You manage an Azure Machine Learning workspace. You build a custom model you must log with MLflow. The custom model includes the following: - The model is not natively supported by MLflow. - The model

Question

Options

• Amodel loader
• Bartifacts
• Cmodel wrapper
• Dcustom signatures

• A
  69% (18)
• B
  4% (1)
• C
  15% (4)
• D
  12% (3)

Explanation

When creating a custom MLflow flavor for a model that cannot be serialized with Pickle and is not natively supported by MLflow, you define a loader_module - the model loader (A). The loader module is a Python module that MLflow calls to load the model from its saved artifacts directory. It exposes a _load_pyfunc function that reconstructs the model object at inference time, bypassing the need for Pickle serialization. This is the required component for a fully custom MLflow flavor. Artifacts (B) are the files saved alongside the model (weights, configs) but do not define how the model is loaded. A model wrapper (C) (mlflow.pyfunc.PythonModel subclass) is an alternative pyfunc approach but does not constitute a custom flavor with its own loader module. Custom signatures (D) define input/output schemas and are optional metadata, not a loading mechanism.

No community discussion yet for this question.