Infer.NET user guide : Controlling how inference is performed

Structure of generated inference code

Infer.NET generates a C# class for performing inference on your model. By default, the source of this class is placed at: (Debug/Release)\bin\GeneratedSource\[ModelName].cs. There are circumstances where it is useful to use this generated code directly, for example, to use a precompiled inference algorithm or to manually modify the message-passing schedule or some other aspect of the code. To help with this, we will now describe the structure of the generated class. Generated source code files also contain full documentation for the class, fields and methods so you can also refer to these directly. To make the code readable, you should attach names to variables in the model.

The Generated Class

The class is an ordinary C# class of name [ModelName]_[Algorithm], for example Model0_EP. It typically contains a number of public fields, properties, and methods. A property is created to hold each observed variable in the model. Fields are created for:

• all requested marginals
• all forward and backwards messages needed to compute requested marginals (x_uses_F[i] is a message from a variable x to its ith use, while x_uses_B[i] is a message to a variable x from its ith use. A gate block counts as a single use.)
• invalidation flags to track which parts of the schedule need to be recomputed when observed values change

Methods are created for:

• Setting observed values and initialising messages

• Performing inference by updating messages

• Retrieving posterior marginals and output messages

Each field and method has a documentation comment describing its purpose.

Generated Methods for Performing Inference

The methods in the table below are used when performing inference. The Reset, Execute and Update methods can all be called through the IGeneratedAlgorithm interface.