Wlkr.SafePaddleOCR
1.0.2307.2307
dotnet add package Wlkr.SafePaddleOCR --version 1.0.2307.2307
NuGet\Install-Package Wlkr.SafePaddleOCR -Version 1.0.2307.2307
<PackageReference Include="Wlkr.SafePaddleOCR" Version="1.0.2307.2307" />
paket add Wlkr.SafePaddleOCR --version 1.0.2307.2307
#r "nuget: Wlkr.SafePaddleOCR, 1.0.2307.2307"
// Install Wlkr.SafePaddleOCR as a Cake Addin #addin nuget:?package=Wlkr.SafePaddleOCR&version=1.0.2307.2307 // Install Wlkr.SafePaddleOCR as a Cake Tool #tool nuget:?package=Wlkr.SafePaddleOCR&version=1.0.2307.2307
Wlkr.Core.ThreadUtils
项目背景
早在PaddleOCR 2.2版本时期,认识了周杰大佬的PaddleSharp项目,试用其中PaddleOCR时,发现它在改为web api调用时会报错,大概意思是OCR实例的内存只能由其创建的线程才具有访问权限,于是就有了本项目的雏形。
潜伏于大佬Q群中很长时间,这个问题更是老生常谈。虽然后来大佬实现了基于BlockingCollection
的线程安全示例,不过估计因为README全是英文,还是出现了很多星际玩家。
食用方式
项目中的SafeThreadRunner,为了实现更直观的调用方式(var res = ocr.run(mat)
),使用了3个信号量SemaphoreSlim
实现了线程安全的轮询方法,它们的作用分别是否空闲,唤醒线程,返回结果。
SafeThreadRunner<Cls, In, Out>
,可以从泛型的名字猜测,Cls对应OCR的实例(如All、Rec、Det等任务),In
为输入即Mat,Out
为输出即Restful规范的返回结果RestResult<Out>
。
核心代码
代码很简单,下面这段代码,通过信号量负责检查线程是否空闲。如果空闲, 则设置入参,唤醒线程。
public RestResult<Out> Run(In src)
{
//是否空闲
safeSrcSlim.Wait();
//设置Source
Source = src;
//恢复线程,运行runFunc
safeRunSlim.Release();
//等待runFunc结果
safeResSlim.Wait();
//释放信号量,设为空闲
safeSrcSlim.Release();
return Result;
}
唤醒后则执行识别,告诉调用者识别完成,输出结果。(Dispose同理)
private void RunByThread()
{
using Cls cls = initFunc();
while (true)
{
safeRunSlim.Wait();
if (IsDisposed)
return;
try
{
Result = runFunc(cls, Source);
}
catch (Exception ex)
{
Result = new RestResult<Out>()
{
code = "500",
msg = ex.Message
};
}
finally
{
safeResSlim.Release();
}
}
}
nuget安装参考命令
# 新建一个console项目
dotnet new console
# 添加nuget包
dotnet add package Wlkr.SafePaddleOCR
CPU加速示例
本项目实现的SafePaddleOCR为PaddleOcrAll开启Mkldnn的实例,使用方式如下:
//Warmup
SafePaddleOCR safePaddleOCR = new SafePaddleOCR();
string imgPath = @"../../../../vx_images/DimTechStudio-Logo.png";
var res = safePaddleOCR.Run(imgPath);
Console.Write(@"res: {res.data.Text}");
定制示例
如需要定制自己的线程安全实例,可参考:
//实例的初始化方法
Func<PaddleOcrAll> initFuc = () =>
{
Action<PaddleConfig> device = PaddleDevice.Mkldnn();
var poa = new PaddleOcrAll(LocalFullModels.ChineseV3, device)
{
Enable180Classification = true,
AllowRotateDetection = true,
};
return poa;
};
//实例的执行方法
Func<PaddleOcrAll, Mat, RestResult<PaddleOcrResult>> mthdFunc = (cls, source) =>
{
var res = cls.Run(source);
return new RestResult<PaddleOcrResult>(res);
};
//声明
SafeThreadRunner<PaddleOcrAll, Mat, PaddleOcrResult> safeThreadRunner = new SafeThreadRunner<PaddleOcrAll, Mat, PaddleOcrResult>(OCRFactory.BuildAllWithMkldnn, OCRFactory.RunAll);
//运行
string imgPath = @"../../../../vx_images/DimTechStudio-Logo.png";
using var mat = Cv2.ImRead(filePath, ImreadModes.AnyColor);
var res = safeThreadRunner.Run(mat);
SemaphoreSlim
与BlockingCollection
对比
- 单实例测试:性能几乎一样,没有明显差异
- 多实例测试:报错!!!
测试用的机器是笔记本 CPU R7 5800H,内存32G。
两种方式均会报错,实例数越多,报错概率越高,错误提示依然内存错误的问题。
System.AccessViolationException: Attempted to read or write protected memory. This is often an indication that other memory is corrupt.
另外SemaphoreSlim的方式比BlockingCollection的方式出现的更频繁,尤其在4实例时基本无法完成10240次OCR测试。
两者在出现代码位置也不经相同,Det、Cls、Rec三种模型预测时均可能错误。
- 周杰大佬的项目优势:实现生产者消费者模式
- 本项目优势:单实例Dispose
由于我也重构过周杰大佬的QueuedPaddleOcrAll.cs,其Dispose方式只能释放所有实例。虽然我增加了动态添加/删除实例的功能,但其使用了Task作为轮询的载体,Task不能像Thread那样有真正意义上的取消动作。
CancellationToken
实现的取消最大缺陷是在阻塞时是无效的。即便我实现了取消,它也必须从blockingCollection.GetConsumingEnumerable()获取到消息执行一次OCR识别,才能释放OCR实例,极端情况下等于无法是释放。
而本项目使用了SemaphoreSlim
,执行Dispose时只要线程是空闲即可触发OCR实例的释放。
测试数据
- 硬件配置: CPU R7 5800H(8核16线程主频3.2GHz),内存32G
- 测试图片:10张,数字0000~0009,宽高160*80,类型png
- 风扇转速常开最大,排除CPU温度影响
- OCR实例参数:All,CPU,Mkldnn
由于10240次大概率报错,无法完成测试,这里改为256次。
平均毫秒 | 平均毫秒 | 平均毫秒 | |
---|---|---|---|
OCR次数 | 256 | 256 | 256 |
实例数 | 1 | 2 | 4 |
SemaphoreSlim | 27.36328125 | 17.640625 | 12.6328125 |
BlockingCollection | 27.74609375 | 17.8984375 | 12.640625 |
- 思路转换
从单进程4实例,改为4进程1实例测试,测试了1次没有报错,每个进程10240次,平均50ms。
- 测试缺陷:
- 居然没用web api来测试而是用了console测试,与项目背景背道而驰……
- 由于用的图片较少,没法作为内存压力测试的参考
总结
基于思路转换的测试,虽然测试次数少了点,不过目前来看,当作为web api时,1进程1实例,多开进程,利用负载均衡来提高并发和服务器利用率,为最优方案。
Author Info
DimWalker ©2023 广州市增城区黯影信息科技部 https://www.dimtechstudio.com/
Product | Versions Compatible and additional computed target framework versions. |
---|---|
.NET | net6.0 is compatible. net6.0-android was computed. net6.0-ios was computed. net6.0-maccatalyst was computed. net6.0-macos was computed. net6.0-tvos was computed. net6.0-windows was computed. net7.0 was computed. net7.0-android was computed. net7.0-ios was computed. net7.0-maccatalyst was computed. net7.0-macos was computed. net7.0-tvos was computed. net7.0-windows was computed. net8.0 was computed. net8.0-android was computed. net8.0-browser was computed. net8.0-ios was computed. net8.0-maccatalyst was computed. net8.0-macos was computed. net8.0-tvos was computed. net8.0-windows was computed. |
-
net6.0
- OpenCvSharp4 (>= 4.8.0.20230708)
- OpenCvSharp4.runtime.win (>= 4.8.0.20230708)
- Sdcb.PaddleInference (>= 2.4.1.4)
- Sdcb.PaddleInference.runtime.win64.mkl (>= 2.4.1)
- Sdcb.PaddleOCR (>= 2.6.0.5)
- Sdcb.PaddleOCR.Models.LocalV3 (>= 2.6.0.5)
- Wlkr.Core.ThreadUtils (>= 1.0.2307.2307)
NuGet packages
This package is not used by any NuGet packages.
GitHub repositories
This package is not used by any popular GitHub repositories.
Version | Downloads | Last updated |
---|---|---|
1.0.2307.2307 | 570 | 7/23/2023 |