Exporting Models with DataRobot Prime

cancel
Showing results for 
Search instead for 
Did you mean: 

Exporting Models with DataRobot Prime

(Updated February 2021)

This article showcases how you can use DataRobot Prime to generate rules that approximate models built with DataRobot.

Overview

DataRobot Prime can be initiated on most DataRobot models (see caveats below). It works by creating a series of rules that approximate the original model. Once created, the Prime model can be exported as a Python module or Java class and then be deployed outside of DataRobot.

Creating a DataRobot Prime model

To create a DataRobot Prime model, click on the model of interest and select Predict > DataRobot Prime.

lhaviland_0-1612469025734.png

In the displayed tab, click Run DataRobot Prime. This will initiate a new model building job; the completed model will appear on the Leaderboard (Figure 2).

lhaviland_1-1612469486181.png

The newly created model includes some functionality that is not included in other, non-Prime DataRobot models. To access that functionality, navigate to the DataRobot Prime tab for the related DataRobot Prime model (Figure 3).

lhaviland_3-1612469821269.png

In that tab, you see a graph for rules accuracy. The X-axis indicates the number of rules and the Y-axis indicates the accuracy metric currently selected. DataRobot will have optimized the number of rules for the current model based on the accuracy metric but you can manually choose how many rules the final model should have and then kick off another modeling job (Figure 4).

lhaviland_4-1612470272733.png

Exporting the model

Once satisfied with the results, you can download the scoring code by clicking Generate and download code. The code can then be used in the environment of your preference.

Caveats

Some important limitations to consider when trying to create a DataRobot Prime model:

  • A DataRobot Prime model cannot be created when the original model: was trained into the validation data; was run using a feature list that contains any user transformations other than var type transformations; has only a single text feature in the feature list; or is part of a multiclass project.
  • Date/time partitioning is not available.
  • Are not displayed on the Learning Curve but do display on Speed vs Accuracy.
  • Must be run on the same feature list, and at the same sample size, as the original model.
  • You cannot manually launch cross-validation from a DataRobot Prime model.
  • You must run the model with enough data left to include a validation set.
  • Date-time format checking is not the same as the other prediction mechanisms. Make sure to verify the date-time formats are the same between training data and prediction data before running predictions.
  • Not available when Exposure or Offset parameters are set (Show advanced options > Additional).

More information

If you're a licensed DataRobot customer, search in-app Platform Documentation for DataRobot Prime tab or DataRobot Prime.

Labels (2)
Comments
andreww
Computer Board

We have DataRobot on-prem, when I try to export the model i only see options for Download and Deploy. I do not see the deploy to hadoop or deploy to prime options. Is this option not available on on-prem?

doyouevendata
DC Motor

Hi @andreww - Deploy to Hadoop is only available when the DataRobot installation is actually made on top of Hadoop.  If your install is on separate machines, it won't be available; however there are ways to score data from Hadoop.  This could include moving a csv from HDFS out to elsewhere to submit it for a scoring job for example.  It could also be queried out via Hive.  Another option could be using the scoring code/codegen functionality to take a java jar and use it to score via Spark dataframes.  It and DataRobot Prime are both licensed options, although your admin may have also not granted you permission to use them.  There are feature flags under the settings for your user account that they could flip on to allow you to leverage them.

Version history
Last update:
‎02-04-2021 03:36 PM
Updated by:
Contributors