MATLAB Interface for MLflow

This package consists of three interfaces which can be used independently:

1. A fluent MATLAB interface for MLflow tracking based on the fluent interface in the Python MLflow package. Apart from the package itself this interface also requires Python and the official mlflow Python package.

2. A MATLAB interface for creating MATLAB flavor MLflow models. This also requires Python and the official mlflow Python package. This interface is often used in combination with the fluent MATLAB interface for MLflow tracking, such that you can for example log a MATLAB flavor model as part of a run logged using that interface.

3. A MATLAB interface for interacting with MLflow tracking servers through their RESTful interface. This interface will work out-the-box when the package has been installed in MATLAB.

Note that option 1 and 3 offer similar functionality, so why choose the one over the other?

• The fluent interface is more convenient to work with, but does require a Python installation. Whereas the RESTful interface is more verbose and requires you to write more code to achieve the same results but it does not require a Python installation and should just immediately work out-of-the-box.

• The MATLAB flavor feature only works in combination with the fluent tracking interface. I.e. MATLAB flavor MLflow models can be logged as part of a run started with the fluent interface but not with runs as started using the RESTful interface. This is of course only relevant if you want to be able to log models as part of runs. Saving models is always possible independent of any tracking interface.

• The fluent interface supports working with MLflow (Python based) plugins. For example, if the azureml-mlflow package has been installed, the fluent interface can directly work with Azure Machine Learning azureml:// style MLflow tracking URIs, including authentication. Whereas with the RESTful interface you would have to manually work with the https:// equivalent and obtain- and configure authentication manually.

Concluding, the fluent interface is easier to work with and provides more features than the RESTful interface and is therefore recommended over the RESTful interfaces if it is indeed possible to setup the Python environment it requires. If it is not possible to- or you prefer not to- setup the Python environment, choose the RESTful interface.

Regardless of the interface, to get started with this package, clone the repository:

$ git clone https://github.com/mathworks-ref-arch/matlab-mlflow.git

Then consult Installation (fluent interface and MATLAB flavor) and/or Installation (RESTful interface) for further installation instructions and see the documentation outline below for further information.

Tracking (fluent interface)

• Installation and Configuration
  • Requirements
  • Python Environment
  • MATLAB Configuration
  • MLflow Tracking servers
• Fluent MATLAB Interface for Tracking
  • Examples
  • Advanced usage of the Python API directly
• MLflow MATLAB Fluent API
  • mlflow package

MATLAB Flavor Interface

• Installation and Configuration
  • Requirements
  • Python Environment
  • MATLAB Configuration
  • MLflow Tracking servers
• Introduction
  • Sub Flavors
  • MATLAB Interface
• Usage
  • MATLAB Functions
  • Example

Subflavors

• none sub flavor
• compiler_sdk_python sub flavor
  • Workflow overview
  • Function in- and output handling
  • MATLAB Runtime Requirements
  • MATLAB Function and Class References
• onnx sub flavor

Tracking (REST interface)

• Installation
• Quick Start
  • Installation
  • Authentication
  • Create a new experiment
  • Connect to an existing experiment
  • Create a new run for the experiment
  • Log a parameter
  • Log a metric
  • Conclusion
• Authentication

References

• MATLAB Interface for MLflow - API Reference
  • Index
  • Help