PMML4S

PMML4S is a PMML (Predictive Model Markup Language) scoring library for Scala. It provides both Scala and Java Evaluator API for PMML.

Features
- Models support
- Transformations support
Installation
- SBT users
- Maven users
Use in Scala
Use in Java
Use PMML in Spark
Use PMML in PySpark
Use PMML in Python
Deploy PMML as REST API
Deploy and Manage AI/ML models at scale
Attentions
Support
License

Features

PMML4S is a lightweight, clean and efficient implementation based on the PMML specification from 2.0 through to the latest 4.4.1.

Models support

It supports the following models:

Not yet supported models:

Transformations support

It supports the following transformations:

Normalization
Discretization
Value mapping
Text Indexing
Functions

Not yet supported transformations:

Aggregation
Lag

Installation

SBT users

libraryDependencies += "org.pmml4s" %%  "pmml4s" % pmml4sVersion

Maven users

<dependency>
  <groupId>org.pmml4s</groupId>
  <artifactId>pmml4s_${scala.version}</artifactId>
  <version>${pmml4s.version}</version>
</dependency>

Use in Scala

PMML4S is really easy to use. Just do one or more of the following:

Load model.

import org.pmml4s.model.Model
import scala.io.Source

// load a model from an IO source that supports various sources, e.g. from a URL locates a PMML model.
val model = Model(Source.fromURL(new java.net.URL("http://dmg.org/pmml/pmml_examples/KNIME_PMML_4.1_Examples/single_iris_dectree.xml")))

import org.pmml4s.model.Model

// load a model from those help methods, e.g. pathname, file object, a string, an array of bytes, or an input stream.
val model = Model.fromFile("single_iris_dectree.xml")

Call predict(values) to predict new values that can be in different types, and the type of results is always same as inputs.

values in a Map:

scala> val result = model.predict(Map("sepal_length" -> 5.1, "sepal_width" -> 3.5, "petal_length" -> 1.4, "petal_width" -> 0.2))
result: Map[String,Any] = Map(probability -> 1.0, probability_Iris-versicolor -> 0.0, probability_Iris-setosa -> 1.0, probability_Iris-virginica -> 0.0, predicted_class -> Iris-setosa, node_id -> 1)

values in a list of pairs of keys and values:

scala> val result = model.predict("sepal_length" -> 5.1, "sepal_width" -> 3.5, "petal_length" -> 1.4, "petal_width" -> 0.2)
result: Seq[(String, Any)] = ArraySeq((predicted_class,Iris-setosa), (probability,1.0), (probability_Iris-setosa,1.0), (probability_Iris-versicolor,0.0), (probability_Iris-virginica,0.0), (node_id,1))

values in an Array:

The order of those values is supposed as same as the input fields list, and the order of results is same as the output fields list.

scala> val inputNames = model.inputNames
inputNames: Array[String] = Array(sepal_length, sepal_width, petal_length, petal_width)

scala> val result = model.predict(Array(5.1, 3.5, 1.4, 0.2))
result: Array[Any] = Array(Iris-setosa, 1.0, 1.0, 0.0, 0.0, 1)

scala> val outputNames = model.outputNames
outputNames: Array[String] = Array(predicted_class, probability, probability_Iris-setosa, probability_Iris-versicolor, probability_Iris-virginica, node_id)

values in the JSON format:

It supports the following styles, and the JSON string can take more than more records to predict, and the results are still a string in JSON with the same format as input.

 - `records` : list like [{column -> value}, … , {column -> value}]
 - `split` : dict like {‘columns’ -> [columns], ‘data’ -> [values]}

scala> val result = model.predict("""[{"sepal_length": 5.1, "sepal_width": 3.5, "petal_length": 1.4, "petal_width": 0.2}, {"sepal_length": 7, "sepal_width": 3.2, "petal_length": 4.7, "petal_width": 1.4}]""")
result: String = [{"probability":1.0,"probability_Iris-versicolor":0.0,"probability_Iris-setosa":1.0,"probability_Iris-virginica":0.0,"predicted_class":"Iris-setosa","node_id":"1"},{"probability":0.9074074074074074,"probability_Iris-versicolor":0.9074074074074074,"probability_Iris-setosa":0.0,"probability_Iris-virginica":0.09259259259259259,"predicted_class":"Iris-versicolor","node_id":"3"}]

scala> val result = model.predict("""{"columns": ["sepal_length", "sepal_width", "petal_length", "petal_width"], "data":[[5.1, 3.5, 1.4, 0.2], [7, 3.2, 4.7, 1.4]]}""")
result: String = {"columns":["predicted_class","probability","probability_Iris-setosa","probability_Iris-versicolor","probability_Iris-virginica","node_id"],"data":[["Iris-setosa",1.0,1.0,0.0,0.0,"1"],["Iris-versicolor",0.9074074074074074,0.0,0.9074074074074074,0.09259259259259259,"3"]]}

values in the PMML4S's Series:

import org.pmml4s.data.Series
import org.pmml4s.util.Utils

// The input schema contains a list of input fields with its name and data type, you can prepare data based on it.
scala> val inputSchema = model.inputSchema
inputSchema: org.pmml4s.common.StructType = StructType(StructField(sepal_length,double), StructField(sepal_width,double), StructField(petal_length,double), StructField(petal_width,double))

// Suppose the row is a record in map from an external columnar data, e.g. a CSV file, or relational database.
val row = Map("sepal_length" -> "5.1", "sepal_width" -> "3.5", "petal_length" -> "1.4", "petal_width" -> "0.2")

// You need to convert the data to the desired type defined by PMML, and keep the same order as defined in the input schema.
val values = inputSchema.map(x => Utils.toDataVal(row(x.name), x.dataType))

scala> val result = model.predict(Series.fromSeq(values))
result: org.pmml4s.data.Series = [Iris-setosa,1.0,1.0,0.0,0.0,1],[(predicted_class,string),(probability,double),(probability_Iris-setosa,double),(probability_Iris-versicolor,double),(probability_Iris-virginica,double),(node_id,string)]

// You can also create a Series with schema, so that values will be accessed by names, the order of values is trivial, e.g.
scala> val result = model.predict(Series.fromSeq(values.reverse, org.pmml4s.common.StructType(inputSchema.fields.reverse)))
result: org.pmml4s.data.Series = [Iris-setosa,1.0,1.0,0.0,0.0,1], [(predicted_class,string),(probability,double),(probability_Iris-setosa,double),(probability_Iris-versicolor,double),(probability_Iris-virginica,double),(node_id,string)]

Which format to use?

You can use any formats of values according to your environment. Except of the Series that need to convert the data explicitly, you don't need to call Utils.toVal explicitly to convert data to ones defined by PMML for others, the conversion will be operated properly automatically. e.g. those input values are string, not double, you can still get the same correct results.

scala> val result = model.predict(Map("sepal_length" -> "5.1", "sepal_width" -> "3.5", "petal_length" -> "1.4", "petal_width" -> "0.2"))
result: Map[String,Any] = Map(probability -> 1.0, probability_Iris-versicolor -> 0.0, probability_Iris-setosa -> 1.0, probability_Iris-virginica -> 0.0, predicted_class -> Iris-setosa, node_id -> 1)

scala> val result = model.predict(Array("5.1", "3.5", "1.4", "0.2"))
result: Array[Any] = Array(Iris-setosa, 1.0, 1.0, 0.0, 0.0, 1)

Understand the result values.

You can see the names of output fields, like predicted_class, probability, actually those names are trivial, the PMML can use any names for those output fields.

The most important attribute of OutputField is feature that specifies the value the output field takes from the computed mining result, about other attributes, see DMG

The Output is optional for PMML, if it's not present, PMML4S will produce some predefined output fields, e.g.

Name	Description
predicted_{target}	Predicted value of the specified target for the supervised model
probability	Probability of the predicted categorical value for the classification model
probability_{category}	Probability of the specified category for the classification model
confidence	Confidence of the specified category for the classification model
node_id	ID of the hit node for the tree model
cluster	Identifier of the winning cluster for the clustering model
cluster_name	Name of the winning cluster for the clustering model
distance	Distance to the predicted entity for the clustering model
similarity	Similarity to the predicted entity for the clustering model
reason_code_{rank}	Reason code of rank for the scorecard model
anomalyScore	Anomaly score of the anomaly detection model

Call outputFields() to get a list of output fields describe result values.

// The output fields describe the result values.
val outputFields = model.outputFields

scala> outputFields.foreach(println)
OutputField(name=predicted_class, displayName=Some(Predicted value of class), dataType=string, opType=nominal, feature=predictedValue, targetField=None, value=None, ruleFeature=consequent, algorithm=exclusiveRecommendation, rank=1, rankBasis=confidence, rankOrder=descending, isMultiValued=false, segmentId=None, isFinalResult=true, decisions=None, expr=None)
OutputField(name=probability, displayName=Some(probability of predicted value), dataType=double, opType=continuous, feature=probability, targetField=None, value=None, ruleFeature=consequent, algorithm=exclusiveRecommendation, rank=1, rankBasis=confidence, rankOrder=descending, isMultiValued=false, segmentId=None, isFinalResult=true, decisions=None, expr=None)
OutputField(name=probability_Iris-setosa, displayName=Some(probability of Iris-setosa), dataType=double, opType=continuous, feature=probability, targetField=None, value=Some(Iris-setosa), ruleFeature=consequent, algorithm=exclusiveRecommendation, rank=1, rankBasis=confidence, rankOrder=descending, isMultiValued=false, segmentId=None, isFinalResult=true, decisions=None, expr=None)
OutputField(name=probability_Iris-versicolor, displayName=Some(probability of Iris-versicolor), dataType=double, opType=continuous, feature=probability, targetField=None, value=Some(Iris-versicolor), ruleFeature=consequent, algorithm=exclusiveRecommendation, rank=1, rankBasis=confidence, rankOrder=descending, isMultiValued=false, segmentId=None, isFinalResult=true, decisions=None, expr=None)
OutputField(name=probability_Iris-virginica, displayName=Some(probability of Iris-virginica), dataType=double, opType=continuous, feature=probability, targetField=None, value=Some(Iris-virginica), ruleFeature=consequent, algorithm=exclusiveRecommendation, rank=1, rankBasis=confidence, rankOrder=descending, isMultiValued=false, segmentId=None, isFinalResult=true, decisions=None, expr=None)
OutputField(name=node_id, displayName=Some(ID of hit node), dataType=string, opType=nominal, feature=entityId, targetField=None, value=None, ruleFeature=consequent, algorithm=exclusiveRecommendation, rank=1, rankBasis=confidence, rankOrder=descending, isMultiValued=false, segmentId=None, isFinalResult=true, decisions=None, expr=None)

Use in Java

It's also easy to use and similar as Scala.

Load model.

import org.pmml4s.model.Model;

Model model = Model.fromFile("single_iris_dectree.xml");

Call predict(values) to predict new values that can be in different types, and the type of results is always same as inputs.

values in a Map of Java:

import java.util.Map;
import java.util.HashMap;

Map result = model.predict(new HashMap<String, Object>() {{
            put("sepal_length", 5.1);
            put("sepal_width", 3.5);
            put("petal_length", 1.4);
            put("petal_width", 0.2);
        }});

values in an Array:

The order of those values is supposed as same as the input fields list, and the order of results is same as the output fields list.

String[] inputNames = model.inputNames();
Object[] result = model.predict(new Double[]{5.1, 3.5, 1.4, 0.2});

values in the JSON format:

It's same as Scala

values in the PMML4S's Series:

import org.pmml4s.data.Series;
import org.pmml4s.util.Utils;
import org.pmml4s.common.StructType;
import org.pmml4s.common.StructField;

// The input schema contains a list of input fields with its name and data type, you can prepare data based on it.
StructType inputSchema = model.inputSchema();

// Suppose the row is a record in map from an external columnar data, e.g. a CSV file, or relational database.
Map row = new HashMap<String, String>() {{
            put("sepal_length", "5.1");
            put("sepal_width", "3.5");
            put("petal_length", "1.4");
            put("petal_width", "0.2");
        }}

// You need to convert the data to the desired type defined by PMML, and keep the same order as defined in the input schema.
Object[] values = new Object[inputSchema.size()];
for (int i = 0; i < values.length; i++) {
  StructField sf = inputSchema.apply(i);
  values[i] = Utils.toVal(row.get(sf.name()), sf.dataType());
}

Series result = model.predict(Series.fromArray(values))

// You can also create a Series with schema, so that values will be accessed by names, the order of values is trivial, e.g.
Series result = model.predict(Series.fromArray(values, inputSchema)))

Understand the result values. See details in Scala above

Use PMML in Spark

See the PMML4S-Spark project. PMML4S-Spark is a PMML scoring library for Spark as SparkML Transformer.

Use PMML in PySpark

See the PyPMML-Spark project. PyPMML-Spark is a Python PMML scoring library for PySpark as SparkML Transformer, it really is the Python API for PMML4s-Spark.

Use PMML in Python

See the PyPMML project. PyPMML is a Python PMML scoring library, it really is the Python API for PMML4S.

Deploy PMML as REST API

See the AI-Serving project. AI-Serving is serving AI/ML models in the open standard formats PMML and ONNX with both HTTP (REST API) and gRPC endpoints.

Deploy and Manage AI/ML models at scale

See the DaaS system that deploys AI/ML models in production at scale on Kubernetes.

Attentions

Thread Safe.

Except of the Association model with transaction, all model objects are read-only, so you are free to use them in a multi-threaded environment.
Scoring Assumptions.

For the regression model with an integer target, float numeric predictions will be produced when there are no round methods specified explicitly.
Use PMML4S library in Android.

You need to use the Scala 2.11 version in Andriod, the Scala core library scala-library of 2.12 and 2.13 depend on the class java.lang.ClassValue that is missing in the java.lang package.

Support

If you have any questions about the PMML4S library, please open issues on this repository.

Feedback and contributions to the project, no matter what kind, are always very welcome.

License

PMML4S is licensed under APL 2.0.

autodeployai/pmml4s