This repository contains a jupyter notebook that demonstrates how the Infer.NET framework for graph Bayesian inference can be invoked in python using pythonnet on Ubuntu 18.04.

• Install Docker
• Download or clone this repository
• Move into the repository and build the Docker image: type sudo docker build --rm -t infer_pythonnet .
• Run the image by typing sudo docker run -it -p 8888:8888 -v ${PWD}:/root/dev infer_pythonnet
• The container will start a Jupyter notebook server. Go to http://localhost:8888 and login with the token that is printed out in the terminal

Python.net cannot handle overloads of generic methods properly (see issue). Calling Variable.Observed(Array[Double]([...])) returns Variable[Double[]] instead of VariableArray[Double]. To enforce a correct return type, we could do Variable.Observed[Double](Array[Double]([...]))), but python.net throws an exception.

Possible soultions:

• Do the reflections manually, but this doesn't seem to work because Variable is an abstract class in Infer.Net and Observed is a static method. So the following does not work:

lib = clr.AddReference("Microsoft.ML.Probabilistic")
var_type = lib.GetType("Microsoft.ML.Probabilistic.Models.Variable")
method = var_type.GetMethod("Observed")
...

GetType returns a None-type, since Variable is abstract.

• Write C# library that wraps the function in a non-generic function. This seems to work, but in the end it is not needed (yet) because usage of Variable.Observed can be avoided in the simple examples.

• Make sure to create Variable objects and assign to them in the same cell. Infer.Net doesn't allow to assign to a same Variable multiple times.

Try more complicated examples with visualizations of the network and inference results in python.