ADVANCED.md

Advanced Usage

Editing generated FMU source

The command mlfmu build will both generate the C++ source code for the mlfmu and compile it automatically. However, it is possible to split this into two steps where it is possible to edit the source code to change the behavior of the resulting FMU.

mlfmu codegen --interface-file Interface.json --model-file model.onnx --fmu-source-path path/to/generated/source

This will result in a folder containing the source structured as below.

[FmuName]
├── resources
│   └── *.onnx
├── sources
│   ├── fmu.cpp
│   └── model_definitions.h
└── modelDescription.xml

Of these generated files, it is only recommended to modify fmu.cpp. In this file one can e.g. modify the DoStep function of the generated FMU class.

class FmuName : public OnnxFmu
{
public:
    FmuName(cppfmu::FMIString fmuResourceLocation)
        : OnnxFmu(fmuResourceLocation)
    { }

    bool DoStep(cppfmu::FMIReal currentCommunicationPoint, cppfmu::FMIReal dt, cppfmu::FMIBoolean newStep,
        cppfmu::FMIReal& endOfStep) override
    {
        // Implement custom behavior here
        // ...

        // Call the base class implementation
        return OnnxFmu::DoStep(currentCommunicationPoint, dt, newStep, endOfStep);
    }
private:
};

After doing the modification to the source code, one can simply run the compile command to complete the process.

mlfmu compile --fmu-source-path path/to/generated/source

Using class

In addition to the command line interface, one can use the same functionality of the tool through a Python class.

1. Import MlFmuBuilder and create an instance of it:

from mlfmu.api import MlFmuBuilder
from pathlib import Path

builder = MlFmuBuilder(
    ml_model_file = Path("path/to/model.onnx")
    interface_file = Path("path/to/interface.json")
)

2. Call the same commands using the class:

• Run build

builder.build()

• Run codegen and then compile

builder.generate()

# Do something ...

builder.compile()