Control Message Documentation

(Latest Version)

The control message is a JSON object used in the Morpheus pipeline workflow. It is wrapped in a ControlMessage object and passed between the Morpheus stages.

The primary component inputs are used to group control messages together, treating them as if they were inputs in a pipeline. In the form of an array, the inputs component is made up of individual control message objects, each representing a distinct input to the pipeline. Each control message object has the following structure:

Control Message

Copy
Copied!
            

{ "tasks": [ // Array of task objects ], "metadata": { // Metadata object } }

Tasks

The tasks component of each control message object is an array of task objects, each of which represents a separate task to be executed on the input data. Each task object has the following structure:

Copy
Copied!
            

{ "type": "string", "properties": { // Properties object } }

  • type : The type field of the task object is a string indicating the type of task to be executed. Currently, the following task types are supported:

    • load : Load input data from a specified file or files

    • training : Train a machine learning model on the input data

    • inference : Perform inference using a trained machine learning model on the input data

  • properties : The properties field of the task object is an object containing task-specific properties. The specific properties required for each task type are described below.

    • The properties object for a load task has the following structure:

      Copy
      Copied!
                  

      { "loader_id": "string", "files": [ "string" ] }

      • loader_id : The ID of the loader to be used for loading the input data. Currently, only the fsspec and file_to_df loaders are supported. The user has the option to register custom loaders in the dataloader registry and utilize them in the pipeline.

      • files : An array of file paths or glob patterns specifying the input data to be loaded.

    • Incorporate key and value updates to properties objects as required for training and inference tasks. There is no specified format.

Metadata

The metadata component of each input object is an object containing metadata information. Properties defined in this metadata component can be accessed anywhere across the stages that consume ControlMessage objects.

  • data_type : which is a string indicates how to process the data. The supported data types are:

    • payload : Arbitrary input data

    • Streaming : Streaming data

This example demonstrates how to add various parameters to control message JSON. Below message contains an array of three task objects: a load task, a training task, and an inference task. The load task loads input data from two files specified in the files array to a dataframe using the fsspec loader. The training task trains a neural network model with three layers and ReLU activation. The inference task performs inference using the trained model with ID model_001. The metadata component of the input object indicates that the input data type is payload.

Copy
Copied!
            

{ "inputs": [ { "tasks": [ { "type": "load", "properties": { "loader_id": "fsspec", "files": [ "/path/to/file1", "/path/to/file2" ] } }, { "type": "training", "properties": { "model_type": "neural_network", "model_params": { "num_layers": 3, "activation": "relu" } } }, { "type": "inference", "properties": { "model_id": "model_001" } } ], "metadata": { "data_type": "payload" } } ] }

© Copyright 2023, NVIDIA. Last updated on Apr 11, 2023.