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 os
import requests
import asyncio
import logging
import re
from logging.handlers import RotatingFileHandler
from transformers import AutoModelForVision2Seq, AutoProcessor
class qwen_thread_batch:
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
# 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):
model = AutoModelForVision2Seq.from_pretrained(
model_path,
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)
# 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})")
# 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_a,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
desc_list = self.image_desc(res_a)
risk_description = ""
suggestion = ""
if desc_list:
for desc, res in zip(desc_list, res_a):
# 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'], ragurl)
# 生成隐患描述
risk_description = self.image_rule_chat_with_detail(filedata, res['waning_value'], ragurl, rag_mode)
# 生成处理建议
suggestion = self.image_rule_chat_suggestion(res['waning_value'], ragurl, rag_mode)
else:
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"],
"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']
}
# 保存到milvus
image_id = self.collection.upsert(data).primary_keys
logging.info(image_id)
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))
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_data = []
# 1. 并行加载图像
def _load_image(path):
return Image.open(path).convert("RGB").resize((448, 448), Image.Resampling.LANCZOS)
with ThreadPoolExecutor(max_workers=4) as executor:
image_data = list(executor.map(
_load_image,
[res['image_desc_path'] for res in res_data]
))
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] * len(image_data),
images=[image_data],
padding=True,
return_tensors="pt",
)
inputs = inputs.to(model.device)
with torch.inference_mode():
outputs = model.generate(**inputs,max_new_tokens=100)
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 = []
for text in image_text:
image_des.append(text)
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')
# self.logger.info(results)
ret = re.sub(r'.*?', '', results, flags=re.DOTALL)
ret = ret.replace(" ", "").replace("\t", "").replace("\n", "")
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, ollama_url, ollama_mode="qwen2.5vl:3b"):
# API调用
response = requests.post(
# ollama地址
url=f"{ollama_url}/chat",
json={
"prompt":"",
# 请求内容
"messages": [
{
"role": "user",
"content": f"请根据规章制度[{filedata}]\n查找[{rule_text}]的安全隐患描述，不进行推理和think。返回信息小于800字"
}
],
# 指定模型
"llm_name": "qwen3:8b",
"stream": False, # 关闭流式输出
"gen_conf": {
"temperature": 0.7, # 控制生成随机性
"max_tokens": 800 # 最大输出长度
}
}
)
# 从json提取data字段内容
ret = response.json()["data"]
#result = response.json()
#ret = result.get("data") or result.get("message", {}).get("content", "")
# 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(ret)
return ret
#处理建议
def image_rule_chat_suggestion(self, rule_text, ollama_url, ollama_mode="qwen2.5vl:3b"):
self.logger.info("----------------------------------------------------------------")
# 请求内容
content = (
f"请根据违规内容[{rule_text}]\n进行返回处理违规建议，不进行推理和think。返回精准信息")
response = requests.post(
# ollama地址
url=f"{ollama_url}/chat",
json={
# 指定模型
"llm_name": "qwen3:8b",
"messages": [
{"role": "user", "content": content}
],
"stream": False # 关闭流式输出
}
)
# 从json提取data字段内容
ret = response.json()["data"]
#result = response.json()
#ret = result.get("data") or result.get("message", {}).get("content", "")
# 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(ret)
return ret
# RAG服务发送请求,获取知识库内容
def get_filedata(self, searchtext, 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": 100,
# 调密向量搜索权重
"weight_dense": 0.7,
# 稀疏向量搜索权重
"weight_sparse": 0.3,
# 空字符串
"filter_expr": "",
# 只返回 text 字段
"output_fields": ["text"]
}
# 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"
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)