Building the Python API for nGraph

Building nGraph Python Wheels

If you want to try a newer version of nGraph's Python API than is available from PyPI, you can build the latest version from source code. This process is very similar to what is outlined in our ngraph_build instructions with two important differences:

You must specify: -DNGRAPH_PYTHON_BUILD_ENABLE=ON and -DNGRAPH_ONNX_IMPORT_ENABLE=ON when running cmake.
Instead of running make, use the command make python_wheel.

$ cmake ../ -DNGRAPH_PYTHON_BUILD_ENABLE=ON -DNGRAPH_ONNX_IMPORT_ENABLE=ON

$ make python_wheel

After this procedure completes, the ngraph/build/python/dist directory should contain the Python packages of the version you cloned. For example, if you checked out and built 0.21 for Python 3.7, you might see something like:

$ ls python/dist/
ngraph-core-0.21.0rc0.tar.gz
ngraph_core-0.21.0rc0-cp37-cp37m-linux_x86_64.whl

Building nGraph Python Wheels on Windows

The build process on Windows consists of 3 steps:

Configure the build with the following cmake invocation:

cmake ..
      -G"Visual Studio 15 2017 Win64"
      -DCMAKE_BUILD_TYPE=Release
      -DCMAKE_INSTALL_PREFIX="C:\temporary_install_dir"
      -DNGRAPH_PYTHON_BUILD_ENABLE=TRUE
      -DNGRAPH_ONNX_IMPORT_ENABLE=TRUE
      -DCMAKE_CXX_COMPILER="C:\Program Files (x86)\Microsoft Visual Studio\2017\Professional\VC\Tools\MSVC\14.16.27023\bin\Hostx64\x64"

There are a couple of things to notice here. One is that the full path to the x64 version of MSVC compiler has to be specified. This is because DNNL requires a 64-bit version and cmake may fail to detect it correctly. The other equally important thing to note is that the temporary directory where the build is to be installed can be specified. If the installation directory is not specified, the default location is C:\Program Files\OpenVINO. This examples uses C:\temporary_install_dir however, a subdirectory of openvino\build works as well. The final Python wheel will contain the contents of this temporary directory so it's very important to set it.

To specify an exact Python version, use the following options:

-DPYTHON_EXECUTABLE="C:\Program Files\Python37\python.exe"
-DPYTHON_LIBRARY="C:\Program Files\Python37\libs\python37.lib"
-DPYTHON_INCLUDE_DIR="C:\Program Files\Python37\include"

Build the install target:

cmake --build . --target install --config Release -j 8

In this step nGraph will be built and installed to the temporary directory specified above. You can adjust the number of threads used in the building process to your machine's capabilities.

Build the Python wheel itself:

cmake --build . --target python_wheel --config Release -j 8

The final wheel should be located in ngraph\python\dist directory.

Configure the environment for the Inference Engine Python API:

call <CMAKE_INSTALL_PREFIX>\bin\setupvars.bat

NOTE: Skip this step if you want to use the nGraph Wheel by itself. This step is required for most usage scenarios, like running unit tests.

Using a virtualenv (optional)

You may wish to use a virutualenv for your installation.

$ virtualenv -p $(which python3) venv
$ source venv/bin/activate
(venv) $

Installing the wheel

You may wish to use a virutualenv for your installation.

(venv) $ pip install ngraph/build/python/dist/ngraph_core-0.21.0rc0-cp37-cp37m-linux_x86_64.whl

Running tests

Unit tests require additional packages be installed:

(venv) $ cd ngraph/python
(venv) $ pip install -r test_requirements.txt

Then run tests:

(venv) $ pytest tests

3.7 KiB Raw Blame History