Java bindings for PyProb. The code is based loosely on the PyProb C++ bindings.

Building

The project uses Gradle as a build system. It has been tested using Gradle 6.3 and Oracle OpenJDK 14.

Additional Requirements:

• Flatbuffers to compile the PyProb PPX interface.

Necessary build commands:

• gradle createFlatBuffers to generate the Java files from the PPX description
• gradle generateBuildConfig to generate static information from git, used in the program.
• gradle build to compile the Java project.

Basic usage

We have ported the Marsaglia Gaussian model to Java as an example usage. The example model is available in src/test/MarsagliaGaussian.java The core idea is that one writes a model in Java by inheriting from the edu.diku.pyprob_java.Model and implementing the call() method which runs the computations in the model. Inside call() one can use sample() and observe() statements to respectively probabilistically sample and condition the model on data.

To use the model from the Java side, one needs to create a MessageHandler to communicate using PPX and ZeroMQ to the Python-based PyProb inference module. This is usually done as follows:

 try (var messageHandler = new MessageHandler()) {
       var model = new MarsagliaGaussian("Marsaglia Gaussian", messageHandler);
       model.run(serverAddress);
    } catch (InterruptedException e) {
        System.out.println(e.toString());
    }

Where serverAddress is some valid ZeroMQ address, e.g. tcp://127.0.0.1:5555 or ipc://@MarsagliaGaussian.

From the Python-based PyProb point of view one needs to create a RemoteModel that connects to the same serverAddress. For the Marsaglia Gaussian example notebook, one can simply replace the GaussianUnknownMean class and model = GaussianUnknownMean() lines with model = RemoteModel(serverAddress) and run the notebook.