ncdhz / yolo-libtorch

一个 LibTorch 版本的 Yolo 封装库

文档

Yolo-LibTorch旨在使用LibTorch轻松调用Yolo的各种版本YoloV5、YoloV6、YoloV7和YoloV8等（以下结果通过yolov8n模型预测）。此项目需要torchscript的模型格式。

快速开始

Windows 中使用（cmake）

1. 下载所需环境

   • cmake
   • libtorch
   • opencv
   • visual studio需要它提供的c++环境

2. 配置环境变量

   # 添加libtorch的lib目录到path
   \path\to\libtorch\lib
   # 添加opencv的bin目录到环境变量
   # vc14 vc15 vc16 都可以
   \path\to\opencv\build\x64\vc16\bin

3. 修改根目录下 CMakeLists.txt 文件

   # Torch_DIR 路径为 libtorch 下面的 share/cmake/Torch/
   set(Torch_DIR /path/to/libtorch/share/cmake/Torch/)
   # OpenCV_DIR 路径为 opencv/build/
   set(OpenCV_DIR /path/to/opencv/build/)

4. 编译

   mkdir build
   cd build
   cmake ..
   
   # 根据你 LibTorch 版本选择执行下面语句
   # Debug
   cmake --build . --config Debug
   cmake --install . --config Debug
   
   # Release
   cmake --build . --config Release
   cmake --install . --config Release

5. 测试

   cd bin
   # 默认会打开电脑摄像头捕捉画面
   # 如果电脑没有摄像头可以通过参数引入视频或图片
   test.exe

yolo-libtorch's Issues

讀取pt時報錯

您好，麻煩一下
利用cpu運行時沒問題
使用gpu運行至讀取pt檔的時候發生以下報錯(有使用gpu進行export torchscript檔案)

0x00007FFA55ECCF19 中 (於 Libtorch_yolov5_DLLtest.exe) 發生未處理的例外狀況: Microsoft C++ 發生例外狀況: c10::NotImplementedError (記憶體位置 0x000000C0766F9530)。

此報錯發生在OperatorEntry.h底下的
const KernelFunction& lookup(DispatchKey k) const {
const auto& kernel = dispatchTable_[static_cast<uint8_t>(k)];
// A valid kernel always has a boxed kernel and may have an
// unboxed kernel. However, we typically do unboxed calls in at::
// APIs, where the kernel 1) will very likely be valid and 2)
// should have an unboxed kernel. Checking the unboxed kernel
// first will allow us to avoid touching the boxed kernel at all
// in the common case.
if (C10_UNLIKELY(!kernel.isValidUnboxed())) {
if (!kernel.isValid()) {
reportError(k);
}
}
return kernel;
}

想請問這是甚麼問題，謝謝

为啥这么多相同函数名的函数

想问下大佬.h文件里那么多相同函数名的函数有用到吗？有时只是多一个两参数。 我小白一个。

Crop and save bounding box libtorch c++

Hi, I want to add crop and save bounding box libtorch c ++ to the following code, but unfortunately I do not know exactly, I can crop one person, but if there are several objects in an image, I can not crop or save them with different boundix. Thank you very much for helping us or writing the code.?

https://github.com/Nebula4869/YOLOv5-LibTorch

大佬, 什么时候支持segment呀

感谢大佬的作品!
请问有计划增加对segment和pose的检测么,还有分类预测,如果都有,那就厉害了!

未来会出其他项目的c++推理吗，比如stylegan系列

something wrong

大佬，这里报错咋回事

Ubuntu20.04无法应用生成的libyolov5.so库

编译环境
Ubuntu 20.04
gcc-5
cmake 3.21.3
cuda:10.2
cudnn:8.2.1
pytorch：libtorch-shared-with-deps-1.9.1+cu102
opencv：4.5.3

-- Found CUDA: /usr/local/cuda (found suitable exact version "10.2")
-- Found CUDA: /usr/local/cuda (found version "10.2")
-- Caffe2: CUDA detected: 10.2
-- Caffe2: CUDA nvcc is: /usr/local/cuda/bin/nvcc
-- Caffe2: CUDA toolkit directory: /usr/local/cuda
-- Caffe2: Header version is: 10.2
-- Found cuDNN: v8.2.1 (include: /usr/local/cuda/include, library: /usr/local/cuda/lib64/libcudnn.so)
-- /usr/local/cuda/lib64/libnvrtc.so shorthash is 08c4863f
-- Autodetected CUDA architecture(s): 5.0
-- Added CUDA NVCC flags for: -gencode;arch=compute_50,code=sm_50
-- Found CUDA: /usr/local/cuda (found suitable exact version "10.2")
-- Found CUDA: /usr/local/cuda (found version "10.2")
-- Caffe2: CUDA detected: 10.2
-- Caffe2: CUDA nvcc is: /usr/local/cuda/bin/nvcc
-- Caffe2: CUDA toolkit directory: /usr/local/cuda
-- Caffe2: Header version is: 10.2
-- Found cuDNN: v8.2.1 (include: /usr/local/cuda/include, library: /usr/local/cuda/lib64/libcudnn.so)
-- /usr/local/cuda/lib64/libnvrtc.so shorthash is 08c4863f
-- Autodetected CUDA architecture(s): 5.0
-- Added CUDA NVCC flags for: -gencode;arch=compute_50,code=sm_50
-- Configuring done
-- Generating done
-- Build files have been written to: /home/du/SoftWave/Project/YoloV5-LibTorch/cmake-build-debug

生成test的过程中，libyolov5.so正常生成但连接到test时导致opencv库引用失败。

/home/du/SoftWave/clion-2021.2.2/bin/cmake/linux/bin/cmake --build /home/du/SoftWave/Project/YoloV5-LibTorch/cmake-build-debug --target test -- -j 3
Scanning dependencies of target YoloV5
[ 25%] Building CXX object src/CMakeFiles/YoloV5.dir/YoloV5.cpp.o
/home/du/SoftWave/Project/YoloV5-LibTorch/src/YoloV5.cpp: In member function ‘cv::Mat YoloV5::drawRectangle(cv::Mat, at::Tensor, std::map<int, cv::Scalar_ >, std::map<int, std::basic_string >, int)’:
/home/du/SoftWave/Project/YoloV5-LibTorch/src/YoloV5.cpp:321:22: warning: passing NULL to non-pointer argument 1 of ‘cv::Scalar_<Tp>::Scalar(_Tp) [with _Tp = double]’ [-Wconversion-null]
cv::Scalar color = NULL;
^
[ 50%] Linking CXX shared library libYoloV5.so
[ 50%] Built target YoloV5
Scanning dependencies of target test
[ 75%] Linking CXX executable test
/usr/bin/ld: CMakeFiles/test.dir/test.cpp.o: in function main': /home/du/SoftWave/Project/YoloV5-LibTorch/test/test.cpp:49: undefined reference to cv::imshow(std::string const&, cv::_InputArray const&)'
/usr/bin/ld: ../src/libYoloV5.so: undefined reference to cv::putText(cv::_InputOutputArray const&, std::string const&, cv::Point_<int>, int, double, cv::Scalar_<double>, int, int, bool)' /usr/bin/ld: ../src/libYoloV5.so: undefined reference to cv::imread(std::string const&, int)'
collect2: error: ld returned 1 exit status
make[3]: *** [test/CMakeFiles/test.dir/build.make:238：test/test] 错误 1
make[2]: *** [CMakeFiles/Makefile2:142：test/CMakeFiles/test.dir/all] 错误 2
make[1]: *** [CMakeFiles/Makefile2:149：test/CMakeFiles/test.dir/rule] 错误 2
make: *** [Makefile:182：test] 错误 2

为什么会有这些报错信息

C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\src\Yolo.cpp(11,8): error C2039: "device": 不是 "Yolo" 的
成员 [C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\build\src\Yolo.vcxproj]
C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\include\Yolo.h(53,7):
参见“Yolo”的声明

C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\src\Yolo.cpp(15,8): error C2039: "confThres": 不是 "Yolo
" 的成员 [C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\build\src\Yolo.vcxproj]
C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\include\Yolo.h(53,7):
参见“Yolo”的声明

C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\src\Yolo.cpp(16,8): error C2039: "isHalf": 不是 "Yolo" 的
成员 [C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\build\src\Yolo.vcxproj]
C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\include\Yolo.h(53,7):
参见“Yolo”的声明

C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\src\Yolo.cpp(23,34): error C2039: "non_max_suppression
": 不是 "Yolo" 的成员 [C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\build\src\Yolo.vcxproj]
C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\include\Yolo.h(53,7):
参见“Yolo”的声明

C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\src\Yolo.cpp(40,23): error C3861: “xywh2xyxy”: 找不到标识符
[C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\build\src\Yolo.vcxproj]
C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\src\Yolo.cpp(55,22): error C3861: “nms”: 找不到标识符 [C:\de
vResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\build\src\Yolo.vcxproj]
C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\src\Yolo.cpp(61,18): error C2039: "getRandScalar": 不是

"Yolo" 的成员 [C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\build\src\Yolo.vcxproj]
C:\devResources\gitproject\Yolo-LibTorch-main\Yolo-LibTorch-main\include\Yolo.h(53,7):
参见“Yolo”的声明

关于非极大值抑制函数的一个问题

请问
std::vectortorch::Tensor YoloV5::non_max_suppression(torch::Tensor prediction, float confThres, float iouThres)
函数里面，torch::Tensor c = x.slice(1, 5, 6) * maxWh;执行的作用是什么？

ubuntu系统cmake ..会报错，期待出个ubuntu版本教程

用自己训练出来的yolo8的推理，程序崩溃；用您给出来的yolov8n.cuda.torchscript可以正常出结果

为什么在我用自己训练出的person.torchscript运行到下面
if (strcmp(this->version.c_str(), V8) == 0)
{
// 处理V8版本的预测结果
torch::Tensor pT = pred.toTensor();
torch::Tensor score = std::get<0>(pT.slice(1, 4, -1).max(1, true));
data = torch::cat({pT.slice(1, 0, 4), score, pT.slice(1, 4, -1)}, 1).permute({0, 2, 1});
} 会崩溃呢？

大佬、你的libtorch是哪个版本呢？

我用的是1.12.1版本，在torch::jit::load时会崩溃，不知道是不是版本的原因？

你好，为啥报错c10:error

在
inline c10::intrusive_ptr<ivalue::Tuple> IValue::toTuple() && { AT_ASSERT(isTuple(), "Expected Tuple but got ", tagKind()); return moveToIntrusivePtr<ivalue::Tuple>();
报错：0x00007FFA5F0FCD29 处(位于 kuka_socket_new_1.exe 中)有未经处理的异常: Microsoft C++ 异常: c10::Error，位于内存位置 0x000000A1EE6FF160 处。

请问batch size的问题，以及如何才能支持GPU半精度模型？[问题请教]

请忽略开始的问题，发现自己犯了个错，YOLO export的时候用错了一些设置。
我希望检测一个视频，不知道怎样才能最大利用GPU。Batch size是可以大一点吗？
如果Yolo导出一个batch size 大于1的模型，不知道怎么在这里使用。
没有看到关于batch size的设定，是否在test.cpp里面预读几个image再塞到模型里面？
另外，如何才能支持GPU半精度模型？
谢谢指教！

yolov8和yolov5区别

作者您好，我在导出torchscript格式的文件时，发现Yolov8和yolov5在第三维（代表各种属性信息）少了一行。 Yolov8 导出尺寸（1、10、8400）； 而 yolov5 导出尺寸 (1 ,11,XXX)。 我的数据包含五个类别。 缺失的一行数据代表什么呢？我把您代码中关于Yolov8的处理转移到我的代码中，但是发现不能检测最后一类缺陷。

Output function for yolox?

None of output function work for yolox. Do u have suitable code for the yolox output that can be pass to the nms function of your work?

关于图片输入模型时的预处理

你好就是我最近在尝试部署我的yoloV5模型，我阅读你的源码后认为你的letterbox操作跟python源码的好像不太一样。但是我按照官方的思路写letterbox的话输入模型会报错，有时还会产生python端不会预测的框，按照你的思路写的话就不会报那个错误，产生的框也是python端产生的框。

cmake编译不通过

cuda 11.3 nvcc8.3 libtorch 1.10 opencv4.5.5

源码是CPU推理的吗，支持GPU推理吗

您好这个项目很不错，不知道是不是支持GPU推理

ubuntu 18.04 make 报错

CMake Error at test/CMakeLists.txt:6 (add_executable):
The target name "test" is reserved or not valid for certain CMake features,
such as generator expressions, and may result in undefined behavior.

CMake Error at test/CMakeLists.txt:8 (target_link_libraries):
Cannot specify link libraries for target "test" which is not built by this
project.

无法正确读取pt文件

• 麻烦，请教下
• py那边训练好的pt，c++程序读取后，运行，卡在

torch::Tensor pred = model.forward({ data }).toTuple()->elements()[0].toTensor();

• 请问可能是什么问题
• 已经用export.py对pt文件进行了转换

std::vector<torch::Tensor> YoloV5::prediction(torch::Tensor data)
{
	if (!data.is_cuda() && this->isCuda)
	{
		data = data.cuda();
	}
	if (data.is_cuda() && !this->isCuda)
	{
		data = data.cpu();
	}
	torch::Tensor pred = model.forward({ data }).toTuple()->elements()[0].toTensor();
	return non_max_suppression(pred, confThres, iouThres);
}

关于yolov5版本和精度问题

您好，请问你这个项目针对yolov5的哪个版本呢？还是通用的？
另外一个问题是转换后的精度相比较yolov5有下降么？

niu bi ! this project is worth reading

this issue just gratitude to author。i ❤️ u

你好，这个代码已经可以部署yolov8了吗，对libtorch和opencv版本有要求吗？

请问矩形框标注的label会出现重复几个置信度的情况如何解决？

编译的时候找不到OpenCVConfig.cmake opencv-config.cmake？

已经在Cmake list中添加了相应路径：
cmake_minimum_required(VERSION 3.10)

Windows 系统下指定 pytorch 路径和 opencv 路径

set(F:/programs/opencv/build)
set(F:/programs/libtorch/hare/cmake/Torch)

project(YoloV5LibTorch)

include_directories(./include)

add_subdirectory(./src)
add_subdirectory(./test)

install(TARGETS test YoloV5
RUNTIME DESTINATION ${CMAKE_SOURCE_DIR}/bin
ARCHIVE DESTINATION ${CMAKE_SOURCE_DIR}/lib
)

进行第三步的时候，提示下面错误。
CMake Error at src/CMakeLists.txt:1 (find_package):
By not providing "FindOpenCV.cmake" in CMAKE_MODULE_PATH this project has
asked CMake to find a package configuration file provided by "OpenCV", but
CMake did not find one.

Could not find a package configuration file provided by "OpenCV" with any
of the following names:

OpenCVConfig.cmake
opencv-config.cmake

Add the installation prefix of "OpenCV" to CMAKE_PREFIX_PATH or set
"OpenCV_DIR" to a directory containing one of the above files. If "OpenCV"
provides a separate development package or SDK, be sure it has been
installed.
-- Configuring incomplete, errors occurred!
See also "F:/programs/YoloV5-LibTorch-main/build/CMakeFiles/CMakeOutput.log".

多gpu推理

您好，代码可以实现多gpu加载yolo模型，实现多gpu推理吗