#2025/7/12 #milvus的upsert方法在多线程调用时多产生重复记录，修正为先删除在新增

shidong

2025-07-12 a7e7050585325bdb873c6d312ea89de94215e11e

 qwen_thread.py

@@ -17,55 +17,42 @@


class qwen_thread:
    def __init__(self, max_workers,config,model_path):
        self.executor = ThreadPoolExecutor(max_workers=max_workers)
        self.semaphore = threading.Semaphore(max_workers)
        self.max_workers = max_workers
    def __init__(self, config,logger):
        self.config = config
        self.max_workers = int(config.get("threadnum"))
        self.executor = ThreadPoolExecutor(max_workers=int(config.get("threadnum")))
        self.semaphore = threading.Semaphore(int(config.get("threadnum")))
        self.logger = logger

        # 初始化Milvus集合
        connections.connect("default", host=config.get("milvusurl"), port=config.get("milvusport"))
        # 加载集合
        self.collection = Collection(name="smartobject")
        self.collection.load()

        self.config = config
        self.model_pool = []
        self.lock_pool = [threading.Lock() for _ in range(max_workers)]
        for i in range(max_workers):
        self.lock_pool = [threading.Lock() for _ in range(int(config.get("threadnum")))]
        for i in range(int(config.get("threadnum"))):
            model = AutoModelForVision2Seq.from_pretrained(
                model_path,
                device_map="cuda:1",
                config.get("qwenaddr"),
                device_map=f"cuda:{config.get('cuda')}",
                trust_remote_code=True,
                use_safetensors=True,
                torch_dtype=torch.float16

            ).eval()
            self.model_pool.append(model)

        # 共享的处理器 (线程安全)
        self.processor = AutoProcessor.from_pretrained(model_path,use_fast=True)
        self.processor = AutoProcessor.from_pretrained(config.get("qwenaddr"), use_fast=True)


        # 创建实例专属logger
        self.logger = logging.getLogger(f"{self.__class__}_{id(self)}")
        self.logger.setLevel(logging.INFO)
        # 避免重复添加handler
        if not self.logger.handlers:
            handler = RotatingFileHandler(
                filename=os.path.join("logs", 'thread_log.log'),
                maxBytes=10 * 1024 * 1024,
                backupCount=3,
                encoding='utf-8'
            )
            formatter = logging.Formatter(
                '%(asctime)s - %(filename)s:%(lineno)d - %(funcName)s() - %(levelname)s: %(message)s'
            )
            handler.setFormatter(formatter)
            self.logger.addHandler(handler)

    def submit(self,res_a):
        # 尝试获取信号量（非阻塞）
        acquired = self.semaphore.acquire(blocking=False)

        if not acquired:
            self.logger.info(f"线程池已满，等待空闲线程... (当前活跃: {self.max_workers - self.semaphore._value}/{self.max_workers})")
            #self.logger.info(f"线程池已满，等待空闲线程... (当前活跃: {self.max_workers - self.semaphore._value}/{self.max_workers})")
            # 阻塞等待直到有可用线程
            self.semaphore.acquire(blocking=True)

@@ -73,32 +60,46 @@
        future.add_done_callback(self._release_semaphore)
        return future

    def _wrap_task(self, res):
    def _wrap_task(self, res_a):
        try:
            #self.logger.info(f"处理: { res['id']}开始")
            current_time = datetime.now()
            image_id = self.tark_do(res, self.config.get("ragurl"), self.config.get("ragmode"), self.config.get("max_tokens"))
            self.logger.info(f"处理: { res['id']}完毕{image_id}:{datetime.now() - current_time}")
            return image_id
            self.tark_do(res_a, self.config.get("ragurl"), self.config.get("ragmode"), self.config.get("max_tokens"))
        except Exception as e:
            self.logger.info(f"任务 { res['id']} 处理出错: {e}")
            self.logger.info(f"处理出错: {e}")
            raise

    def tark_do(self,res,ragurl,rag_mode,max_tokens):
        try :
        try:
            # 1. 从集合A获取向量和元数据
            is_waning = 0
            image_des = self.image_desc(f"{res['image_desc_path']}")
            self.logger.info(image_des)
            if image_des:
                # rule_text = self.get_rule(ragurl)
                is_waning = self.image_rule_chat(image_des,res['waning_value'],ragurl,rag_mode,max_tokens)
            is_desc = 2

            # 生成图片描述
            ks_time = datetime.now()
            desc = self.image_desc(res)
            desc_time = datetime.now() - ks_time
            current_time = datetime.now()
            risk_description = ""
            suggestion = ""
            # 图片描述生成成功
            if desc:
                is_desc = 2
                # 调用规则匹配方法,判断是否预警
                is_waning = self.image_rule_chat(desc, res['waning_value'], ragurl,rag_mode,max_tokens)
                # 如果预警,则生成隐患描述和处理建议
                if is_waning == 1:
                    # 获取规章制度数据
                    filedata = self.get_filedata(res['waning_value'],res['suggestion'], ragurl)
                    # 生成隐患描述
                    risk_description = self.image_rule_chat_with_detail(filedata, res['waning_value'], ragurl,rag_mode,max_tokens)
                    # 生成处理建议
                    suggestion = self.image_rule_chat_suggestion(filedata, res['waning_value'], ragurl,rag_mode,max_tokens)
                    self.logger.info(
                        f"{res['video_point_id']}执行完毕：{res['id']}:是否预警{is_waning},安全隐患：{risk_description}\n处理建议：{suggestion}")
            else:
                is_waning = 0
                is_desc = 3

            # 数据组
            data = {
                "id": res['id'],
                "event_level_id": res['event_level_id'],  # event_level_id
                "event_level_name": res['event_level_name'],  # event_level_id
                "rule_id": res["rule_id"],
@@ -106,7 +107,7 @@
                "video_point_name": res['video_point_name'],
                "is_waning": is_waning,
                "is_desc": is_desc,
                "zh_desc_class": image_des,  # text_vector
                "zh_desc_class": desc,  # text_vector
                "bounding_box": res['bounding_box'],  # bounding_box
                "task_id": res['task_id'],  # task_id
                "task_name": res['task_name'],  # task_id
@@ -117,35 +118,38 @@
                "image_path": res['image_path'],  # image_path
                "image_desc_path": res['image_desc_path'],  # image_desc_path
                "video_path": res['video_path'],
                "text_vector": res['text_vector']
                "text_vector": res['text_vector'],
                "risk_description": risk_description,
                "suggestion": suggestion,
                "knowledge_id": res['knowledge_id']
            }
            self.collection.delete(f"id == {res['id']}")
            # 保存到milvus
            image_id = self.collection.upsert(data).primary_keys
            if is_desc == 2:
                data = {
                    "id": str(image_id[0]),
                    "video_point_id": res['video_point_id'],
                    "video_path": res["video_point_name"],
                    "zh_desc_class": image_des,
                    "detect_time": res['detect_time'],
                    "image_path": f"{res['image_path']}",
                    "task_name": res["task_name"],
                    "event_level_name": res["event_level_name"],
                    "rtsp_address": f"{res['video_path']}"
                }
                # 调用rag
                asyncio.run(self.insert_json_data(ragurl, data))
            return image_id
            image_id = self.collection.insert(data).primary_keys
            data = {
                "id": str(image_id[0]),
                "video_point_id": res['video_point_id'],
                "video_path": res["video_point_name"],
                "zh_desc_class": desc,
                "detect_time": res['detect_time'],
                "image_path": f"{res['image_path']}",
                "task_name": res["task_name"],
                "event_level_name": res["event_level_name"],
                "rtsp_address": f"{res['video_path']}"
            }
            # 调用rag
            asyncio.run(self.insert_json_data(ragurl, data))
            rag_time = datetime.now() - current_time
            self.logger.info(f"{res['video_point_id']}执行完毕：{image_id}运行结束总体用时:{datetime.now() - ks_time},图片描述用时{desc_time}，RAG用时{rag_time}")
        except Exception as e:
            self.logger.info(f"线程：执行模型解析时出错:任务：{e}")
            self.logger.info(f"线程：执行模型解析时出错::{e}")
            return 0

    def image_desc(self, image_path):
    def image_desc(self, res_data):
        try:
            model, lock = self._acquire_model()
            # 2. 处理图像
            image = Image.open(image_path).convert("RGB")  # 替换为您的图片路径
            image = image.resize((600, 600), Image.Resampling.LANCZOS)  # 高质量缩放

            image = Image.open(res_data['image_desc_path']).convert("RGB").resize((600, 600), Image.Resampling.LANCZOS)
            messages = [
                {
                    "role": "user",
@@ -167,15 +171,9 @@
                padding=True,
                return_tensors="pt",
            )
            inputs = inputs.to("cuda:1")
            current_time = datetime.now()
            outputs = model.generate(**inputs,
                                               max_new_tokens=300,
                                               do_sample=True,
                                               temperature=0.7,
                                               renormalize_logits=True
                                               )
            print(f"处理完毕:{datetime.now() - current_time}")
            inputs = inputs.to(model.device)
            with torch.inference_mode():
                outputs = model.generate(**inputs,max_new_tokens=200)
            generated_ids = outputs[:, len(inputs.input_ids[0]):]
            image_text = self.processor.batch_decode(
                generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
@@ -222,7 +220,6 @@
        try:
            content = (
                f"图片描述内容为：\n{image_des}\n规则内容：\n{rule_text}。\n请验证图片描述中是否有符合规则的内容，不进行推理和think。返回结果格式为[xxx符合的规则id]，如果没有返回[]")
            # self.logger.info(content)
            #self.logger.info(len(content))
            search_data = {
                "prompt": "",
@@ -241,7 +238,6 @@
            }
            response = requests.post(ragurl + "/chat", json=search_data)
            results = response.json().get('data')
            #self.logger.info(len(results))
            # self.logger.info(results)
            ret = re.sub(r'<think>.*?</think>', '', results, flags=re.DOTALL)
            ret = ret.replace(" ", "").replace("\t", "").replace("\n", "")
@@ -252,6 +248,99 @@
        except Exception as e:
            self.logger.info(f"线程：执行规则匹配时出错:{image_des, rule_text, ragurl, rag_mode,e}")
            return None

    # 隐患描述
    def image_rule_chat_with_detail(self,filedata, rule_text, ragurl, rag_mode,max_tokens):
        # API调用
        content = (
            f"规章制度为：[{filedata}]\n违反内容为：[{rule_text}]\n请查询违反内容在规章制度中的安全隐患，不进行推理和think，返回简短的文字信息")
        # self.logger.info(len(content))
        search_data = {
            "prompt": "",
            "messages": [
                {
                    "role": "user",
                    "content": content
                }
            ],
            "llm_name": rag_mode,
            "stream": False,
            "gen_conf": {
                "temperature": 0.7,
                "max_tokens": max_tokens
            }
        }
        response = requests.post(ragurl + "/chat", json=search_data)
        # 从json提取data字段内容
        ret = response.json()["data"]
        # 移除<think>标签和内容
        ret = re.sub(r'<think>.*?</think>', '', ret, flags=re.DOTALL)
        # 字符串清理,移除空格,制表符,换行符,星号
        ret = ret.replace(" ", "").replace("\t", "").replace("\n", "").replace("**","")
        #print(f"安全隐患:{ret}")
        return ret
    #处理建议
    def image_rule_chat_suggestion(self,filedata, rule_text, ragurl, rag_mode,max_tokens):
        # 请求内容
        content = (
            f"规章制度为：[{filedata}]\n违反内容为：[{rule_text}]\n请查询违反内容在规章制度中的处理建议，不进行推理和think，返回简短的文字信息")
        response = requests.post(
            # ollama地址
            url=f"{ragurl}/chat",
            json={
                # 指定模型
                "llm_name": rag_mode,
                "messages": [
                    {"role": "user", "content": content}
                ],
                "stream": False,  # 关闭流式输出
                "gen_conf": {
                    "temperature": 0.7,
                    "max_tokens": max_tokens
                }
            }
        )
        # 从json提取data字段内容
        ret = response.json()["data"]
        # 移除<think>标签和内容
        ret = re.sub(r'<think>.*?</think>', '', ret, flags=re.DOTALL)
        # 字符串清理,移除空格,制表符,换行符,星号
        ret = ret.replace(" ", "").replace("\t", "").replace("\n", "").replace("**","")
        #print(f"处理建议:{ret}")
        return ret

    # RAG服务发送请求,获取知识库内容
    def get_filedata(self, searchtext,filter_expr, ragurl):
        search_data = {
            # 知识库集合
            "collection_name": "smart_knowledge",
            # 查询文本
            "query_text": searchtext,
            # 搜索模式
            "search_mode": "hybrid",
            # 最多返回结果
            "limit": 10,
            # 调密向量搜索权重
            "weight_dense": 0.9,
            # 稀疏向量搜索权重
            "weight_sparse": 0.1,
            # 空字符串
            "filter_expr": f"docnm_kwd in {filter_expr}",
            # 只返回 text 字段
            "output_fields": ["text"]
        }
        #print(search_data)
        # 向 ragurl + "/search" 端点发送POST请求
        response = requests.post(ragurl + "/search", json=search_data)
        # 从响应中获取'results'字段
        results = response.json().get('results')
        # 初始化 text
        text = ""
        # 遍历所有结果规则(rule)，将每条规则的'entity'中的'text'字段取出.
        for rule in results:
            text = text + rule['entity'].get('text') + ";\n"
        #print(text)
        return text

    async def insert_json_data(self, ragurl, data):
        try:
@@ -264,7 +353,7 @@

    def _release_semaphore(self, future):
        self.semaphore.release()
        self.logger.info(f"释放线程 (剩余空闲: {self.semaphore._value}/{self.max_workers})")
        #self.logger.info(f"释放线程 (剩余空闲: {self.semaphore._value}/{self.max_workers})")

    def shutdown(self):
        """安全关闭"""