ML Pipelines

\[ \newcommand{\R}{\mathbb{R}} \newcommand{\E}{\mathbb{E}} \newcommand{\x}{\mathbf{x}} \newcommand{\y}{\mathbf{y}} \newcommand{\wv}{\mathbf{w}} \newcommand{\av}{\mathbf{\alpha}} \newcommand{\bv}{\mathbf{b}} \newcommand{\N}{\mathbb{N}} \newcommand{\id}{\mathbf{I}} \newcommand{\ind}{\mathbf{1}} \newcommand{\0}{\mathbf{0}} \newcommand{\unit}{\mathbf{e}} \newcommand{\one}{\mathbf{1}} \newcommand{\zero}{\mathbf{0}} \]

In this section, we introduce the concept of ML Pipelines. ML Pipelines provide a uniform set of high-level APIs built on top of DataFrames that help users create and tune practical machine learning pipelines.

Table of Contents

Main concepts in Pipelines

MLlib standardizes APIs for machine learning algorithms to make it easier to combine multiple algorithms into a single pipeline, or workflow. This section covers the key concepts introduced by the Pipelines API, where the pipeline concept is mostly inspired by the scikit-learn project.

DataFrame

Machine learning can be applied to a wide variety of data types, such as vectors, text, images, and structured data. This API adopts the DataFrame from Spark SQL in order to support a variety of data types.

DataFrame supports many basic and structured types; see the Spark SQL datatype reference for a list of supported types. In addition to the types listed in the Spark SQL guide, DataFrame can use ML Vector types.

A DataFrame can be created either implicitly or explicitly from a regular RDD. See the code examples below and the Spark SQL programming guide for examples.

Columns in a DataFrame are named. The code examples below use names such as “text”, “features”, and “label”.

Pipeline components

Transformers

A Transformer is an abstraction that includes feature transformers and learned models. Technically, a Transformer implements a method transform(), which converts one DataFrame into another, generally by appending one or more columns. For example:

Estimators

An Estimator abstracts the concept of a learning algorithm or any algorithm that fits or trains on data. Technically, an Estimator implements a method fit(), which accepts a DataFrame and produces a Model, which is a Transformer. For example, a learning algorithm such as LogisticRegression is an Estimator, and calling fit() trains a LogisticRegressionModel, which is a Model and hence a Transformer.

Properties of pipeline components

Transformer.transform()s and Estimator.fit()s are both stateless. In the future, stateful algorithms may be supported via alternative concepts.

Each instance of a Transformer or Estimator has a unique ID, which is useful in specifying parameters (discussed below).

Pipeline

In machine learning, it is common to run a sequence of algorithms to process and learn from data. E.g., a simple text document processing workflow might include several stages:

MLlib represents such a workflow as a Pipeline, which consists of a sequence of PipelineStages (Transformers and Estimators) to be run in a specific order. We will use this simple workflow as a running example in this section.

How it works

A Pipeline is specified as a sequence of stages, and each stage is either a Transformer or an Estimator. These stages are run in order, and the input DataFrame is transformed as it passes through each stage. For Transformer stages, the transform() method is called on the DataFrame. For Estimator stages, the fit() method is called to produce a Transformer (which becomes part of the PipelineModel, or fitted Pipeline), and that Transformer’s transform() method is called on the DataFrame.

We illustrate this for the simple text document workflow. The figure below is for the training time usage of a Pipeline.

ML Pipeline Example

Above, the top row represents a Pipeline with three stages. The first two (Tokenizer and HashingTF) are Transformers (blue), and the third (LogisticRegression) is an Estimator (red). The bottom row represents data flowing through the pipeline, where cylinders indicate DataFrames. The Pipeline.fit() method is called on the original DataFrame, which has raw text documents and labels. The Tokenizer.transform() method splits the raw text documents into words, adding a new column with words to the DataFrame. The HashingTF.transform() method converts the words column into feature vectors, adding a new column with those vectors to the DataFrame. Now, since LogisticRegression is an Estimator, the Pipeline first calls LogisticRegression.fit() to produce a LogisticRegressionModel. If the Pipeline had more Estimators, it would call the LogisticRegressionModel’s transform() method on the DataFrame before passing the DataFrame to the next stage.

A Pipeline is an Estimator. Thus, after a Pipeline’s fit() method runs, it produces a PipelineModel, which is a Transformer. This PipelineModel is used at test time; the figure below illustrates this usage.

ML PipelineModel Example

In the figure above, the PipelineModel has the same number of stages as the original Pipeline, but all Estimators in the original Pipeline have become Transformers. When the PipelineModel’s transform() method is called on a test dataset, the data are passed through the fitted pipeline in order. Each stage’s transform() method updates the dataset and passes it to the next stage.

Pipelines and PipelineModels help to ensure that training and test data go through identical feature processing steps.

Details

DAG Pipelines: A Pipeline’s stages are specified as an ordered array. The examples given here are all for linear Pipelines, i.e., Pipelines in which each stage uses data produced by the previous stage. It is possible to create non-linear Pipelines as long as the data flow graph forms a Directed Acyclic Graph (DAG). This graph is currently specified implicitly based on the input and output column names of each stage (generally specified as parameters). If the Pipeline forms a DAG, then the stages must be specified in topological order.

Runtime checking: Since Pipelines can operate on DataFrames with varied types, they cannot use compile-time type checking. Pipelines and PipelineModels instead do runtime checking before actually running the Pipeline. This type checking is done using the DataFrame schema, a description of the data types of columns in the DataFrame.

Unique Pipeline stages: A Pipeline’s stages should be unique instances. E.g., the same instance myHashingTF should not be inserted into the Pipeline twice since Pipeline stages must have unique IDs. However, different instances myHashingTF1 and myHashingTF2 (both of type HashingTF) can be put into the same Pipeline since different instances will be created with different IDs.

Parameters

MLlib Estimators and Transformers use a uniform API for specifying parameters.

A Param is a named parameter with self-contained documentation. A ParamMap is a set of (parameter, value) pairs.

There are two main ways to pass parameters to an algorithm:

  1. Set parameters for an instance. E.g., if lr is an instance of LogisticRegression, one could call lr.setMaxIter(10) to make lr.fit() use at most 10 iterations. This API resembles the API used in spark.mllib package.
  2. Pass a ParamMap to fit() or transform(). Any parameters in the ParamMap will override parameters previously specified via setter methods.

Parameters belong to specific instances of Estimators and Transformers. For example, if we have two LogisticRegression instances lr1 and lr2, then we can build a ParamMap with both maxIter parameters specified: ParamMap(lr1.maxIter -> 10, lr2.maxIter -> 20). This is useful if there are two algorithms with the maxIter parameter in a Pipeline.

ML persistence: Saving and Loading Pipelines

Often times it is worth it to save a model or a pipeline to disk for later use. In Spark 1.6, a model import/export functionality was added to the Pipeline API. As of Spark 2.3, the DataFrame-based API in spark.ml and pyspark.ml has complete coverage.

ML persistence works across Scala, Java and Python. However, R currently uses a modified format, so models saved in R can only be loaded back in R; this should be fixed in the future and is tracked in SPARK-15572.

Backwards compatibility for ML persistence

In general, MLlib maintains backwards compatibility for ML persistence. I.e., if you save an ML model or Pipeline in one version of Spark, then you should be able to load it back and use it in a future version of Spark. However, there are rare exceptions, described below.

Model persistence: Is a model or Pipeline saved using Apache Spark ML persistence in Spark version X loadable by Spark version Y?

Model behavior: Does a model or Pipeline in Spark version X behave identically in Spark version Y?

For both model persistence and model behavior, any breaking changes across a minor version or patch version are reported in the Spark version release notes. If a breakage is not reported in release notes, then it should be treated as a bug to be fixed.

Code examples

This section gives code examples illustrating the functionality discussed above. For more info, please refer to the API documentation (Scala, Java, and Python).

Example: Estimator, Transformer, and Param

This example covers the concepts of Estimator, Transformer, and Param.

Refer to the Estimator Python docs, the Transformer Python docs and the Params Python docs for more details on the API.

from pyspark.ml.linalg import Vectors
from pyspark.ml.classification import LogisticRegression

# Prepare training data from a list of (label, features) tuples.
training = spark.createDataFrame([
    (1.0, Vectors.dense([0.0, 1.1, 0.1])),
    (0.0, Vectors.dense([2.0, 1.0, -1.0])),
    (0.0, Vectors.dense([2.0, 1.3, 1.0])),
    (1.0, Vectors.dense([0.0, 1.2, -0.5]))], ["label", "features"])

# Create a LogisticRegression instance. This instance is an Estimator.
lr = LogisticRegression(maxIter=10, regParam=0.01)
# Print out the parameters, documentation, and any default values.
print("LogisticRegression parameters:\n" + lr.explainParams() + "\n")

# Learn a LogisticRegression model. This uses the parameters stored in lr.
model1 = lr.fit(training)

# Since model1 is a Model (i.e., a transformer produced by an Estimator),
# we can view the parameters it used during fit().
# This prints the parameter (name: value) pairs, where names are unique IDs for this
# LogisticRegression instance.
print("Model 1 was fit using parameters: ")
print(model1.extractParamMap())

# We may alternatively specify parameters using a Python dictionary as a paramMap
paramMap = {lr.maxIter: 20}
paramMap[lr.maxIter] = 30  # Specify 1 Param, overwriting the original maxIter.
# Specify multiple Params.
paramMap.update({lr.regParam: 0.1, lr.threshold: 0.55})  # type: ignore

# You can combine paramMaps, which are python dictionaries.
# Change output column name
paramMap2 = {lr.probabilityCol: "myProbability"}
paramMapCombined = paramMap.copy()
paramMapCombined.update(paramMap2)  # type: ignore

# Now learn a new model using the paramMapCombined parameters.
# paramMapCombined overrides all parameters set earlier via lr.set* methods.
model2 = lr.fit(training, paramMapCombined)
print("Model 2 was fit using parameters: ")
print(model2.extractParamMap())

# Prepare test data
test = spark.createDataFrame([
    (1.0, Vectors.dense([-1.0, 1.5, 1.3])),
    (0.0, Vectors.dense([3.0, 2.0, -0.1])),
    (1.0, Vectors.dense([0.0, 2.2, -1.5]))], ["label", "features"])

# Make predictions on test data using the Transformer.transform() method.
# LogisticRegression.transform will only use the 'features' column.
# Note that model2.transform() outputs a "myProbability" column instead of the usual
# 'probability' column since we renamed the lr.probabilityCol parameter previously.
prediction = model2.transform(test)
result = prediction.select("features", "label", "myProbability", "prediction") \
    .collect()

for row in result:
    print("features=%s, label=%s -> prob=%s, prediction=%s"
          % (row.features, row.label, row.myProbability, row.prediction))
Find full example code at "examples/src/main/python/ml/estimator_transformer_param_example.py" in the Spark repo.

Refer to the Estimator Scala docs, the Transformer Scala docs and the Params Scala docs for details on the API.

import org.apache.spark.ml.classification.LogisticRegression
import org.apache.spark.ml.linalg.{Vector, Vectors}
import org.apache.spark.ml.param.ParamMap
import org.apache.spark.sql.Row

// Prepare training data from a list of (label, features) tuples.
val training = spark.createDataFrame(Seq(
  (1.0, Vectors.dense(0.0, 1.1, 0.1)),
  (0.0, Vectors.dense(2.0, 1.0, -1.0)),
  (0.0, Vectors.dense(2.0, 1.3, 1.0)),
  (1.0, Vectors.dense(0.0, 1.2, -0.5))
)).toDF("label", "features")

// Create a LogisticRegression instance. This instance is an Estimator.
val lr = new LogisticRegression()
// Print out the parameters, documentation, and any default values.
println(s"LogisticRegression parameters:\n ${lr.explainParams()}\n")

// We may set parameters using setter methods.
lr.setMaxIter(10)
  .setRegParam(0.01)

// Learn a LogisticRegression model. This uses the parameters stored in lr.
val model1 = lr.fit(training)
// Since model1 is a Model (i.e., a Transformer produced by an Estimator),
// we can view the parameters it used during fit().
// This prints the parameter (name: value) pairs, where names are unique IDs for this
// LogisticRegression instance.
println(s"Model 1 was fit using parameters: ${model1.parent.extractParamMap}")

// We may alternatively specify parameters using a ParamMap,
// which supports several methods for specifying parameters.
val paramMap = ParamMap(lr.maxIter -> 20)
  .put(lr.maxIter, 30)  // Specify 1 Param. This overwrites the original maxIter.
  .put(lr.regParam -> 0.1, lr.threshold -> 0.55)  // Specify multiple Params.

// One can also combine ParamMaps.
val paramMap2 = ParamMap(lr.probabilityCol -> "myProbability")  // Change output column name.
val paramMapCombined = paramMap ++ paramMap2

// Now learn a new model using the paramMapCombined parameters.
// paramMapCombined overrides all parameters set earlier via lr.set* methods.
val model2 = lr.fit(training, paramMapCombined)
println(s"Model 2 was fit using parameters: ${model2.parent.extractParamMap}")

// Prepare test data.
val test = spark.createDataFrame(Seq(
  (1.0, Vectors.dense(-1.0, 1.5, 1.3)),
  (0.0, Vectors.dense(3.0, 2.0, -0.1)),
  (1.0, Vectors.dense(0.0, 2.2, -1.5))
)).toDF("label", "features")

// Make predictions on test data using the Transformer.transform() method.
// LogisticRegression.transform will only use the 'features' column.
// Note that model2.transform() outputs a 'myProbability' column instead of the usual
// 'probability' column since we renamed the lr.probabilityCol parameter previously.
model2.transform(test)
  .select("features", "label", "myProbability", "prediction")
  .collect()
  .foreach { case Row(features: Vector, label: Double, prob: Vector, prediction: Double) =>
    println(s"($features, $label) -> prob=$prob, prediction=$prediction")
  }
Find full example code at "examples/src/main/scala/org/apache/spark/examples/ml/EstimatorTransformerParamExample.scala" in the Spark repo.

Refer to the Estimator Java docs, the Transformer Java docs and the Params Java docs for details on the API.

import java.util.Arrays;
import java.util.List;

import org.apache.spark.ml.classification.LogisticRegression;
import org.apache.spark.ml.classification.LogisticRegressionModel;
import org.apache.spark.ml.linalg.VectorUDT;
import org.apache.spark.ml.linalg.Vectors;
import org.apache.spark.ml.param.ParamMap;
import org.apache.spark.sql.Dataset;
import org.apache.spark.sql.Row;
import org.apache.spark.sql.RowFactory;
import org.apache.spark.sql.types.DataTypes;
import org.apache.spark.sql.types.Metadata;
import org.apache.spark.sql.types.StructField;
import org.apache.spark.sql.types.StructType;

// Prepare training data.
List<Row> dataTraining = Arrays.asList(
    RowFactory.create(1.0, Vectors.dense(0.0, 1.1, 0.1)),
    RowFactory.create(0.0, Vectors.dense(2.0, 1.0, -1.0)),
    RowFactory.create(0.0, Vectors.dense(2.0, 1.3, 1.0)),
    RowFactory.create(1.0, Vectors.dense(0.0, 1.2, -0.5))
);
StructType schema = new StructType(new StructField[]{
    new StructField("label", DataTypes.DoubleType, false, Metadata.empty()),
    new StructField("features", new VectorUDT(), false, Metadata.empty())
});
Dataset<Row> training = spark.createDataFrame(dataTraining, schema);

// Create a LogisticRegression instance. This instance is an Estimator.
LogisticRegression lr = new LogisticRegression();
// Print out the parameters, documentation, and any default values.
System.out.println("LogisticRegression parameters:\n" + lr.explainParams() + "\n");

// We may set parameters using setter methods.
lr.setMaxIter(10).setRegParam(0.01);

// Learn a LogisticRegression model. This uses the parameters stored in lr.
LogisticRegressionModel model1 = lr.fit(training);
// Since model1 is a Model (i.e., a Transformer produced by an Estimator),
// we can view the parameters it used during fit().
// This prints the parameter (name: value) pairs, where names are unique IDs for this
// LogisticRegression instance.
System.out.println("Model 1 was fit using parameters: " + model1.parent().extractParamMap());

// We may alternatively specify parameters using a ParamMap.
ParamMap paramMap = new ParamMap()
  .put(lr.maxIter().w(20))  // Specify 1 Param.
  .put(lr.maxIter(), 30)  // This overwrites the original maxIter.
  .put(lr.regParam().w(0.1), lr.threshold().w(0.55));  // Specify multiple Params.

// One can also combine ParamMaps.
ParamMap paramMap2 = new ParamMap()
  .put(lr.probabilityCol().w("myProbability"));  // Change output column name
ParamMap paramMapCombined = paramMap.$plus$plus(paramMap2);

// Now learn a new model using the paramMapCombined parameters.
// paramMapCombined overrides all parameters set earlier via lr.set* methods.
LogisticRegressionModel model2 = lr.fit(training, paramMapCombined);
System.out.println("Model 2 was fit using parameters: " + model2.parent().extractParamMap());

// Prepare test documents.
List<Row> dataTest = Arrays.asList(
    RowFactory.create(1.0, Vectors.dense(-1.0, 1.5, 1.3)),
    RowFactory.create(0.0, Vectors.dense(3.0, 2.0, -0.1)),
    RowFactory.create(1.0, Vectors.dense(0.0, 2.2, -1.5))
);
Dataset<Row> test = spark.createDataFrame(dataTest, schema);

// Make predictions on test documents using the Transformer.transform() method.
// LogisticRegression.transform will only use the 'features' column.
// Note that model2.transform() outputs a 'myProbability' column instead of the usual
// 'probability' column since we renamed the lr.probabilityCol parameter previously.
Dataset<Row> results = model2.transform(test);
Dataset<Row> rows = results.select("features", "label", "myProbability", "prediction");
for (Row r: rows.collectAsList()) {
  System.out.println("(" + r.get(0) + ", " + r.get(1) + ") -> prob=" + r.get(2)
    + ", prediction=" + r.get(3));
}
Find full example code at "examples/src/main/java/org/apache/spark/examples/ml/JavaEstimatorTransformerParamExample.java" in the Spark repo.

Example: Pipeline

This example follows the simple text document Pipeline illustrated in the figures above.

Refer to the Pipeline Python docs for more details on the API.

from pyspark.ml import Pipeline
from pyspark.ml.classification import LogisticRegression
from pyspark.ml.feature import HashingTF, Tokenizer

# Prepare training documents from a list of (id, text, label) tuples.
training = spark.createDataFrame([
    (0, "a b c d e spark", 1.0),
    (1, "b d", 0.0),
    (2, "spark f g h", 1.0),
    (3, "hadoop mapreduce", 0.0)
], ["id", "text", "label"])

# Configure an ML pipeline, which consists of three stages: tokenizer, hashingTF, and lr.
tokenizer = Tokenizer(inputCol="text", outputCol="words")
hashingTF = HashingTF(inputCol=tokenizer.getOutputCol(), outputCol="features")
lr = LogisticRegression(maxIter=10, regParam=0.001)
pipeline = Pipeline(stages=[tokenizer, hashingTF, lr])

# Fit the pipeline to training documents.
model = pipeline.fit(training)

# Prepare test documents, which are unlabeled (id, text) tuples.
test = spark.createDataFrame([
    (4, "spark i j k"),
    (5, "l m n"),
    (6, "spark hadoop spark"),
    (7, "apache hadoop")
], ["id", "text"])

# Make predictions on test documents and print columns of interest.
prediction = model.transform(test)
selected = prediction.select("id", "text", "probability", "prediction")
for row in selected.collect():
    rid, text, prob, prediction = row
    print(
        "(%d, %s) --> prob=%s, prediction=%f" % (
            rid, text, str(prob), prediction   # type: ignore
        )
    )
Find full example code at "examples/src/main/python/ml/pipeline_example.py" in the Spark repo.

Refer to the Pipeline Scala docs for details on the API.

import org.apache.spark.ml.{Pipeline, PipelineModel}
import org.apache.spark.ml.classification.LogisticRegression
import org.apache.spark.ml.feature.{HashingTF, Tokenizer}
import org.apache.spark.ml.linalg.Vector
import org.apache.spark.sql.Row

// Prepare training documents from a list of (id, text, label) tuples.
val training = spark.createDataFrame(Seq(
  (0L, "a b c d e spark", 1.0),
  (1L, "b d", 0.0),
  (2L, "spark f g h", 1.0),
  (3L, "hadoop mapreduce", 0.0)
)).toDF("id", "text", "label")

// Configure an ML pipeline, which consists of three stages: tokenizer, hashingTF, and lr.
val tokenizer = new Tokenizer()
  .setInputCol("text")
  .setOutputCol("words")
val hashingTF = new HashingTF()
  .setNumFeatures(1000)
  .setInputCol(tokenizer.getOutputCol)
  .setOutputCol("features")
val lr = new LogisticRegression()
  .setMaxIter(10)
  .setRegParam(0.001)
val pipeline = new Pipeline()
  .setStages(Array(tokenizer, hashingTF, lr))

// Fit the pipeline to training documents.
val model = pipeline.fit(training)

// Now we can optionally save the fitted pipeline to disk
model.write.overwrite().save("/tmp/spark-logistic-regression-model")

// We can also save this unfit pipeline to disk
pipeline.write.overwrite().save("/tmp/unfit-lr-model")

// And load it back in during production
val sameModel = PipelineModel.load("/tmp/spark-logistic-regression-model")

// Prepare test documents, which are unlabeled (id, text) tuples.
val test = spark.createDataFrame(Seq(
  (4L, "spark i j k"),
  (5L, "l m n"),
  (6L, "spark hadoop spark"),
  (7L, "apache hadoop")
)).toDF("id", "text")

// Make predictions on test documents.
model.transform(test)
  .select("id", "text", "probability", "prediction")
  .collect()
  .foreach { case Row(id: Long, text: String, prob: Vector, prediction: Double) =>
    println(s"($id, $text) --> prob=$prob, prediction=$prediction")
  }
Find full example code at "examples/src/main/scala/org/apache/spark/examples/ml/PipelineExample.scala" in the Spark repo.

Refer to the Pipeline Java docs for details on the API.

import java.util.Arrays;

import org.apache.spark.ml.Pipeline;
import org.apache.spark.ml.PipelineModel;
import org.apache.spark.ml.PipelineStage;
import org.apache.spark.ml.classification.LogisticRegression;
import org.apache.spark.ml.feature.HashingTF;
import org.apache.spark.ml.feature.Tokenizer;
import org.apache.spark.sql.Dataset;
import org.apache.spark.sql.Row;

// Prepare training documents, which are labeled.
Dataset<Row> training = spark.createDataFrame(Arrays.asList(
  new JavaLabeledDocument(0L, "a b c d e spark", 1.0),
  new JavaLabeledDocument(1L, "b d", 0.0),
  new JavaLabeledDocument(2L, "spark f g h", 1.0),
  new JavaLabeledDocument(3L, "hadoop mapreduce", 0.0)
), JavaLabeledDocument.class);

// Configure an ML pipeline, which consists of three stages: tokenizer, hashingTF, and lr.
Tokenizer tokenizer = new Tokenizer()
  .setInputCol("text")
  .setOutputCol("words");
HashingTF hashingTF = new HashingTF()
  .setNumFeatures(1000)
  .setInputCol(tokenizer.getOutputCol())
  .setOutputCol("features");
LogisticRegression lr = new LogisticRegression()
  .setMaxIter(10)
  .setRegParam(0.001);
Pipeline pipeline = new Pipeline()
  .setStages(new PipelineStage[] {tokenizer, hashingTF, lr});

// Fit the pipeline to training documents.
PipelineModel model = pipeline.fit(training);

// Prepare test documents, which are unlabeled.
Dataset<Row> test = spark.createDataFrame(Arrays.asList(
  new JavaDocument(4L, "spark i j k"),
  new JavaDocument(5L, "l m n"),
  new JavaDocument(6L, "spark hadoop spark"),
  new JavaDocument(7L, "apache hadoop")
), JavaDocument.class);

// Make predictions on test documents.
Dataset<Row> predictions = model.transform(test);
for (Row r : predictions.select("id", "text", "probability", "prediction").collectAsList()) {
  System.out.println("(" + r.get(0) + ", " + r.get(1) + ") --> prob=" + r.get(2)
    + ", prediction=" + r.get(3));
}
Find full example code at "examples/src/main/java/org/apache/spark/examples/ml/JavaPipelineExample.java" in the Spark repo.

Model selection (hyperparameter tuning)

A big benefit of using ML Pipelines is hyperparameter optimization. See the ML Tuning Guide for more information on automatic model selection.