Plugin system

While GraphPPL is backend agnostic, specific inference backends might require additional functionality or data saved in nodes. To accommodate these needs, GraphPPL exposes a plugin system that allows users to extend the functionality of the package. Plugins allow the core package to remain lightweight, as well as allowing backend-specific functionality. For example, a node does not need to know by which line of code it was originally created. As an example, for debugging purposes, it might be useful to save this information in the node. GraphPPL implements a plugin that, when enabled on a model, saves this information in every node. This allows for useful debugging, while switching this functionality off when not needed saves the memory footprint of the model.

Creating a plugin

A plugin is a structure that contains a set of functions that are called at specific points in the model creation process. The plugin is implemented with the GraphPPL.preprocess_plugin and GraphPPL.postprocess_plugin functions:

GraphPPL.preprocess_plugin — Function

preprocess_plugin(plugin, model, context, label, nodedata, options)

Call a plugin specific logic for a node with label and nodedata upon their creation.

source

GraphPPL.postprocess_plugin — Function

postprocess_plugin(plugin, model)

Calls a plugin specific logic after the model has been created. By default does nothing.

source

Within these functions, the plugin can modify the model, add new nodes, or modify existing nodes. Also, additional data can be passed to nodes in the preprocess_plugin function.

Available plugins

The following plugins are available by default in GraphPPL:

GraphPPL.VariationalConstraintsPlugin: adds constraints to the model that are used in variational inference.
GraphPPL.MetaPlugin: adds arbitrary metadata to nodes in the model.
GraphPPL.NodeCreatedByPlugin: adds information about the line of code that created the node.
GraphPPL.NodeIdPlugin: allows attaching an id to factor nodes for later inspection.

Using a plugin

To use a plugin, call the with_plugins function when constructing a model:

GraphPPL.with_plugins — Function

with_plugins(generator::ModelGenerator, plugins::PluginsCollection)

Attaches the plugins to the generator. For example:

plugins = GraphPPL.PluginsCollection(GraphPPL.NodeCreatedByPlugin())
new_generator = GraphPPL.with_plugins(generator, plugins)

source

The PluginsCollection is a collection of plugins that will be applied to the model. The order of plugins in the collection is important, as the preprocess_plugin and postprocess_plugin functions are called in the order of the plugins in the collection.