Note

Starting with MLJ 0.23.0, MLJFlow.jl methods are not immediately available, as the package has been removed as a direct dependency. Previous behaviour will require using MLJFlow. See also the "Warning" below.

Logging Workflows

In principle, the following workflows can log their outcomes to an external machine learning tracking platform, such as mlflow:

Estimating model performance using evaluate/evaluate!.
Model tuning, using the TunedModel wrapper, as described under Tuning Models.

To enable logging one must create a logger object for the relevant tracking platform, and either:

Provide logger as an explicit keyword argument in the workflow, as in evaluate(...; logger=...) or TunedModel(...; logger=...); or
Set a global default logger with the call default_logger(logger).

MLJ logging examples are given in the MLJFlow.jl documentation.

Supported tracking platforms

Warning

Due to issues with the mlflow REST API, the current model for MLJ-mlflow integration is being reassessed. Use the existing tools at your own risk.

To use mlflow with MLJ you will need to add MLJFlow to your package environment and call using MLJFlow. You additionally need to install mlflow itself, and separately launch an mlflow service; see the mlflow docs on how to do this. The service can immediately be wrapped to create a logger object, as demonstrated in the MLJFlow.jl documentation.