Importing MOJO Pipelines from Driverless AI¶
MOJO scoring pipeline artifacts, created in Driverless AI, can be used in Spark to carry out predictions in parallel using the Sparkling Water API. This section shows how to load and run predictions on the MOJO scoring pipeline in Sparkling Water.
Note: Sparkling Water is backward compatible with MOJO versions produced by different Driverless AI 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 also important to mention that the format of prediction on MOJOs
from Driverless AI differs from predictions on H2O-3 MOJOs. The format of Driverless AI prediction is explained bellow.
To use the MOJO scoring pipeline, a Driverless AI license has to be passed to Spark.
This can be achieved via
--jars argument of the Spark launcher scripts.
Note: In Local Spark mode, please use
--driver-class-path to specify the path to the license file.
We also need Sparkling Water distribution which can be obtained from H2O Download page. After we downloaded the Sparkling Water distribution, extract it, and go to the extracted directory.
The MOJO scoring pipeline is included in the
mojo.zip archive downloaded from Driverless AI. The archive also contains
a mojo runtime library, examples and other files. The only file which is important for scoring with Sparkling Water is
pipeline.mojo. Thus before running Sparkling Water, extract the
mojo.zip archive and copy
pipeline.mojo to a
location of your choice.
unzip /path/to/mojo.zip -d /tmp/mojo.zip.extracted cp /tmp/mojo.zip.extracted/mojo-pipeline/pipeline.mojo /path/to/pipeline.mojo
Starting a Scoring Environment¶
First, we need to start a scoring environment for the desired language with a Driverless AI license. 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.
./bin/spark-shell --jars license.sig,jars/sparkling-water-assembly_2.11-126.96.36.199-1-2.1-all.jar
./bin/sparkling-shell --jars license.sig
./bin/pyspark --jars license.sig --py-files py/build/dist/h2o_pysparkling_2.1-188.8.131.52-1-2.1.zip
./bin/pysparkling --jars license.sig
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 Driveless AI MOJO Model¶
The Pipeline MOJO model could be loaded as:
import ai.h2o.sparkling.ml.models.H2OMOJOPipelineModel val settings = H2OMOJOSettings(predictionCol = "fruit_type", convertUnknownCategoricalLevelsToNa = true) val mojo = H2OMOJOPipelineModel.createFromMojo("file:///path/to/pipeline.mojo", settings)
from pysparkling.ml import H2OMOJOPipelineModel settings = H2OMOJOSettings(predictionCol = "fruit_type", convertUnknownCategoricalLevelsToNa = True) mojo = H2OMOJOPipelineModel.createFromMojo("file:///path/to/pipeline.mojo", settings)
In the examples above
settings is an optional argument. If it’s not specified, the default values are used.
Prepare the dataset to score on:
val dataFrame = spark.read.option("header", "true").option("inferSchema", "true").csv("file:///path/to/data.csv")
dataFrame = spark.read.option("header", "true").option("inferSchema", "true").csv("file:///path/to/data.csv")
And finally, score the mojo on the loaded dataset:
val predictions = mojo.transform(dataFrame)
predictions = mojo.transform(dataFrame)
We can select the predictions as:
The output data frame contains all the original columns plus the prediction column which is by default named
prediction. The prediction column contains all the prediction detail. Its name can be modified via the
When the option
namedMojoOutputColumns is enabled on
predictionCol contains sub-columns with
names corresponding to the columns Driverless AI identified as output columns. For example, if Driverless API MOJO
pipeline contains one output column AGE ( for example regression problem), the prediction column contains another sub-column
named AGE. If The MOJO pipeline contains multiple output columns, such as VALID.0 and VALID.1 (for example classification problems),
the prediction column contains two sub-columns with the aforementioned names.
If this option is disabled, the
predictionCol contains the array of predictions without
the column names. For example, if the Driverless API MOJO pipeline contains one output column AGE ( for example regression problem),
the prediction column contains an array of size 1 with the predicted value.
If The MOJO pipeline contains multiple output columns, such as VALID.0 and VALID.1 (for example classification problems),
the prediction column contains an array of size 2 containing predicted probabilities for each class.
By default, this option is enabled.
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.
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.
namedMojoOutputColumns- Enables or disables named output columns. By default, it is enabled.
If you see the following exception during loading the MOJO pipeline:
java.io.IOException: MOJO doesn't contain resource mojo/pipeline.pb, then it means you are adding incompatible mojo-runtime.jar on your classpath. It is not required and also not suggested to put the JAR on the classpath as Sparkling Water already bundles the correct dependencies.
java.io.IOException: None of 2 available pipeline factories [pbuf, toml] can read this mojo., then you most-likely passed the whole
mojo.ziparchive to the createFromMojo method instead of the
pipeline.mojofile, which is contained in the archive.
A similar error to
java.lang.ClassCastException: Mojo column of type Float32 can be assigned Java values only from the following types: [class java.lang.Short, class java.lang.Double, class java.lang.Byte, class java.lang.Integer, class java.lang.Float], Java class on the input was: Long, then call the method
getFeatureTypes()to get a map/dictionary from feature names to expected types. Identify a feature with the expected type
Float32` and ``Long` type in the dataset for scoring and manually cast the feature column from ``Longto