Train XGBoost Model in Sparkling Water

Sparkling Water provides API for H2O XGBoost in Scala and Python. The following sections describe how to train the XGBoost model in Sparkling Water in both languages. See also Parameters of H2OXGBoost and Details of H2OXGBoostMOJOModel.

Scala

First, let’s start Sparkling Shell as

./bin/sparkling-shell

Start H2O cluster inside the Spark environment

import ai.h2o.sparkling._
import java.net.URI
val hc = H2OContext.getOrCreate()

Parse the data using H2O and convert them to Spark Frame

import org.apache.spark.SparkFiles
spark.sparkContext.addFile("https://raw.githubusercontent.com/h2oai/sparkling-water/master/examples/smalldata/prostate/prostate.csv")
val rawSparkDF = spark.read.option("header", "true").option("inferSchema", "true").csv(SparkFiles.get("prostate.csv"))
val sparkDF = rawSparkDF.withColumn("CAPSULE", $"CAPSULE" cast "string")
val Array(trainingDF, testingDF) = sparkDF.randomSplit(Array(0.8, 0.2))

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

import ai.h2o.sparkling.ml.algos.H2OXGBoost
val estimator = new H2OXGBoost().setLabelCol("CAPSULE")
val model = estimator.fit(trainingDF)

By default, the H2OXGBoost algorithm distinguishes between a classification and regression problem based on the type of the label column of the training dataset. If the label column is a string column, a classification model will be trained. If the label column is a numeric column, a regression model will be trained. If you don’t want be worried about column data types, you can explicitly identify the problem by using ai.h2o.sparkling.ml.algos.classification.H2OXGBoostClassifier or ai.h2o.sparkling.ml.algos.regression.H2OXGBoostRegressor instead.

Run Predictions

model.transform(testingDF).show(false)

You can also get model details via calling methods listed in Details of H2OXGBoostMOJOModel.

Python

First, let’s start PySparkling Shell as

./bin/pysparkling

Start H2O cluster inside the Spark environment

from pysparkling import *
hc = H2OContext.getOrCreate()

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")
sparkDF = hc.asSparkFrame(frame)
sparkDF = sparkDF.withColumn("CAPSULE", sparkDF.CAPSULE.cast("string"))
[trainingDF, testingDF] = sparkDF.randomSplit([0.8, 0.2])

Train the model. You can configure all the available XGBoost arguments using provided setters or constructor parameters, such as the label column.

from pysparkling.ml import H2OXGBoost
estimator = H2OXGBoost(labelCol = "CAPSULE")
model = estimator.fit(trainingDF)

By default, the H2OXGBoost algorithm distinguishes between a classification and regression problem based on the type of the label column of the training dataset. If the label column is a string column, a classification model will be trained. If the label column is a numeric column, a regression model will be trained. If you don’t want to be worried about column data types, you can explicitly identify the problem by using H2OXGBoostClassifier or H2OXGBoostRegressor instead.

Run Predictions

model.transform(testingDF).show(truncate = False)

You can also get model details via calling methods listed in Details of H2OXGBoostMOJOModel.

XGBoost Memory Configuration

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

When running on YARN or Kubernetes, please make sure to set the spark.yarn.executor.memoryOverhead so XGBoost has enough native memory on executors. It’s recommended to set the property to 12O% of the value set in spark.executor.memory.

Note: spark.yarn.executor.memoryOverhead must be set in MiB.

Example

If you set spark.executor.memory to 10g, spark.yarn.executor.memoryOverhead should be set to 12288. The size of the corresponding YARN or Kubernetes container will be at least 22 GiB.

Note: In case of Pysparkling, the YARN container will be bigger about the memory required by the Python process.

Memory Overhead on Spark driver

If you enabled a H2O client (a special H2O node representing an entry point for the communication with the H2O cluster) to run on the Spark driver, you should also set the following properties in the same way as spark.yarn.executor.memoryOverhead.

  • spark.yarn.am.memoryOverhead - in case of deploying to YARN in the client mode

  • spark.yarn.driver.memoryOverhead - in case of deploying to YARN in the cluster mode and other deployments (Kubernetes, Mesos)

Note: A H2O client can run on the Spark driver only with Sparkling Water in Scala/Java API and the property spark.ext.h2o.rest.api.based.client set to false. The default value of the property is true.