202306161114-mlflow_with_hydra

MLflow with Hydra

Hydra param sweeps:

mlflow run . -P hydra_options="-m parameters.a=3,4 parameters.b=range(2,4,1)"

Easy parallelization (requires the pip-installable hydra-joblib-launcher):

mlflow run . -P hydra_options="hydra/launcher=joblib parameters.a=3,4 parameters.b=range(2,4,1) -m"

NOTE: use experiment configs instead? This setup does not yet use the hierarchical nature of Hydra.

MLflow imposes a limit on the length of each option string of 250 characters. Sometimes if you are changing several parameters you might hit that limit on the command line. You can solve this problem by putting your ranges directly in the configuration file (remember to make a copy of the original so you can always revert back) and then run without any parameter override.