/jpmml-sparkml

Java library and command-line application for converting Apache Spark ML pipelines to PMML

Primary LanguageJavaGNU Affero General Public License v3.0AGPL-3.0

JPMML-SparkML

Java library and command-line application for converting Apache Spark ML pipelines to PMML.

Features

Prerequisites

  • Apache Spark version 1.5.X, 1.6.X, 2.0.X, 2.1.X, 2.2.X or 2.3.X.

Installation

Library

JPMML-SparkML library JAR file (together with accompanying Java source and Javadocs JAR files) is released via Maven Central Repository.

The current version is 1.4.5 (26 June, 2018).

<dependency>
	<groupId>org.jpmml</groupId>
	<artifactId>jpmml-sparkml</artifactId>
	<version>1.4.5</version>
</dependency>

Compatibility matrix:

JPMML-SparkML version Apache Spark version PMML version
1.0.0 through 1.0.9 1.5.X and 1.6.X 4.2
1.1.0 2.0.X 4.2
1.1.1 through 1.1.20 2.0.X 4.3
1.2.0 through 1.2.12 2.1.X 4.3
1.3.0 through 1.3.8 2.2.X 4.3
1.4.0 through 1.4.5 2.3.X 4.3

JPMML-SparkML depends on the latest and greatest version of the JPMML-Model library, which is in conflict with the legacy version that is part of the Apache Spark distribution.

This conflict is documented in SPARK-15526.

Modifying Apache Spark installation

The embodiment of the legacy version of the JPMML-Model library:

  • $SPARK_HOME/jars/pmml-model-1.2.15.jar
  • $SPARK_HOME/jars/pmml-schema-1.2.15.jar

Removing these two JAR files will solve all conflicts for all applications forever.

Compile-time conflict resolution

Excluding the legacy version of the JPMML-Model library:

<dependency>
	<groupId>org.apache.spark</groupId>
	<artifactId>spark-mllib_2.11</artifactId>
	<version>${spark.version}</version>
	<scope>provided</scope>
	<exclusions>
		<exclusion>
			<groupId>org.jpmml</groupId>
			<artifactId>pmml-model</artifactId>
		</exclusion>
	</exclusions>
</dependency>

Run-time conflict resolution

Using the Maven Shade Plugin to relocate all org.dmg.pmml.* and org.jpmml.* classes of the latest and greatest version of the JPMML-Model library to a different namespace (aka "shading"):

<plugin>
	<groupId>org.apache.maven.plugins</groupId>
	<artifactId>maven-shade-plugin</artifactId>
	<version>${maven.shade.version}</version>
	<executions>
		<execution>
			<phase>package</phase>
			<goals>
				<goal>shade</goal>
			</goals>
			<configuration>
				<relocations>
					<relocation>
						<pattern>org.dmg.pmml</pattern>
						<shadedPattern>org.shaded.dmg.pmml</shadedPattern>
					</relocation>
					<relocation>
						<pattern>org.jpmml</pattern>
						<shadedPattern>org.shaded.jpmml</shadedPattern>
					</relocation>
				</relocations>
			</configuration>
		</execution>
	</executions>
</plugin>

The downside of shading is that such relocated classes are incompatible with other JPMML APIs. For example, the PMMLBuilder#build() builder method would start returning org.shaded.dmg.pmml.PMML object instances, which are not valid substitutes for org.dmg.pmml.PMML object instances.

Example application

Enter the project root directory and build using Apache Maven:

mvn clean install

The build produces two JAR files:

  • target/jpmml-sparkml-1.4-SNAPSHOT.jar - Library JAR file.
  • target/jpmml-sparkml-executable-1.4-SNAPSHOT.jar - Example application JAR file.

Usage

Library

Fitting a Spark ML pipeline that only makes use of supported Transformer types:

DataFrame irisData = ...;

StructType schema = irisData.schema();

RFormula formula = new RFormula()
	.setFormula("Species ~ .");

DecisionTreeClassifier classifier = new DecisionTreeClassifier()
	.setLabelCol(formula.getLabelCol())
	.setFeaturesCol(formula.getFeaturesCol());

Pipeline pipeline = new Pipeline()
	.setStages(new PipelineStage[]{formula, classifier});

PipelineModel pipelineModel = pipeline.fit(irisData);

Converting the Spark ML pipeline to PMML using the org.jpmml.sparkml.PMMLBuilder builder class:

PMML pmml = new PMMLBuilder(schema, pipelineModel)
	.build();

// Viewing the result
JAXBUtil.marshalPMML(pmml, new StreamResult(System.out));

Example application

The example application JAR file contains an executable class org.jpmml.sparkml.Main, which can be used to convert a pair of serialized org.apache.spark.sql.types.StructType and org.apache.spark.ml.PipelineModel objects to PMML.

The example application JAR file does not include Apache Spark runtime libraries. Therefore, this executable class must be executed using Apache Spark's spark-submit helper script.

For example, converting a pair of Spark ML schema and pipeline serialization files src/test/resources/schema/Iris.json and src/test/resources/pipeline/DecisionTreeIris.zip, respectively, to a PMML file DecisionTreeIris.pmml:

spark-submit --master local --class org.jpmml.sparkml.Main target/jpmml-sparkml-executable-1.4-SNAPSHOT.jar --schema-input src/test/resources/schema/Iris.json --pipeline-input src/test/resources/pipeline/DecisionTreeIris.zip --pmml-output DecisionTreeIris.pmml

Getting help:

spark-submit --master local --class org.jpmml.sparkml.Main target/jpmml-sparkml-executable-1.4-SNAPSHOT.jar --help

License

JPMML-SparkML is dual-licensed under the GNU Affero General Public License (AGPL) version 3.0, and a commercial license.

Additional information

JPMML-SparkML is developed and maintained by Openscoring Ltd, Estonia.

Interested in using JPMML software in your application? Please contact info@openscoring.io