Planning

[jan-meyer-1986]

State of Development on feat/getml_mlflow:

• Logging of Input parameters in the Parameters tab. Includes all parameters of objects in the pipeline: predictors, feature learners, preprocessors etc
• Logging of model performance parameters (currently during fitting, but should be moved to evaluate/scoring)
• Logging of live engine performance metrics (memory usage, CPU usage)
• Logging of system metrics

On log-dataset-info:

• logging of data set Meta data: we can govern the displayed names of dataframes and their contexts. But runs on PandasDataSet under the hood
• Scoring using mlflow.evaluate(). But only to limited extent (produces only prediction based metrics and plots, for probability based plots, model provision is necessary)

Under development locally:

• getMLDataSet: avoids reliance on PandasDataSet, enables the controlling of the "profile" (the dictionary in the dataset view). so far unable to add roles to the schema

Under development:

• Enahnced communication between engine and API: detailed logging of progress of Feature Learner and Predictor progress. Soeren developed communication protocol on engine side already.

Future developments:

• Integration of improved engine communication on API side
• Input example and signature (challenge: getML uses containers)
• enable mlflow.evaluate() to use the model for on the spot predictions
• sensible naming and grouping of tags
• one-pager or more complex interactive interface to report logged information

Stale/obsolete less relevant branches and PRs:

#1 (project-folder-identifier) : identifies correct getML installation folder. Might be less relevant due to correct getml1.5 release
parameter-revision: merged
log-metrics: merged
save-load-getml: obsolete