Technical details, installation#

Benchmarking is an exact science as the results may change depending on the machine used to compute the figures. There is not necessarily an exact correlation between the processing time and the algorithm cost. The results may also depend on the options used to compile a library (CPU, GPU, MKL, …). Next sections gives some details on how it was done.


scikit-learn is usually the current latest stable version except if the test involves a pull request which implies scikit-learn is installed from the master branch.


python install works on Linux. On Windows, protobuf must be compiled first and then referenced before building the wheel. This is what it could look like:

@echo off
set ONNX_ML=1
set PATH=c:\Python370_x64;c:\Python370_x64\Scripts;%PATH%
set PATH=%~dp0..\\..\protobuf\build_msvc\Release;%PATH%
set CMAKE_ARGS=..\\third_party\\pybind11\\tools -DPROTOBUF_INCLUDE_DIRS=C:\\xavierdupre\\__home_\\github_fork\\protobuf\\src -DPROTOBUF_LIBRARIES=C:\\xavierdupre\\__home_\\github_fork\\protobuf\\build_msvc\\Release\\libprotobuf.lib;C:\\xavierdupre\\__home_\\github_fork\\protobuf\\build_msvc\\Release\\libprotoc.lib -DONNX_PROTOC_EXECUTABLE=C:\\xavierdupre\\__home_\\github_fork\\protobuf\\build_msvc\\Release\\protoc.exe
cd onnx
python bdist_wheel
cd ..


onnxruntime is not easy to install on Linux even on CPU. The current implementation requires that Python is built with a specific flags --enable-shared:

./configure --enable-optimizations --with-ensurepip=install --enable-shared --prefix=/opt/bin

This is due to a feature which requests to be able to interpret Python inside a package itself and more specifically: Embedding the Python interpreter. Then the environment variable LD_LIBRARY_PATH must be set to the location of the shard libraries, /opt/bin in the previous example. The following issue might appear:

UserWarning: Cannot load onnxruntime.capi.
Error: ' cannot open shared object file: No such file or directory'

To build onnxruntime:

git clone --recursive
export LD_LIBRARY_PATH=/usr/local/Python-3.6.8
python3.6 ./onnxruntime/tools/ci_build/ --build_dir ./onnxruntime/build/debian36 --config Release --enable_pybind --build_wheel --use_mkldnn --use_openmp --build_shared_lib
export LD_LIBRARY_PATH=/usr/local/Python-3.7.2
python3.7 ./onnxruntime/tools/ci_build/ --build_dir ./onnxruntime/build/debian37 --config Release --enable_pybind --build_wheel --use_mkldnn --use_openmp --build_shared_lib

If the wheel then, it is possible to just copy the files into the python distribution:

cp -r ./onnxruntime/build/debian36/Release/onnxruntime /usr/local/lib/python3.6/site-packages/
cp -r ./onnxruntime/build/debian37/Release/onnxruntime /usr/local/lib/python3.7/site-packages/