Train XGBoost Model in Sparkling Water

Sparkling Water provides API for H2O XGBoost in both Scala and Python. The following sections describe how to train XGBoost model in Sparkling Water in both languages.

Running XGBoost in Scala

First, let’s start Sparkling Shell as

./bin/sparkling-shell

Start H2O cluster inside the Spark environment

import org.apache.spark.h2o._
import java.net.URI
val hc = H2OContext.getOrCreate(spark)

Parse the data using H2O and convert them to Spark Frame

val frame = new H2OFrame(new URI("https://raw.githubusercontent.com/h2oai/sparkling-water/master/examples/smalldata/prostate/prostate.csv"))
val sparkFrame = hc.asDataFrame(frame)(spark.sqlContext)

Train the model. You can configure all the available XGBoost arguments using provided setters, such as the predictions column.

import org.apache.spark.ml.h2o.algos.H2OXGBoost
val estimator = new H2OXGBoost()(hc, spark.sqlContext).setPredictionCol("AGE")
val model = estimator.fit(sparkFrame)

Run Predictions

model.transform(sparkFrame)

Running XGBoost in Python

First, let’s start PySparkling Shell as

./bin/pysparkling

Start H2O cluster inside the Spark environment

from pysparkling import *
hc = H2OContext.getOrCreate(spark)

Parse the data using H2O and convert them to Spark Frame

import h2o
frame = h2o.import_file("https://raw.githubusercontent.com/h2oai/sparkling-water/master/examples/smalldata/prostate/prostate.csv")
spark_frame = hc.as_spark_frame(frame)

Train the model. You can configure all the available XGBoost arguments using provided setters, such as the predictions column.

from pysparkling.ml import H2OXGBoost
estimator = H2OXGBoost(predictionCol="AGE")
model = estimator.fit(spark_frame)

Run Predictions

model.transform(spark_frame)

XGBoost Memory Configuration

H2O XGBoost uses additionally to Java memory, off-heap memory. This means that it requires some additionally memory available on the system.

When running on YARN, please make sure to set the memoryOverhead so XGBoost has enough memory. On Spark, the following properties might be set

  • spark.yarn.am.memoryOverhead - in case of YARN Cluster deployment
  • spark.yarn.driver.memoryOverhead - in case of YARN client and other deployments
  • spark.yarn.executor.memoryOverhead - in all deployment scenarios

On YARN, the container size is determined by application_memory * memory_overhead. Therefore, by specifying the overhead, we are also allocating some additional off-heap memory which XGBoost can use.

In Spark Standalone Mode or IBM Conductor environment, please make sure to configure the following configurations:

  • spark.memory.offHeap.enabled=true
  • spark.memory.offHeap.size=4G - example of setting this property to 4G of off-heap memory