search

m2lines / gz21_ocean_momentum

Stochastic-Deep Learning Parameterization of Ocean Momentum Forcing
MIT License
5 stars 1 forks source link

train: add `--forcing-data-path` CLI option to load forcing data directly instead of through MLflow #84

Closed raehik closed 1 year ago

raehik commented 1 year ago

Fixes #78 .

Should be backwards compatible with --run-id invocations. Note that we remove --exp-id due to it not being used (other than to for logging to screen).

raehik commented 1 year ago

Note that this doesn't touch the rest of the MLflow integration: we still (optionally) execute the steps with MLflow, and save files using mlflow.log_artifact() in the data step.

raehik commented 1 year ago

Cleaned up my to-dos, removed --exp_id. It doesn't get used due to how MLflow searches through its own runs. (Maybe it was required in an earlier version.)

I'm trying to fix up MLflow so that both options are available. Having issues with None in MLproject being parsed as a string...

raehik commented 1 year ago

The MLproject file is restrictive (can't default to None type), so we need a silly workaround to make it work. @MarionBWeinzierl how is this?