JPMML-Spark

PMML evaluator library for the Apache Spark cluster computing system (http://spark.apache.org/).

Features

Full support for PMML specification versions 3.0 through 4.2. The evaluation is handled by the [JPMML-Evaluator] (https://github.com/jpmml/jpmml-evaluator) library.

Prerequisites

Apache Spark version 1.5.X or 1.6.X.

Installation

Enter the project root directory and build using [Apache Maven] (http://maven.apache.org/):

mvn clean install

The build produces two JAR files:

pmml-spark/target/pmml-spark-1.0-SNAPSHOT.jar - Library JAR file.
pmml-spark-example/target/example-1.0-SNAPSHOT.jar - Example application JAR file.

Usage

Library

Building a generic transformer based on a PMML document in local filesystem:

File pmmlFile = ...;

Evaluator evaluator = EvaluatorUtil.createEvaluator(pmmlFile);

TransformerBuilder pmmlTransformerBuilder = new TransformerBuilder(evaluator)
	.withTargetCols()
	.withOutputCols()
	.exploded(false);

Transformer pmmlTransformer = transformerBuilder.build();

Building an Apache Spark ML-style regressor when the PMML document is known to contain a regression model (eg. auto-mpg dataset):

TransformerBuilder pmmlTransformerBuilder = new TransformerBuilder(evaluator)
	.withLabelCol("MPG") // Double column
	.exploded(true);

Building an Apache Spark ML-style classifier when the PMML document is known to contain a classification model (eg. iris-species dataset):

TransformerBuilder pmmlTransformerBuilder = new TransformerBuilder(evaluator)
	.withLabelCol("Species") // String column
	.withProbabilityCol("Species_probability", Arrays.asList("setosa", "versicolor", "virginica")) // Vector column
	.exploded(true);

Scoring data:

DataFrame input = ...;
DataFrame output = pmmlTransformer.transform(input);

In default mode, the transformation appends an intermediary "pmml" column to the data frame, which contains all the requested result columns:

root
 |-- Sepal_Length: double (nullable = true)
 |-- Sepal_Width: double (nullable = true)
 |-- Petal_Length: double (nullable = true)
 |-- Petal_Width: double (nullable = true)
 |-- pmml: struct (nullable = true)
 |    |-- Species: string (nullable = false)
 |    |-- Species_probability: vector (nullable = false)

In exploded mode, the transformation appends all the requested result columns to the data frame:

root
 |-- Sepal_Length: double (nullable = true)
 |-- Sepal_Width: double (nullable = true)
 |-- Petal_Length: double (nullable = true)
 |-- Petal_Width: double (nullable = true)
 |-- Species: string (nullable = false)
 |-- Species_probability: vector (nullable = false)

A note about building and packaging JPMML-Spark applications. The JPMML-Evaluator library depends on JPMML-Model and Google Guava library versions that are in conflict with the ones that are bundled with Apache Spark and/or Apache Hadoop. This conflict can be easily solved by relocating JPMML-Evaluator library dependencies to a different namespace using the [Apache Maven Shade Plugin] (https://maven.apache.org/plugins/maven-shade-plugin/). Please see the JPMML-Spark example application for a worked out example.

Example application

The example application JAR file contains an executable class org.jpmml.spark.EvaluationExample.

This class expects three command-line arguments:

The path of the model PMML file in local filesystem.
The path of the input CSV file in local filesystem.
The path of the output directory in local filesystem.

For example:

spark-submit --master local --class org.jpmml.spark.EvaluationExample example-1.0-SNAPSHOT.jar DecisionTreeIris.pmml Iris.csv /tmp/DecisionTreeIris

License

JPMML-Spark is licensed under the [GNU Affero General Public License (AGPL) version 3.0] (http://www.gnu.org/licenses/agpl-3.0.html). Other licenses are available on request.

Additional information

Please contact [info@openscoring.io] (mailto:info@openscoring.io)

manugarri/jpmml-spark