Migration to SparklingWater 3.32 - Missing Features

11 views
Skip to first unread message

Gary Geniesse

unread,
Mar 4, 2021, 10:59:59 AMMar 4
to H2O Open Source Scalable Machine Learning - h2ostream
I am trying to migrate a legacy H2O Sparkling Water system in Scala to version 3.32.

In the migration guide, it seems to indicate to shift from the old org.apache.spark.h2o, water, and hex package resources such as GBMModel and MojoModel to the new ai.h2o.sparkling package resources like H2OGBM and H2OMOJOModel, as well as alternate versions of H2OConf, H2OContext, H2OFrame, etc.

The issue I'm running into is that much of the information exposed in the hex package does not appear to be available through the new resources. I understood from the migration guide that the support is incomplete based on the deprecation notes on things like H2OContext and that it would likely be completed in 3.34. I just wanted to verify that I wasn't overlooking key functionality/information.

For example, I have not been able to find a way to access _output values such as _scoring_history  and _varimp. Similarly, the metrics available from the H2OMOJOModel seem very minimal and I have not been able to find analogs to ModelMetrics classes like ModelMetricsBinomial where I can access things like the full auc data and confusion matrix. And finally, I did not see any support for things like PCA algorithms.

Assuming I'm not overlooking things in 3.32, are all of these things slated for inclusion in 3.34? I'm a bit confused about how we're supposed to follow the migration guide suggestions without a clear picture of how to access analogous functionality.
Reply all
Reply to author
Forward
0 new messages