Importing H2O MOJOs from H2O-3

When training algorithm using Sparkling Water API, Sparkling Water always produces H2OMOJOModel. It is also possible to import existing MOJO models into the Sparkling Water ecosystem from H2O-3. Such MOJO models then have the same scoring capabilities as MOJO models trained via Sparkling Water API.

Note: Sparkling Water is backward compatible with MOJO versions produced by different H2O-3 versions.

One advantage of scoring the MOJO artifacts is that H2OContext does not have to be created if we only want to run predictions on MOJOs using Spark. It is important to mention that the format of prediction on MOJOs from Driverless AI differs from predictions on H2O-3 MOJOs. The format of H2O-3 predictions is explained bellow.

Starting a Scoring Environment

First, we need to start a scoring environment for the desired language. There are two variants. We can use Sparkling Water prepared scripts which put required dependencies on the Spark classpath or we can use Spark directly and add the dependencies manually.

Scala

./bin/spark-shell --jars jars/sparkling-water-assembly-scoring_2.11-3.38.0.3-1-2.3-all.jar

If there is a need to train H2O-3/SW models at the same time when we score with existing MOJO models, use jars/sparkling-water-assembly_2.11-3.38.0.3-1-2.3-all.jar instead.

./bin/sparkling-shell

Python

export PYTHONPATH="py/h2o_pysparkling_scoring_2.3-3.38.0.3-1-2.3.zip:$PYTHONPATH" # This line is needed only if the Spark distribution contains the bug SPARK-21945.
./bin/pyspark --py-files py/h2o_pysparkling_scoring_2.3-3.38.0.3-1-2.3.zip

If there is a need to train H2O-3/SW models at the same time when we score with existing MOJO models, use py/h2o_pysparkling_2.3-3.38.0.3-1-2.3.zip instead.

./bin/pysparkling

At this point, we have a Spark interactive terminal where we can carry out predictions. If we don’t require an interactive environment, we can deploy our scoring logic with ./bin/spark-submit. The parameters will be the same as in the example above.

Loading and Usage of H2O-3 MOJO Model

H2O MOJOs can be imported to Sparkling Water from all data sources supported by Apache Spark such as a local file, S3 or HDFS and the semantics of the import is the same as in the Spark API.

When creating a MOJO specified by a relative path and HDFS is enabled, the method attempts to load the MOJO from the HDFS home directory of the current user. In case we are not running on a HDFS-enabled system, we create the mojo from a current working directory.

Scala

import ai.h2o.sparkling.ml.models._
val model = H2OMOJOModel.createFromMojo("prostate_mojo.zip")

Python

from pysparkling.ml import *
model = H2OMOJOModel.createFromMojo("prostate_mojo.zip")

R

library(rsparkling)
sc <- spark_connect(master = "local")
model <- H2OMOJOModel.createFromMojo("prostate_mojo.zip")

An absolute local path can also be used. To create a MOJO model from a locally available MOJO, call:

Scala

import ai.h2o.sparkling.ml.models._
val model = H2OMOJOModel.createFromMojo("/Users/peter/prostate_mojo.zip")

Python

from pysparkling.ml import *
model = H2OMOJOModel.createFromMojo("/Users/peter/prostate_mojo.zip")

R

library(rsparkling)
sc <- spark_connect(master = "local")
model <- H2OMOJOModel.createFromMojo("/Users/peter/prostate_mojo.zip")

Absolute paths on Hadoop can also be used. To create a MOJO model from a MOJO stored on HDFS, call:

Scala

import ai.h2o.sparkling.ml.models._
val model = H2OMOJOModel.createFromMojo("/user/peter/prostate_mojo.zip")

Python

from pysparkling.ml import *
model = H2OMOJOModel.createFromMojo("/user/peter/prostate_mojo.zip")

R

library(rsparkling)
sc <- spark_connect(master = "local")
model <- H2OMOJOModel.createFromMojo("/user/peter/prostate_mojo.zip")

The call loads the mojo file from the following location hdfs://{server}:{port}/user/peter/prostate_mojo.zip, where {server} and {port} is automatically filled in by Spark.

We can also manually specify the type of data source we need to use, in that case, we need to provide the schema:

Scala

import ai.h2o.sparkling.ml.models._
// HDFS
val modelHDFS = H2OMOJOModel.createFromMojo("hdfs:///user/peter/prostate_mojo.zip")
// Local file
val modelLocal = H2OMOJOModel.createFromMojo("file:///Users/peter/prostate_mojo.zip")

Python

from pysparkling.ml import *
# HDFS
modelHDFS = H2OMOJOModel.createFromMojo("hdfs:///user/peter/prostate_mojo.zip")
# Local file
modelLocal = H2OMOJOModel.createFromMojo("file:///Users/peter/prostate_mojo.zip")

R

library(rsparkling)
sc <- spark_connect(master = "local")
 # HDFS
modelHDFS <- H2OMOJOModel.createFromMojo("hdfs:///user/peter/prostate_mojo.zip")
# Local file
modelLocal <- H2OMOJOModel.createFromMojo("file:///Users/peter/prostate_mojo.zip")

The loaded model is an immutable instance, so it’s not possible to change the configuration of the model during its existence. On the other hand, the model can be configured during its creation via H2OMOJOSettings:

Scala

import ai.h2o.sparkling.ml.models._
val settings = H2OMOJOSettings(convertUnknownCategoricalLevelsToNa = true, convertInvalidNumbersToNa = true)
val model = H2OMOJOModel.createFromMojo("prostate_mojo.zip", settings)

Python

from pysparkling.ml import *
settings = H2OMOJOSettings(convertUnknownCategoricalLevelsToNa = True, convertInvalidNumbersToNa = True)
model = H2OMOJOModel.createFromMojo("prostate_mojo.zip", settings)

R

library(rsparkling)
sc <- spark_connect(master = "local")
settings <- H2OMOJOSettings(convertUnknownCategoricalLevelsToNa = TRUE, convertInvalidNumbersToNa = TRUE)
model <- H2OMOJOModel.createFromMojo("prostate_mojo.zip", settings)

To score the dataset using the loaded mojo, call:

Scala

model.transform(dataset)

Python

model.transform(dataset)

R

model$transform(dataset)

In Scala, the createFromMojo method returns a mojo model instance cast as a base class H2OMOJOModel. This class holds only properties common for all mojo models across different Sparkling Water algorithms.

If a Scala user wants to get a property specific for a given MOJO model type, he/she must utilize casting or call the createFromMojo method on the specific MOJO model type.

import ai.h2o.sparkling.ml.models._
val specificModel = H2OGBMMOJOModel.createFromMojo("prostate_mojo.zip")
println(s"Ntrees: ${specificModel.getNTrees()}")

The list of specific MOJO models:

  • H2OXGBoostMOJOModel

  • H2OGBMMOJOModel

  • H2ODRFMOJOModel

  • H2OGLMMOJOModel

  • H2OGAMMOJOModel

  • H2ODeepLearningMOJOModel

  • H2OKMeansMOJOModel

  • H2OIsolationForestMOJOModel

  • H2OExtendedIsolationForestMOJOModel

  • H2OCoxPHMOJOModel

  • H2OTargetEncoderMOJOModel

  • H2OAutoEncoderMOJOModel

Exporting the loaded MOJO model using Sparkling Water

To export the MOJO model, call model.write.save("/some/path"). In case of a Hadoop-enabled system, the command by default uses HDFS. To reference a path on the local file system of the Spark driver, the path must be prefixed with file:// when HDFS is enabled.

Importing the previously exported MOJO model from Sparkling Water

To import the MOJO model, call H2OMOJOModel.read.load("/some/path"). In case of a Hadoop-enabled system, the command by default uses HDFS. To reference a path on the local file system of the Spark driver, the path must be prefixed with file:// when HDFS is enabled.

Accessing additional prediction details

After computing predictions, the prediction column contains in case of classification problem the predicted label and in case regression problem the predicted number. If we need to access more details for each prediction, see the content of a detailed prediction column. By default, the column is named named detailed_prediction. It could contain, for example, predicted probabilities for each predicted label in case of classification problem, Shapley values, and other information.

Customizing the MOJO Settings

We can configure the output and format of predictions via the H2OMOJOSettings. The available options are

  • predictionCol - Specifies the name of the generated prediction column. The default value is prediction.

  • detailedPredictionCol - Specifies the name of the generated detailed prediction column. The detailed prediction column, if enabled, contains additional details, such as probabilities, Shapley values etc. The default value is detailed_prediction.

  • convertUnknownCategoricalLevelsToNa - Enables or disables conversion of unseen categoricals to NAs. By default, it is disabled.

  • convertInvalidNumbersToNa - Enables or disables conversion of invalid numbers to NAs. By default, it is disabled.

  • withContributions - Enables or disables computing Shapley values. Shapley values are generated as a sub-column for the detailed prediction column. Shapley values are supported only by tree-based binomial and regression models.

  • withLeafNodeAssignments - When enabled, a user can obtain the leaf node assignments after the model training has finished. By default, it is disabled.

  • withStageResults - When enabled, a user can obtain the stage results for tree-based models. By default, it is disabled and also it’s not supported by XGBoost although it’s a tree-based algorithm.

  • dataFrameSerializer - A full name of a serializer used for serialization and deserialization of Spark DataFrames to a JSON value within NullableDataFrameParam.

Methods available on MOJO Model

  • See Model Details for methods available on particular model types.

Obtaining Domain Values

To obtain domain values of the trained model, we can run getDomainValues() on the model. This call returns a mapping from a column name to its domain in a form of an array.

Obtaining Model Category

The method getModelCategory can be used to get the model category (such as binomial, multinomial etc).

Obtaining Feature Types

The method getFeatureTypes returns a map/dictionary from a feature name to a corresponding feature type [enum (categorical), numeric, string, etc.]. These pieces helps to understand how individual columns of the training dataset were treated during the model training.

Obtaining Feature Importances

The method getFeatureImportances returns a data frame describing importance of each feature. The importance is expressed by several numbers (Relative Importance, Scaled Importance and Percentage). H2O-3 documentation describes how the numbers are calculated.

Obtaining Scoring History

The method getScoringHistory returns a data frame describing how the model evolved during the training process according to a certain training and validation metrics.

Obtaining Metrics

There are two sets of methods to obtain metrics from the MOJO model.

  1. The first set of methods return a map from the metric name to its double value.

  • getTrainingMetrics() - to obtain training metrics

  • getValidationMetrics() - to obtain validation metrics

  • getCrossValidationMetrics() - to obtain metrics combined from cross-validation holdouts

There is also the method getCurrentMetrics() which gets one of the metrics above based on the following algorithm:

If cross-validation was used, ie, setNfolds was called and the value was higher than zero, this method returns cross-validation metrics. If cross-validation was not used, but the validation frame was used, the method returns validation metrics. The validation frame is used if setSplitRatio was called with the value lower than one. If neither cross-validation nor validation frame was used, this method returns the training metrics.

2. The second set of methods returns typed instances. The instances make individual metrics available via getter methods and the metrics could be also of a complex type. (see Metric Classes for details)

  • getTrainingMetricsObject() - to obtain training metrics

  • getValidationMetricsObject() - to obtain validation metrics

  • getCrossValidationMetricsObject() - to obtain metrics combined from cross-validation holdouts

There is also the method getCurrentMetricsObject() working a similar way as getCurrentMetrics().

Obtaining Cross Validation Metrics Summary

The getCrossValidationMetricsSummary method returns data frame with information about performance of individual folds according to various model metrics.

Obtaining Cross Validation Models

If the model was trained with SW API (i.e. the model wasn’t loaded with the method H2OMOJOModel.createFromMojo()), the algorithm parameter keepCrossValidationModels was set to true and cross-validation was enabled during the training phase, a user can access the sequence cross-validation models by calling the method getCrossValidationModels(). The returned models are regular Sparkling Water MOJO models with model metrics and other important information. [This feature is not available in SW R API.]

Obtaining Leaf Node Assignments

To obtain the leaf node assignments, please first make sure to set withLeafNodeAssignments to true on your MOJO settings object. The leaf node assignments are now stored in the ${detailedPredictionCol}.leafNodeAssignments column on the dataset obtained from the prediction. Please replace ${detailedPredictionCol} with the actual value of your detailed prediction col. By default, it is detailed_prediction.

Obtaining Stage Probabilities

To obtain the stage results, please first make sure to set withStageResults to true on your MOJO settings object. The stage results for regression and anomaly detection problems are stored in the ${detailedPredictionCol}.stageResults on the dataset obtained from the prediction. The stage results for classification (binomial, multinomial) problems are stored under ${detailedPredictionCol}.stageProbabilities Please replace ${detailedPredictionCol} with the actual value of your detailed prediction col. By default, it is detailed_prediction.

The stage results are an array of values, where a value at the position t is the prediction/probability combined from contributions of trees T1, T2, …, Tt. For t equal to a number of model trees, the value is the same as the final prediction/probability. The stage results (probabilities) for the classification problem are represented by a list of columns, where one column contains stage probabilities for a given prediction class.