Open issues

Importing H2O Binary Models from older versions

Description

The problem

Various H2O Models can be downloaded in two forms:

  1. Binary format ([documentation|http://docs.h2o.ai/h2o/latest-stable/h2o-docs/save-and-load-model.html]),

  2. MOJO ([documentation|http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/mojo_import.html]).

When a model is downloaded in the above-mentioned binary format, it is tied to a specific version of H2O. An attempt to import such model into any other version of H2O is going to result in an error. However, there are cases when the model is finished and is only required to be used for scoring, or extraction of information about the model. For information on how to import an existing model into a different version of H2O and use a subset of the original H2O model’s functionality, please refer to the next section named “The solution”.


Reasoning


There are multiple reasons to for incompatibility of H2O models among different versions of H2O. Primarily, parameters of models between versions may change and the behavior of any of the available models improved, resulting in change of expected behavior and mutable parameter incompatibilities. Secondly, the are many other features link to the model, like checkpointing. Using a checkpoint and continuing on a different level is not possible.

The solution

H2O Models can be exported in a[ MOJO format|http://docs.h2o.ai/h2o/latest-stable/h2o-docs/save-and-load-model.html], primarily aimed towards productionizing H2O model. H2O supports loading MOJO models back into H2O and use a *subset* of features available to a full-fledged model. This way, the version limitation of traditional binary model can be evaded.

For the MOJOs imported into H2O, scoring is available. Basic model information is also available for all the importable MOJOs. In case of selected models, H2O is even able to display model metrics.

For more information about MOJOs, please refer to the [H2O Open Source documentation | http://docs.h2o.ai/h2o/latest-stable/h2o-docs/productionizing.html#about-pojos-and-mojos]. Or refer directly to the guide on [how to import MOJOs into H2O|http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/mojo_import.html].

Environment

None

Status

Assignee

Pavel Pscheidl

Reporter

Pavel Pscheidl

Labels

None

Release Priority

None

CustomerVisible

None

testcase 1

None

testcase 2

None

testcase 3

None

h2ostream link

None

Affected Spark version

None

AffectedContact

None

AffectedCustomers

None

AffectedPilots

None

AffectedOpenSource

None

Support Assessment

None

Customer Request Type

None

Support ticket URL

None

End date

None

Baseline start date

None

Baseline end date

None

Task progress

None

Task mode

None

Priority

Major