nebula4869/yolov5-libtorch

Real time object detection with deployment of YOLOv5 through LibTorch C++ API

YOLOv5-LibTorch

Real time object detection with deployment of YOLOv5 through LibTorch C++ API

Environment

• Ubuntu 18.04
• OpenCV 3.2.0
• LibTorch 1.6.0
• CMake 3.10.2

Getting Started

1. Install OpenCV.

   sudo apt-get install libopencv-dev

2. Install LibTorch.

   wget https://download.pytorch.org/libtorch/nightly/cpu/libtorch-shared-with-deps-latest.zip
   unzip libtorch-shared-with-deps-latest.zip

3. Edit "CMakeLists.txt" to configure OpenCV and LibTorch correctly.

4. Compile and run.

   cd build
   cmake ..
   make
   ./../bin/YOLOv5LibTorch

Note: COCO-pretrained YOLOv5s model has been provided. For more pretrained models, see yolov5.

yolov5-libtorch's Issues

How to extract the raw results without using the func non_max_suppression?

First of all , thanks for you share.
I'd like to know how to extract the raw results after the forward step?
because i need to debug for the middle result before using the nmx
Thanks!

How to change bounded box & text color?

Im new to libtorch. How to change bounded box & text color?

Can we use OpenCL for AMD GPUs?

自己的模型怎样生成.torchscript.pt

jittrace总是报错 请问怎么回事

make error ...how to solve it ?

[ 50%] Building CXX object CMakeFiles/YOLOv5LibTorch.dir/src/YOLOv5LibTorch.cpp.o
[100%] Linking CXX executable ../bin/YOLOv5LibTorch
CMakeFiles/YOLOv5LibTorch.dir/src/YOLOv5LibTorch.cpp.o：在函数‘main’中：
YOLOv5LibTorch.cpp:(.text.startup+0x14b4)：对‘cv::format(char const*, ...)’未定义的引用
YOLOv5LibTorch.cpp:(.text.startup+0x1587)：对‘cv::putText(cv::InputOutputArray const&, std::string const&, cv::Point, int, double, cv::Scalar_, int, int, bool)’未定义的引用
YOLOv5LibTorch.cpp:(.text.startup+0x19c3)：对‘cv::putText(cv::InputOutputArray const&, std::string const&, cv::Point, int, double, cv::Scalar_, int, int, bool)’未定义的引用
YOLOv5LibTorch.cpp:(.text.startup+0x1a3f)：对‘cv::imshow(std::string const&, cv::_InputArray const&)’未定义的引用
collect2: error: ld returned 1 exit status
CMakeFiles/YOLOv5LibTorch.dir/build.make:132: recipe for target '../bin/YOLOv5LibTorch' failed
make[2]: *** [../bin/YOLOv5LibTorch] Error 1
CMakeFiles/Makefile2:94: recipe for target 'CMakeFiles/YOLOv5LibTorch.dir/all' failed
make[1]: *** [CMakeFiles/YOLOv5LibTorch.dir/all] Error 2
Makefile:102: recipe for target 'all' failed
make: *** [all] Error 2

how about yolov5 v6?

Only Run on CPU.......

Only Run on CPU.......
How to run on GPU?

Why does the NMS algorithm take several minutes to slow down to a graph?

请问一下处理速度和Python相比如何啊？

I need help!

//I use the way of torch.jit.trace to convert original pt file to torchscript pt file , but it generates many bugs.....
I find it !! There is a export.py file converting original pt file to torchscript pt file in the YOLOv5 project .

which version of yolov5 ?

how to use fp16?

Crop and save bounding box. I found yolov4 can do and any for yolov5 in c++

The torchscript weight from YoLoV5 is not workable

Thanks very much for creating this repo.

Now I am facing one problem that the converted torchscript weight from YoLoV5 is not working with the code you provided.

The conversion is followed on this:
ultralytics/yolov5#251

The image size is set to [640, 640] and I also modified your code to be 640.

When loading your existing torchscript weight, it works perfectly.

Do you know which could be the reasons?

Thanks.

Opencv Error when running make

When i run make, it returns a linking error.

/usr/bin/ld: CMakeFiles/YOLOv5LibTorch.dir/src/YOLOv5LibTorch.cpp.o: in function main': YOLOv5LibTorch.cpp:(.text.startup+0xed7): undefined reference to cv::putText(cv::InputOutputArray const&, std::string const&, cv::Point, int, double, cv::Scalar_, int, int, bool)'
/usr/bin/ld: YOLOv5LibTorch.cpp:(.text.startup+0xf5a): undefined reference to cv::imshow(std::string const&, cv::_InputArray const&)' /usr/bin/ld: YOLOv5LibTorch.cpp:(.text.startup+0x1621): undefined reference to cv::format(char const*, ...)'
/usr/bin/ld: YOLOv5LibTorch.cpp:(.text.startup+0x16f6): undefined reference to `cv::putText(cv::InputOutputArray const&, std::string const&, cv::Point, int, double, cv::Scalar_, int, int, bool)'
collect2: error: ld returned 1 exit status
make[2]: *** [CMakeFiles/YOLOv5LibTorch.dir/build.make:160: ../bin/YOLOv5LibTorch] Error 1
make[1]: *** [CMakeFiles/Makefile2:95: CMakeFiles/YOLOv5LibTorch.dir/all] Error 2
make: *** [Makefile:103: all] Error 2

it is can not work when I use my own trained model

Hi, i get a problem when i use my model to inffer. I converted the official yolov5s.pt to yolov5s.torchscript.pt by export.py , but it cannot work in your code, and also my own model can not work too, do you know how can i sovle it ? thank you so much!

memory error

Hi, I've tried your code on your provided weights its working fine, but when I tried my custom weight exported using ultralytics repo, even using older tags is not working. can you please help me with it

这个NMS好像是有点问题的

Error in inference on other size except (640,640)

Hi,
I want inference my .torchscript.pt which is trained by this command and switch:

python train.py --img 128 128 --rect ...

but, by change resize function in this repo to (128,128), i got error:

RuntimeError: The size of tensor a (32) must match the size of tensor b (80) at non-singleton dimension 3

but, i don't know where should i change to work for me.
please help me.

Question: Have you tried this code with CUDA?

I am pursuing a similar path, and have been able to loop multiple inference calls to module::forward() when using CPU version of libtorch and PT file. CUDA version not so much - the inference call makes it through the first time, but subsequent call results in an exception. Have you tried this, and what are your results? FYI I am using CUDA 10.1.

I also logged an issue on this with pytorch group:
pytorch/pytorch#47590

Thanks,
Rob