部署散记1（增强）

拉取modelscope镜像

sudo docker pull registry.cn-hangzhou.aliyuncs.com/modelscope-repo/modelscope:ubuntu20.04-cuda11.3.0-py37-torch1.11.0-tf1.15.5-1.1.2

拉取增强代码

git clone https://gitee.com/binghai228/Bringing-Old-Photos-Back-to-Life.git
cd Bringing-Old-Photos-Back-to-Life

安装增强模块：

cd Face_Enhancement/models/networks/
git clone https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
cp -rf Synchronized-BatchNorm-PyTorch/sync_batchnorm .
cd ../../../

cd Global/detection_models
git clone https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
cp -rf Synchronized-BatchNorm-PyTorch/sync_batchnorm .
cd ../../

cd Face_Detection/
wget http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
bzip2 -d shape_predictor_68_face_landmarks.dat.bz2
cd ../

cd Face_Enhancement/
wget https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life/releases/download/v1.0/face_checkpoints.zip
unzip face_checkpoints.zip
cd ../
cd Global/
wget https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life/releases/download/v1.0/global_checkpoints.zip
unzip global_checkpoints.zip
cd ../

创建容器：

sudo nvidia-docker run -dit --net=host --name beauty --shm-size="1g" -v $PWD:/beauty registry.cn-hangzhou.aliyuncs.com/modelscope-repo/modelscope:ubuntu20.04-cuda11.3.0-py37-torch1.11.0-tf1.15.5-1.1.2 bash

进入容器：

sudo docker exec -it beauty bash
cd beauty

安装依赖：

pip install -r requirements.txt

测试修复：

python run.py --input_folder test \
              --output_folder result \
              --GPU 0 \
              --with_scratch \
              --HR

从官网下载redis稳定版到项目根目录下。

tar -zxvf redis-7.0.7.tar.gz

在dockers环境中编译redis：

cd redis-7.0.7
make

编译完成后安装redis：

cd src
make install

启动redis：

redis-server &

测试连接：

redis-cli
redis 127.0.0.1:6379> ping

正常输出如下：

PONG

想要查看所有键值：

keys *

删除某个键值:

del key1

断开连接：

exit

其他相关命令如下：

redis-server stop
redis-server start
redis-server restart

安装数据库驱动：

pip install redis

安装fastapi:

pip install fastapi -i https://mirror.baidu.com/pypi/simple
pip install "uvicorn[standard]" -i https://mirror.baidu.com/pypi/simple
pip install python-multipart -i https://mirror.baidu.com/pypi/simple

后台代码：

import numpy as np
import cv2
import base64
from fastapi import FastAPI, File, UploadFile, Depends
import time
import uuid
import os
from fastapi import Request
import json
from redis import Redis,ConnectionPool

# 导入自定义函数
from tool import base64_encode_image,base64_decode_image 

# 开发前端界面
from fastapi.staticfiles import StaticFiles
from fastapi.templating import Jinja2Templates
from fastapi.responses import HTMLResponse
from fastapi.middleware.cors import CORSMiddleware


# 定义fastapi的启动app
app = FastAPI()

# 解除跨域访问限制
origins = ["*"]
app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# 定义Redis数据库依赖注入
def get_rdb():
    pool = ConnectionPool(host='127.0.0.1',port = 6379)
    rdb = Redis(connection_pool=pool)
    try:
        yield rdb
    finally:
        rdb.close()
        

# 定义访问接口
@app.post("/beauty/")
async def get_image(request: Request,file: UploadFile =File(...), db:Redis=Depends(get_rdb)):
    print("接收到数据")
    
    # 接收前端上传的图片
    imgdata = await file.read()
    imgdata = np.frombuffer(imgdata, np.uint8)
    image = cv2.imdecode(imgdata, cv2.IMREAD_COLOR)
    h,w,c = image.shape

    # 图像存储进redis
    k = str(uuid.uuid4())
    in_data = {"in_img": base64_encode_image(image.astype('float32')),"height":h,"width":w,"channel":c}
    db.set(k, json.dumps(in_data))
    db.expire(k, 50) # 设置失效时间
    
    # 等待返回
    time_out_num = 0
    while True:
        # 尝试获取结果
        output = db.get(k)
        output = json.loads(output.decode("utf-8"))
        if output.get('out_img') is not None:
            #取出结果
            im = base64_decode_image(output["out_img"],"float32",(output.get('height'), output.get('width'),output.get('channel')))
            
            #封装结果返回
            db.delete(k)
            _, buffer_img = cv2.imencode('.jpg', im)
            img64 = base64.b64encode(buffer_img)
            img64 = str(img64, encoding='utf-8')
            print("完成")
            return {"state":1 ,"img": img64}

        time.sleep(1)

        #设置超时机制
        time_out_num = time_out_num + 1
        if (time_out_num > 50):
            db.delete(k)
            print('超时')
            return {"state":-1 ,"code": '当前拥挤，请稍候再试'}


# 挂载静态资源目录
app.mount("/static", StaticFiles(directory="static"), name="static")

# 定义页面模板库位置
templates = Jinja2Templates(directory="templates")

# 定义首页
@app.get("/", response_class=HTMLResponse)
async def home(request: Request):
    return templates.TemplateResponse("index.html", {"request": request})

# if __name__ == "__main__":
#     '''程序启动入口'''
#     import uvicorn
#     uvicorn.run(app, host="0.0.0.0", port=8041)

启动fastapi:

uvicorn main_fastapi:app --host 0.0.0.0 --port 8041 &

其中tool.py文件函数如下：

import numpy as np
import base64

def base64_encode_image(a):
    """将numpy数组进行base64编码"""
    return base64.b64encode(a).decode("utf-8")
 
 
def base64_decode_image(a, dtype, shape):
    """base64转图像"""
    a = bytes(a, encoding="utf-8")
    a = np.frombuffer(base64.decodestring(a), dtype=dtype)
    a = a.reshape(shape)
    return a

模型循环推理代码：

#导入第三方库
import os
import numpy as np
import json
import cv2
import base64
import time
import redis

#导入自定义库
from tool import base64_encode_image,base64_decode_image 

# 导入推理脚本
# from run import inference
from modelscope.outputs import OutputKeys
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
 

if __name__ == "__main__":
    #参数设置
    db = redis.StrictRedis(host="localhost", port=6379, db=0)
    
    skin_retouching = pipeline(Tasks.skin_retouching,model='damo/cv_unet_skin-retouching')

    print('美颜模型开始运行')
    
    #循环处理
    while True:
        # 获取有效键值
        keys = db.keys()
        isfinder = False
        for k in keys:
            output = db.get(k)
            if output is None:
                continue
            output = json.loads(output.decode("utf-8")) 
                
            if output.get('out_img') is None:
                isfinder = True
                in_img = base64_decode_image(output["in_img"],"float32",(output.get('height'), output.get('width'),output.get('channel')))
                break
            else:
                continue
        
        if not isfinder:
            time.sleep(0.2)
            continue
        
        # 解析图像   
        cv2.imwrite('./in/test.jpg',in_img)
        
        # 开始推理
        #inference()
        result = skin_retouching('./in/test.jpg')
        cv2.imwrite('./out/result.png', result[OutputKeys.OUTPUT_IMG])
        
        # 读取推理结果
        out_img = cv2.imread('./out/result.png',cv2.IMREAD_COLOR)
    
        # 将结果写入redis
        out_data = {"out_img": base64_encode_image(out_img.astype('float32'))}
        output.update(out_data)
        db.set(k, json.dumps(output))
        
        print("完成推理")

        time.sleep(0.2)

本地测试脚本client_beauty.py：

# 导入依赖库
import numpy as np
import requests
import cv2
import base64
 
# 定义http接口
url = "http://127.0.0.1:8041/beauty"
 
# 发送请求
files = {'file':open('1.jpg','rb'),}
result = requests.post(url=url, files=files)
 
# 解析返回值
result = result.json()
 
# 解码返回的图像数据
img = result["img"]
img = bytes(img, encoding='utf-8')  # str转bytes
img = base64.b64decode(img)  # base64解码
img = np.asarray(bytearray(img), dtype="uint8")
img = cv2.imdecode(img, cv2.IMREAD_UNCHANGED)
 
# 保存图像到本地
if img is None:
    print('调用失败')
else:
    cv2.imwrite('result.png',img)
    print('完成')