import time from concurrent.futures import ThreadPoolExecutor import threading import torch from PIL import Image from pymilvus import connections, Collection from datetime import datetime import requests import asyncio import re from qwen_vl_utils import process_vision_info from transformers import AutoModelForVision2Seq, AutoProcessor class qwen_thread: 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() if config.get('cuda') == None or config.get('cuda') == '0': self.device = f"cuda" else: self.device = f"cuda:{config.get('cuda')}" self.model_pool = [] self.lock_pool = [threading.Lock() for _ in range(int(config.get("threadnum")))] for i in range(int(config.get("threadnum"))): model = AutoModelForVision2Seq.from_pretrained( config.get("qwenaddr"), device_map=self.device, trust_remote_code=True, use_safetensors=True, torch_dtype=torch.float16 ).eval() model = model.to(self.device) self.model_pool.append(model) # 共享的处理器 (线程安全) self.processor = AutoProcessor.from_pretrained(config.get("qwenaddr"), use_fast=True) 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})") # 阻塞等待直到有可用线程 self.semaphore.acquire(blocking=True) future = self.executor.submit(self._wrap_task, res_a) future.add_done_callback(self._release_semaphore) return future def _wrap_task(self, res_a): try: 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"处理出错: {e}") raise def tark_do(self,res,ragurl,rag_mode,max_tokens): try: # 1. 从集合A获取向量和元数据 is_waning = 0 is_desc = 2 # 生成图片描述 ks_time = datetime.now() desc_time = datetime.now() - ks_time current_time = datetime.now() risk_description = "" suggestion = "" # 调用规则匹配方法,判断是否预警 is_waning = self.image_rule(res) # 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}") # 数据组 data = { "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"], "video_point_id": res['video_point_id'], # video_point_id "video_point_name": res['video_point_name'], "is_waning": is_waning, "is_desc": 1, "zh_desc_class": res['zh_desc_class'], # text_vector "bounding_box": res['bounding_box'], # bounding_box "task_id": res['task_id'], # task_id "task_name": res['task_name'], # task_id "detect_id": res['detect_id'], # detect_id "detect_time": res['detect_time'], # detect_time "detect_num": res['detect_num'], "waning_value": res['waning_value'], "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'], "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.insert(data).primary_keys res['id'] = image_id[0] # 图片描述生成成功 desc = self.image_desc(res) if desc: is_desc = 2 else: is_desc = 3 # 数据组 data = { "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"], "video_point_id": res['video_point_id'], # video_point_id "video_point_name": res['video_point_name'], "is_waning": is_waning, "is_desc": is_desc, "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 "detect_id": res['detect_id'], # detect_id "detect_time": res['detect_time'], # detect_time "detect_num": res['detect_num'], "waning_value": res['waning_value'], "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'], "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.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}") if is_waning == 1: self.logger.info(f"{res['video_point_id']}执行完毕：{image_id},图片描述：{desc}\n隐患：{risk_description}\n建议：{suggestion}") except Exception as e: self.logger.info(f"线程：执行模型解析时出错::{e}") return 0 def image_desc(self, res_data): try: model, lock = self._acquire_model() image = Image.open(res_data['image_desc_path']).convert("RGB").resize((600, 600), Image.Resampling.LANCZOS) messages = [ { "role": "user", "content": [ { "type": "image", }, {"type": "text", "text": "请详细描述图片中的目标信息及特征。返回格式为整段文字描述"}, ], } ] # Preparation for inference text = self.processor.apply_chat_template( messages, add_generation_prompt=True ) inputs = self.processor( text=[text], images=[image], padding=True, return_tensors="pt", ) inputs = inputs.to(model.device) with torch.inference_mode(), torch.amp.autocast(device_type=self.device, dtype=torch.float16): outputs = model.generate(**inputs,max_new_tokens=200,do_sample=False,num_beams=1,temperature=None,top_p=None,top_k=1,use_cache=True,repetition_penalty=1.0) 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 ) image_des = (image_text[0]).strip() #self.logger.info(f"{res_data['video_point_id']}:{res_data['id']}:{res_data['detect_time']}:{image_des}") return image_des except Exception as e: self.logger.info(f"线程：执行图片描述时出错:{e}") finally: # 4. 释放模型 self._release_model(model) torch.cuda.empty_cache() def image_rule(self, res_data): try: model, lock = self._acquire_model() image = Image.open(res_data['image_desc_path']).convert("RGB").resize((600, 600), Image.Resampling.LANCZOS) messages = [ { "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": f"图片中是否有{res_data['waning_value']}?请回答yes或no"}, ], } ] # Preparation for inference text = self.processor.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) image_inputs, video_inputs = process_vision_info(messages) inputs = self.processor( text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt", ) inputs = inputs.to(model.device) with torch.no_grad(): outputs = model.generate(**inputs, max_new_tokens=10) 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 ) image_des = (image_text[0]).strip() return image_des except Exception as e: self.logger.info(f"线程：执行图片描述时出错:{e}") finally: # 4. 释放模型 self._release_model(model) torch.cuda.empty_cache() def get_rule(self,ragurl): try: rule_text = None search_data = { "collection_name": "smart_rule", "query_text": "", "search_mode": "hybrid", "limit": 100, "weight_dense": 0.7, "weight_sparse": 0.3, "filter_expr": "", "output_fields": ["text"] } response = requests.post(ragurl + "/search", json=search_data) results = response.json().get('results') rule_text = "" ruleid = 1 for rule in results: if rule['score'] >= 0: rule_text = rule_text + str(ruleid) + ". " + rule['entity'].get('text') + ";\n" ruleid = ruleid + 1 # self.logger.info(len(rule_text)) else: self.logger.info(f"线程：执行获取规则时出错:{response}") return rule_text except Exception as e: self.logger.info(f"线程：执行获取规则时出错:{e}") return None def image_rule_chat(self, image_des,rule_text, ragurl, rag_mode,max_tokens): 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不符合规则的内容，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(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) results = response.json().get('data') ret = re.sub(r'.*?', '', results, flags=re.DOTALL) ret = ret.replace(" ", "").replace("\t", "").replace("\n", "") #self.logger.info(f"{rule_text}:{ret}") is_waning = 0 if len(ret) > 2: is_waning = 1 return is_waning 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 } } #self.logger.info(content) response = requests.post(ragurl + "/chat", json=search_data) # 从json提取data字段内容 ret = response.json()["data"] # 移除标签和内容 ret = re.sub(r'.*?', '', 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"] # 移除标签和内容 ret = re.sub(r'.*?', '', 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: data = {'collection_name': "smartrag", "data": data, "description": ""} requests.post(ragurl + "/insert_json_data", json=data, timeout=(0.3, 0.3)) #self.logger.info(f"调用录像服务:{ragurl, data}") except Exception as e: #self.logger.info(f"{self._thread_name}线程：调用录像时出错:地址：{ragurl}：{e}") return 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})") def shutdown(self): """安全关闭""" self.executor.shutdown(wait=False) for model in self.model_pool: del model torch.cuda.empty_cache() def _acquire_model(self): """从池中获取一个空闲模型 (简单轮询)""" while True: for i, (model, lock) in enumerate(zip(self.model_pool, self.lock_pool)): if lock.acquire(blocking=False): return model, lock time.sleep(0.1) # 避免CPU空转 def _release_model(self, model): """释放模型回池""" for i, m in enumerate(self.model_pool): if m == model: self.lock_pool[i].release() break def remove_duplicate_lines(self,text): seen = set() result = [] for line in text.split('。'): # 按句号分割 if line.strip() and line not in seen: seen.add(line) result.append(line) return '。'.join(result) def remove_duplicate_lines_d(self,text): seen = set() result = [] for line in text.split(','): # 按句号分割 if line.strip() and line not in seen: seen.add(line) result.append(line) return '。'.join(result) def remove_duplicate_lines_n(self,text): seen = set() result = [] for line in text.split('\n'): # 按句号分割 if line.strip() and line not in seen: seen.add(line) result.append(line) return '。'.join(result)