Download Features

Download Features

Hi ,

Once final model is ready for deployment in DataRobot,  Can we download the features  and the feature creation code only  from DataRobot ? The reason being if I need to score externally then wont I need the feature creations code ?


Labels (2)
0 Kudos
1 Reply

Hi PK -

The feature transformation pipeline for a DataRobot model is included in the model package. For scoring, the only steps you need to replicate are those you did outside of DataRobot when you originally created the project. Steps internal to the project (imputing missing values or one-hot encoding, for example) will be done at scoring time internally in the DataRobot model.

The model itself is most commonly hosted via the Prediction API on dedicated scoring resources on a DataRobot cluster. This API can be used in a real-time approach and can be used for batch scoring input csv files through various supporting tools and capabilities the platform provides.

It is possible to export the model to external infrastructure. If this is performed, additional steps need to take place to monitor that externally deployed model for the metrics that show up under model management, including features like tracking drift over time.

There are 3 methods of exporting this model:

  1. DataRobot Prime - this is a java or python rules based approximation of a model, available for most models.
  2. DataRobot Scoring Code - this is a compiled java binary jar file of the exact model.  The source code is available as well, although not all libraries will be available to compile it.  This works for many models but there is a subset that is not supported, Support Vector Machines (SVMs) as of the moment for example.
  3. A Portable Prediction Server (PPS) - a docker container representing the exact model, with full additional features (this is the only option of this set which can provide prediction explanations, for example.)

Note the exportable options are available depending on the licensing of the platform; although these options are not available in the AI Platform Trial.