vs2017 libtorch cuda11配置yolov5

记录，学习中，c++菜鸡

---

一、libtorch

从官网下载的是debug版本的
Download here for C++ (Debug version):
https://download.pytorch.org/libtorch/cu110/libtorch-win-shared-with-deps-debug-1.7.1%2Bcu110.zip
cuda版本对应libtorch的下载版本，安装并配置路径。

libtorch-yolo5的源码在https://github.com/Nebula4869/YOLOv5-LibTorch

二、vs2017 libtorch opencv

1.配置

选择C/C+±常规-SDL检查，修改为”否“：
选择C/C+±语言，符合模式，修改为”否“：若有std报错
选择C/C+±语言，c++语言标准选ISO C++14 标准 (/std:c++14)，不清楚有没有用

vc++目录 包含目录

D:\libtorch\include
D:\opencv\opencv\build\include\opencv2
D:\opencv\opencv\build\include\opencv
D:\opencv\opencv\build\include
D:\libtorch\include\torch\csrc\api\include

vc++目录 库目录

D:\opencv\opencv\build\x64\vc14\lib
D:\libtorch\lib

链接器 输入 附加依赖项

D:\opencv\opencv\build\x64\vc15\lib\opencv_world341d.lib
D:\libtorch\lib\c10.lib
C:\Program Files\NVIDIA Corporation\NvToolsExt\lib\x64\nvToolsExt64_1.lib
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\lib\x64\cudart_static.lib
D:\libtorch\lib\caffe2_nvrtc.lib
D:\libtorch\lib\c10_cuda.lib
D:\libtorch\lib\torch.lib
D:\libtorch\lib\torch_cuda.lib
D:\libtorch\lib\torch_cpu.lib
-INCLUDE:?warp_size@cuda@at@@YAHXZ
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\lib\x64\cufft.lib
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\lib\x64\curand.lib
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\lib\x64\cublas.lib
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\lib\x64\cudnn.lib
kernel32.lib
user32.lib
gdi32.lib
winspool.lib
shell32.lib
ole32.lib
oleaut32.lib
uuid.lib
comdlg32.lib
advapi32.lib

自行测试libtorch和cuda是否可以用

#include <torch/script.h> 
#include <torch/torch.h> 
#include <iostream>
#include <memory>

int main(int argc, const char* argv[]) {
	std::cout << "cuda::is_available():" << torch::cuda::is_available() << std::endl;
	torch::DeviceType device_type = at::kCPU; // 定义设备类型
	if (torch::cuda::is_available())
		device_type = at::kCUDA;
}

2.部分报错解决方法

参考https://blog.csdn.net/zzz_zzz12138/article/details/109138805
写的很全，很多坑都踩过

模型导出默认为cpu，按下面修改为gpu

# line 29
model.model[-1].export = False
Add GPU support: Note that the current export script in yolov5 uses CPU by default, the "export.py" needs to be modified as following to support GPU:

# line 28
img = torch.zeros((opt.batch_size, 3, *opt.img_size)).to(device='cuda')  
# line 31
model = attempt_load(opt.weights, map_location=torch.device('cuda'))

Export a trained yolov5 model:

cd yolov5
export PYTHONPATH="$PWD"  # add path
python models/export.py --weights yolov5s.pt --img 640 --batch 1  # export

在这要注意导出的模型为input为（640，640），对应c++里面的resize进行修改，不然会报错foward（input）那地方。

GPU推理结果为cuda数据类型，nms之前要转成cpu，否则会报错

std::vector<torch::Tensor> non_max_suppression(torch::Tensor preds, float score_thresh = 0.5, float iou_thresh = 0.5)
增加到函数里pred = pred.to(at::kCPU);注意preds的数据类型，转成cpu进行后处理。

函数中overlaps为cpu数据类型，

总结

总体流程可以参考这个
但是有部分小错误要注意。

release测试通过。