Model versioning

[tovbinm]

Problem Exported models neither have any version information nor verification checks (on loading) that verify that a particular model can be safely executed with the current code version.

Solution

1. Include code version in serialized model json file.
2. Implement verification logic that would check if a particular model can be safely executed with the current code version. If not, a proper error should be raised.

Alternatives N/A

[shenzgang]

How do I deploy my model into production as a predictive service?

[tovbinm]

@shenzgang it depends how would you like to run the inference. You can either use Spark runtime to do it or use transmogrifai-local - https://github.com/salesforce/TransmogrifAI/tree/master/local

[shenzgang]

Why does the data need to include label columns when using models for prediction? The training model sample contains feature columns and tag columns, but the data that needs to be predicted has no tag columns, and I had to use a temporary tag column replacement to be successful？

[tovbinm]

How so? I don't think it does. Simply load the model OpWorkflowModel.load("/path/to/model") and then run score() or prepare scoreFunction for local scoring.

[shenzgang]

I used Titan data to get the model and then load the model to predict, I removed the tag column prediction failure! The prediction data must contain the same number of columns as the model training！ my code： val model = OpWorkflowModel.load(modelPath) val scoreFn = model.scoreFunction(spark) val rawData = Seq(//If i remove the label column 'survived',the prediction will fail!! Map("id"->248,"survived"->1,"pclass"->2,"name"->"Hamalainen Mrs. William (Anna)","sex"->"female","age"->24,"sibsp"->0,"parch"->2,"ticket"->"250649","fare"->14.5,"cabin"->null,"embarked"->"S"), Map("id"->249,"survived"->1,"pclass"->1,"name"->"Beckwith Mr. Richard Leonard","sex"->"male","age"->37,"sibsp"->1,"parch"->1,"ticket"->"11751","fare"->52.5542,"cabin"->"D35","embarked"->"S"), Map("id"->250,"survived"->1,"pclass"->2,"name"->"Carter Rev. Ernest Courtenay","sex"->"male","age"->54,"sibsp"->1,"parch"->0,"ticket"->"244252","fare"->26,"cabin"->null,"embarked"->"S") ) val scores = rawData.map(scoreFn) scores.foreach(println(_))