This guide will walk you through the process of installing OpenCV-Python from the source. Before beginning, make sure to remove any existing OpenCV-Python directories and load the necessary modules:
Step 0: Have cuda and cudnn installed properly on your environment: https://medium.com/geekculture/install-cuda-and-cudnn-on-windows-linux-52d1501a8805#68ce
Next, clone the OpenCV-Python repository and navigate to the newly created directory:
Step 1:
git clone --recursive https://github.com/opencv/opencv-python.git
Step 2:
cd opencv-python
Once inside the OpenCV-Python directory, check out the stable version of the code that you wish to install. As of today, the stable version is 4.7.0.72, which corresponds to commit f9de34ec8bc8bc5d61e5e35c0721676993da7d71. Note that if you attempt to build the latest version, you may encounter a PEP 440 error.
Step 3:
git checkout $STABLE_VERSION
git checkout f9de34ec8bc8bc5d61e5e35c0721676993da7d71
Now it's time to set up your build environment by exporting the necessary CMake arguments, where "path/to/*" is your path to the * directory that is required:
Step 4:
export CMAKE_ARGS="-DOPENCV_EXTRA_MODULES_PATH= $path/to/opencv-python/opencv_contrib/modules$ -DWITH_LAPACK=OFF -DENABLE_FAST_MATH=1 -DCUDA_FAST_MATH=1 -DBUILD_SHARED_LIBS=OFF -DBUILD_TESTS=OFF -DBUILD_PERF_TESTS=OFF -DBUILD_EXAMPLES=OFF -DWITH_OPENEXR=OFF -DWITH_CUDA=ON -DWITH_CUBLAS=ON -DWITH_CUDNN=ON -DOPENCV_DNN_CUDA=ON -DCUDA_GENERATION=Volta -DENABLE_PRECOMPILED_HEADERS=OFF $path/to/opencv-python$"
Before we proceed, I would like to mention that there are many more flags that you can enable or disable, but this was the easiest way for me to get GPU support for the tasks at hand. Let's go through each of these arguments one by one to understand what they do:
-DOPENCV_EXTRA_MODULES_PATH: This option specifies a semicolon-separated list of directories containing extra modules that will be added to the build.-DWITH_LAPACK=OFF: Keeping this flag on may cause errors in conjunction with the imlk library, which is Intel's version of GCC. This flag disables the use of LAPACK (Linear Algebra PACKage) library.-DENABLE_FAST_MATH=1: This flag enables the use of fast math operations.-DCUDA_FAST_MATH=1: This flag enables the use of fast math operations with CUDA.-DBUILD_SHARED_LIBS=OFF: This flag disables building shared libraries and builds static libraries instead.-DBUILD_TESTS=OFF: This flag disables building tests.-DBUILD_PERF_TESTS=OFF: This flag disables building performance tests.-DBUILD_EXAMPLES=OFF: This flag disables building examples.-DWITH_OPENEXR=OFF: This flag disables the use of OpenEXR library.-DWITH_CUDA=ON: This flag enables CUDA support.-DWITH_CUBLAS=ON: This flag enables CUDA Basic Linear Algebra Subprograms (cuBLAS) support.-DWITH_CUDNN=ON: This flag enables CUDA Deep Neural Network (cuDNN) support.-DOPENCV_DNN_CUDA=ON: This flag enables CUDA support for OpenCV’s Deep Neural Network module.-DCUDA_GENERATION=Volta: This flag specifies the target CUDA architecture generation (in this case, Volta).-DENABLE_PRECOMPILED_HEADERS=OFF: This flag disables the use of precompiled headers.
After setting up your build environment, enable contrib by exporting the ENABLE_CONTRIB variable:
Step 5:
export ENABLE_CONTRIB=1
Now you're ready to build OpenCV-Python by running pip wheel . with the --verbose flag:
Step 6:
pip wheel . --verbose
Once the build process is complete, you should have a new wheel file in your current directory. Simply install this wheel using pip install, and you're done!
Step 7:
pip install $*.whl$
pip install opencv_contrib_python-4.7.0.72-cp37-cp37m-linux_x86_64.whl
mine was opencv_contrib_python-4.7.0.72-cp37-cp37m-linux_x86_64.whl
Thank you for reading my first article on Medium! Enjoy your fast operations!