MLflow, the open-source platform that helps developers streamline the machine learning cycle, just saw its latest release go public. MLflow is an API set that can be used with any ML application or library, such as TensorFlow, PyTorch, and XGBoost, in any environment where ML code is run. The set of current components includes:
- The MLflow Tracking API logs ML experiment parameters, code, and results and enables comparisons via an interactive UI.
- The code packaging format, MLflow Projects, facilitates reproducible runs using Conda and Docker, making ML code easily shared.
- MLflow Models is a model packaging format. It includes tools that allow users to deploy a model to batch and real-time scoring on several platforms, including Docker, Apache Spark, Azure ML, and AWS SageMaker.
- A centralized model store, set of APIs, and UI, collectively known as MLflow Model Registry, supports collaborative management of the MLflow Models lifecycle.
The latest MLflow release (2.9.0) introduces several major features and improvements, including the MLflow AI Gateway deprecation in favor of the MLflow deployments API, with a migration guide now available, and the overhaul of The MLflow tracking docs. The complete list of features, security and bug fixes, and documentation updates is available on the 2.9.0 release GitHub page.