# Model Runner

    identifier: 0x140f9a78
    status: experimental

Runs machine learning models.

Only models with a single input tensor and a single output tensor are supported at the moment.
Input is provided by Sensor Aggregator service on the same device.
Multiple instances of this service may be present, if more than one model format is supported by a device.

## Commands

    unique command set_model @ 0x80 {
        model_size: u32 B
    }
    report {
        model_port: pipe_port
    }

Open pipe for streaming in the model. The size of the model has to be declared upfront.
The model is streamed over regular pipe data packets.
The format supported by this instance of the service is specified in `format` register.
When the pipe is closed, the model is written all into flash, and the device running the service may reset.

    command predict @ 0x81 {
        outputs: pipe
    }
    report {
        inputs: pipe_port
    }

Open channel that can be used to manually invoke the model. When enough data is sent over the `inputs` pipe, the model is invoked,
and results are send over the `outputs` pipe.

## Registers

    rw auto_invoke_every: u16 @ 0x80

When register contains `N > 0`, run the model automatically every time new `N` samples are collected.
Model may be run less often if it takes longer to run than `N * sampling_interval`.
The `outputs` register will stream its value after each run.
This register is not stored in flash.

    ro outputs @ reading {
    repeats:
        output: f32
    }

Results of last model invocation as `float32` array.

    ro input_shape @ 0x180 {
    repeats:
        dimension: u16
    }

The shape of the input tensor.

    ro output_shape @ 0x181 {
    repeats:
        dimension: u16
    }

The shape of the output tensor.

    ro last_run_time: u32 us @ 0x182

The time consumed in last model execution.

    ro allocated_arena_size: u32 B @ 0x183

Number of RAM bytes allocated for model execution.

    ro model_size: u32 B @ 0x184

The size of the model in bytes.

    ro last_error: string @ 0x185

Textual description of last error when running or loading model (if any).

    enum ModelFormat: u32 {
        TFLite = 0x334c4654,
        ML4F = 0x30470f62,
        EdgeImpulseCompiled = 0x30564945,
    }
    const format: ModelFormat @ 0x186

The type of ML models supported by this service.
`TFLite` is flatbuffer `.tflite` file.
`ML4F` is compiled machine code model for Cortex-M4F.
The format is typically present as first or second little endian word of model file.

    const format_version: u32 @ 0x187

A version number for the format.

    const parallel?: bool @ 0x188

If present and true this service can run models independently of other
instances of this service on the device.