This is a tutorial to helping compile NCNN library. The content comes from my attempts to complete the Mask-Detection project, so the operation is based on the Yolo-Fastest model. I hope it will help readers deploy the model, and also for the convenience of myself to quickly rebuild the environment in the future.
According to the author nihui, the name of NCNN comes from (New/Next or Naive or Neon or Nihui) + CNN. This is the introduction on its Github repository:
ncnn is a high-performance neural network inference computing framework optimized for mobile platforms. ncnn is deeply considerate about deployment and uses on mobile phones from the beginning of design. ncnn does not have third party dependencies. it is cross-platform, and runs faster than all known open source frameworks on mobile phone cpu.
In addition to the excellent optimization of the mobile devices, its no third library dependency also means that there are many fewer restrictions on the deployment of models. However, the process of compiling NCNN still requires several third-party libraries.
- g++ & gcc
- camke
- protobuf
- opencv
The following content is based on Ubuntu 18.04.5. In fact, I have tried MacOS, but after encountering various problems caused by clang instead of g++, I chose to change it. . .
g++ & gcc
Usually, Ubuntu has built-in this library, you can use the following code in the Terminal to check the installed version:
g++ -v
Or you can install it via apt.
sudo apt install gcc
sudo apt install g++
cmake
Similar to the previous one, we can still use apt.
sudo apt install cmake
Protobuf
The installation of protobuf can be implemented in accordance with the instructions of its github repository. The key is the following steps.
sudo apt-get install autoconf automake libtool curl make g++ unzip
Then download the package from its release page:
https://github.com/protocolbuffers/protobuf/releases/latest
The one I downloaded is protobuf-all-3.14.0.tar.gz. Then unzip the package and enter the folder in Terminal. Then run the following code.
./configure
make
make check
sudo make install
sudo ldconfig # refresh shared library cache.
Hint: If protobuf is installed successfully, and some errors are reported when compiling NCNN, you can try to restart Ubuntu. If you still can’t compile NCNN normally, you can recompile and install protobuf before restarting.
OpenCV
Open https://opencv.org/releases/ and select an appropriate version. What I installed is OpenCV-3.4.12. Click Sources to download the package.
Then unzip the package and enter the folder in Terminal. Then run the following code to install the dependencies.
sudo apt-get install build-essential libgtk2.0-dev libavcodec-dev libavformat-dev libjpeg.dev libtiff4.dev libswscale-dev libjasper-dev
Create a compilation folder under this folder, then cmake.
mkdir build
cd build
cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=/usr/local ..
Then start to compile it.
sudo make
sudo make install
Add the OpenCV library to the path so that the system can find it.
sudo gedit /etc/ld.so.conf.d/opencv.conf
This command opened a file. Add /usr/local/lib at the end of the file and save it. Executing this command makes the configuration path effective.
sudo ldconfig
The following command will open a file.
sudo gedit /etc/bash.bashrc
Add these at the end of the file and save it:
PKG_CONFIG_PATH=$PKG_CONFIG_PATH:/usr/local/lib/pkgconfig
export PKG_CONFIG_PATH
Executing this command makes the configuration path effective.
source /etc/bash.bashrc
sudo updatedb
The configuration process of OpenCV has been completed, let’s test it with its own tools.
Enter opencv-3.4.12/samples/cpp/example_cmake in Terminal.
Execute the following code:
cmake .
make
./opencv_example
If the camera is automatically turned on, and a hello opencv appears in the upper left corner of the image captured by the camera, it means that the configuration is successful.
NCNN
Now that the dependent libraries have been installed, we will start to compile NCNN. You can find the official tutorial here.
If you want to use GPU, please install Vulkan as a dependency with the following code.
wget https://sdk.lunarg.com/sdk/download/1.2.154.0/linux/vulkansdk-linux-x86_64-1.2.154.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.2.154.0.tar.gz
tar -xf vulkansdk-linux-x86_64-1.2.154.0.tar.gz
export VULKAN_SDK=$(pwd)/1.2.154.0/x86_64
However, cause I’m using a virtual machine running on Macbook pro, I choose to forget Vulkan. If you choose to do so, please remember to set the Vulkan option to be OFF.
Then we need to clone the repository of NCNN and start to compile it.
git clone https://github.com/Tencent/ncnn.git
cd ncnn
mkdir -p build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_VULKAN=ON -DNCNN_SYSTEM_GLSLANG=ON -DNCNN_BUILD_EXAMPLES=ON ..
make
make install
So far, the compilation process of NCNN has been completed. It is recommended to carefully read the official documents and issues in Github when compiling, you can find many solutions to errors.
After this, we can start to use various tools in the ncnn/build/tools folder to help us convert the model.
For example, you can copy the .cfg and .weights files of your darknet model to the darknet folder, and use the code to convert to the NCNN model. The details could be found here.
./darknet2ncnn yolo-fastest.cfg best.weights yolo-fastest.param yolo-fastest.bin 1
After getting .param and .bin, we can deploy the model to the mobile devices or even the browser. And the files in ncnn/build/install would be used when you need to run your own model.
Note: Thanks to the articles “pytorch模型的部署（系列一）–ncnn的编译和使用” and “ubuntu16.04安装opencv3.4.1教程” for their help in the compilation process.
