exalate-issue-sync[bot]
commented
1 year ago Tomas Fryda commented: [~accountid:557058:afd6e9a4-1891-4845-98ea-b5d34a2bc42c], am I correct to assume this should work only for blending, i.e., not for CV?
Or for CV, I think I could also check that the user is training with the same data as the checkpointed model was and then just use the stored CV predictions, i.e., the user would not be able to provide more data but (s)he could improve the metalearner, e.g., by adding more trees (trained on same data as before). If the user specifies different frame, I could throw an exception.
What do you think? Or is there some better approach?
h2o-ops
We should add a checkpointing parameter to the Stacked Ensemble function. If you train a Stacked Ensemble with metalearner_algorithm = “GBM” or something that supports checkpointing, we should be able to start-stop the training (feature request from folks who are working on streaming/online models with H2O).
Grab the metalearner model object from the SE object and use it’s checkpointing functionality to re-start training. Then copy the updated metalearner model & metrics into all the right places (including top-level metrics for SE model).