C++ implementation and python bindings of the numeric KL projection.
It contains a tuned implementation for the non-batched projection (DiagCovOnlyProjection.cpp/h) and the batched-projection (BatchedDiagCovOnlyProjection.cpp/h). The latter is parallelizing the former with OpenMP.
The interface to python is provided by pybind11 (PyProject.cpp), the conversion for numpy arrays to Armadillo vec/mat/cube is provided in PyArmaConverter.h.
Note that the C++ part uses column-major layout while the python part uses row-major (nothing you need to care about if you just want to use it).
Tested with python 3.6.8
The following libraries are required: gcc, openmp, gfortran, openblas, lapack, and cmake. Install them using your package manager
(On Ubuntu: sudo apt-get install gcc gfortran libopenblas-dev liblapack-dev cmake ) .
Dowload NLOPT and follow the installation instructions. You do not need to install using sudo but can put the library anywhere.
Change the include_directories and link_directories statements in cpp_projection/CMakeLists.txt such that they point to the NLOPT headers and libraries.
Download Armadillo, unpack, and run ./configure. You do not need to build Armadillo.
Change the include_directories and statement in cpp_projection/CMakeLists.txt such that it points to the Armadillo headers.
Download pybind11, unpack, rename to pybind11, and place the pybind11 folder under cpp_projection/.
You can put it at a different location or install pybind11 using pip, conda, or any other method, but then the cpp_projection/CMakeLists.txt needs to be adpated such that pybind11 is found.
With your virtualenv being active, go to cpp_projection and run
python3 setup.py install --user