Logging Workflows

Currently 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

• MLflow is natively supported by MLJ. You will still 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.