Compile YOLOv4 on Windows

 technik  737  4 mins  

The compilation of a YOLOv4 (GPU version) on Windows seems to be much more complex than in other platforms, for example to configure the environment. The article is aimed to introduce a method, that shows how we prepare the dependences and compile the YOLOv4.

The method is tested on Windows 10 2004, with Nvidia RTX 2070.

I. Prepare the Dependences

1) Install Anaconda

Anaconda is an open source that helps us to manage the python environment. With conda/ pip/ vituralenv can we simply manage and deploy the packages.

The software can be downloaded by —–> Link. Install it with default setting.

2) Install Visual Studio 2017

VS 2017 can be downloaded by —–> Link. Here what we need is the desktop development with C++.

vs2017

3) Install CUDA and cuDNN

CUDA stands for “Compute Unified Device Architecture”. The parallel computing platform makes the GPU possible for complex computation. —–> (Link)

cuDNN is a library provided by Nvidia for deep learning. The library speed up the training and inference processes through the usage of GPU and supports many deep learning frameworks. —–> (Link)

Make sure to download the same version cuDNN for CUDA. Here we download the CUDA 10.1 driver.

The installation of CUDA and cuDNN can be declared as following steps:

  • Install CUDA with default settings.
  • Unzip the downloaded cuDNN zip file.
  • Copy all the unzipped files to CUDA folder. If you never change the CUDA folder by installation, the default folder should be C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1.

4) Install CMake

Download CMake from the official website, or just download cmake-3.18.2-win64 by —–> Link.

Choose to add its directory to the system PATH, so that we can call the program from any folder in command line.

cmake

5) Download and compile OpenCV

Download the OpenCV sources

OpenCV can be download by —–> Link. For the extended functions, we need to download the OpenCV_contrib from —–> Link.

Here we specifically download OpenCV 3.4.11. The version of OpenCV_contrib must match the OpenCV.

The downloaded OpenCV.exe is a self-extracting file, we can extract the file in C:\. Also, we can put the unzipped OpenCV_contrib in C:\opencv.

The folders in file tree:

C:\opencv\

build\

opencv_contrib-3.4.11\

sources\

LICENSE.txt

LICENSE_FFMPEG.txt

README.md.txt

Compile with CMake

  • Open the CMake_GUI. Choose the source and build folders, and then click Configure.

cmake_choose_folder

  • Choose Visual Studio 15 2017 and specify the platform x64, then click Finish.

cmake_vs2017

  • Wait and configure some parameters.

    Parameters Value
    BUILD_EXAMPLES OFF
    BUILD_PERF_TESTS OFF
    BUILD_TESTS OFF
    BUILD_TBB OFF
    INSTALL_CREATE_DISTRIB ON
    INSTALL_C_EXAMPLES OFF
    INSTALL_PYTHON_EXAMPLES OFF
    OPENCV_EXTRA_MODULES_PATH C:\opencv\opencv_contrib-3.4.11\modules
    WITH_CUDA ON
    WITH_LAPACK OFF
    WITH_VTK OFF
    WITH_TBB OFF

  • Click Configure and wait. Before further action, we need to check the compute capability of your GPU from the official website. —–> link

    My GPU is GeForce RTX 2070, the compute capability is 7.5. So we need to configure some more parameters.

    Parameters Value
    CUDA_ARCH_BIN 7.5
    CUDA_ARCH_PTX 7.5
    BUILD_opencv_saliency OFF
    BUILD_opencv_world ON

  • Then click Configure again. Wait and then generate.

  • Go to C:\opencv\build, call command prompt by address bar.

    opencv_cmd

  • Compile OpenCV with the command:

    1

    cmake.exe --build . --config Release --target INSTALL

  • Modify the system environment variables.

    Action Variable Value
    Create OPENCV_DIR C:\opencv\build
    Create OPENCV_VER 3411
    Add PATH %OPENCV_DIR%\bin\Debug
    Add PATH %OPENCV_DIR%\bin\Release
    Add PATH %OPENCV_DIR%\install\x64\vc15\bin

II. Compile Darknet

1) Download source

Download from Github via git or http. —–> Link

2) Edit darknet.vcxproj

Open darknet.vcxproj in folder darknet/build/darknet. Search:~~

1

<Import Project="$(VCTargetsPath)\BuildCustomizations

Check if the following string is CUDA 10.1. Otherwise, change it to CUDA 10.1. You can only find two places.

3) Compile with darknet.sln

  • Open darknet.sln in the same folder, maybe VS 2017 will ask you to update the file to v141, just confirm. Then set the project to Release and choose x64 platform.

  • Set properties of darknet

    1. Level 2. Level String
    VC++ Directories Include Directories C:\opencv\build\include
    VC++ Directories Include Directories C:\opencv\build\include\opencv
    VC++ Directories Include Directories C:\opencv\build\include\opencv2
    VC++ Directories Library Directories C:\opencv\build\x64\vc15\lib
    Linker.Input Additional Dependencies opencv_world3411.lib

  • Build solution and wait for compilation

2) Compile with script build.ps1

  • Check the execution policies, change the policy if necessary

    • run PowerShell as administrator

    • 1

      Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope LocalMachine

  • Run the script build.ps1 to compile darknet

    • run a PowerShell session by typing powershell in the folder’s address bar, in the folder you can see the script build.ps1

    • 1

      .\build.ps1

  • wait till the compilation is done

PS: If you can’t import cv2 in a python script, just copy the .pyd file from C:\opencv\build\python\cv2 to C:\Users\{yourname}\anaconda3\Lib\site-packages.

The .pyd file should match the python version!!!

Enjoy!!!

updatedupdated2021-05-232021-05-23

See Also: