Open christophe-rannou opened 6 years ago
I would like to serialize my preprocessing tasks through PMML.
The "entry point" of the JPMML-Evaluator library is the org.jpmml.evaluator.Evaluator
interface, which requires a backing model element. If your PMML document does not contain any model elements, then you cannot use this "entry point" and either 1) must devise and develop an alternative "entry point" interface (something like org.jpmml.evaluator.Preprocessor
?) or 2) use an alternative library.
Is it possible to have like an Identity MiningModel to return those field without needing a dummy model to evaluate the PMML?
You cannot use the MiningModel
element for that, because it is a wrapper around child model elements.
However, you can use the RegressionModel
element to represent regression-type identity transforms. Just construct the following regression table: y = 1.0 * encoded_feat + 0.0
Reopening this issue, because I might want to do something about it in the upcoming 1.4.X
development branch.
I would like to work on this, is there any pointers you could give me ?
You raised this issue, so you have a use case that needs addressing, not me.
The goal is to design an interface similar to org.jpmml.evaluator.Evaluator
, but for preprocessors. The Evaluator
interface encapsulates models, so it's dealing with model input, target and result fields; every field class has its own specification, etc.
The requirement here is to design a "preprocessor schema". It should have at least two schema query methods Preprocessor#getArgumentFields()
and Preprocessor#getResultFields()
, and the evaluate method Preprocessor#evaluate(Map<FieldName, Object>)
. Anyway, preprocessor's argument fields and result fields are functionally different from standard model fields.
@christophe-rannou : pushing code ? ;-)
Opened a "request for clarification" at DMG.org's issue tracker to have the entry/exit interfaces of transformer-only PMML documents specified: http://mantis.dmg.org/view.php?id=228
Hi, I'd like to ask is this issue resolved? I saw this related PR: https://github.com/jpmml/jpmml-evaluator/pull/116 was closed instead of merged. I wonder if we have any workaround to use transformer-only pmml. Could you please advise? Thanks!
I wonder if we have any workaround to use transformer-only pmml.
The PMML specification does not define such a workflow.
I've asked DMG.org to clarify the situation, but there's been no official response yet (typically takes 2-3 years to obtain it): http://mantis.dmg.org/view.php?id=228
The JPMML software project can always do a vendor extension. But I don't have a clear use case to base my work upon.
I saw this related PR: #116 was closed instead of merged.
We don't do copy&paste programming in this project.
Hi,
I have a question which I am not really sure is PMML related or jpmml-evaluator related. I would like to serialize my preprocessing tasks through PMML. So far I succeded in parsing my preprocessing tasks using the TransformationDictionnary and to output the desired processed field through Output when coupled with a MiningModel (such as a tree). Is it possible to have like an Identity MiningModel to return those field without needing a dummy model to evaluate the PMML ?
The following PMML sums up what I am trying to achieve (not valid since missing a
functionName
):Thanks