(len(text_pairs)): results.append({ "contradiction": probabilities[i][0].item(), "entailment": probabilities[i][1].item(), "neutral": probabilities[i][2].item(), "predicted_label": model.config.id2label[torch.argmax(scores[i]).item()] }) return results

5. 实际应用案例

5.1 新闻分类系统

使用零样本分类功能构建一个新闻分类系统：

# 示例新闻文本 news_article = """ Apple has unveiled its latest iPhone model featuring a revolutionary camera system and improved battery life. The new device will be available starting next month. """ # 定义候选类别 categories = [ "technology", "sports", "politics", "entertainment", "health" ] # 进行分类 classification_result = zero_shot_classification(news_article, categories) print(f"这篇文章最可能属于: {classification_result['best_label']}")

5.2 智能客服问答匹配

使用文本对打分功能评估用户问题与预设答案的匹配度：

# 用户问题 user_question = "How do I reset my password?" # 知识库中的候选答案 knowledge_base = [ "Password reset instructions are sent to your email", "You can change password in account settings", "Contact support for password issues" ] # 评估每个答案的匹配度 for answer in knowledge_base: score = score_text_pair(user_question, answer) print(f"答案: {answer}") print(f"匹配度: {score['entailment']:.3f}") print("---")

6. 服务部署与监控

6.1 使用Gunicorn生产部署

对于生产环境，建议使用Gunicorn作为WSGI服务器：

gunicorn -w 4 -b 0.0.0.0:5000 app:app

6.2 添加Prometheus监控

集成Prometheus客户端监控API性能：

from prometheus_client import start_http_server, Counter, Histogram # 定义指标 REQUEST_COUNT = Counter( 'request_count', 'App Request Count', ['method', 'endpoint', 'http_status'] ) REQUEST_LATENCY = Histogram( 'request_latency_seconds', 'Request latency', ['endpoint'] ) # 修改Flask路由添加监控 @app.route('/score', methods=['POST']) def score_api(): start_time = time.time() REQUEST_COUNT.labels('POST', '/score', '200').inc() data = request.json text_a = data.get('text_a', '') text_b = data.get('text_b', '') result = score_text_pair(text_a, text_b) REQUEST_LATENCY.labels('/score').observe(time.time() - start_time) return jsonify(result)