Migration Guide: MLlib (Machine Learning)
- Upgrading from MLlib 2.4 to 3.0
- Upgrading from MLlib 2.2 to 2.3
- Upgrading from MLlib 2.1 to 2.2
- Upgrading from MLlib 2.0 to 2.1
- Upgrading from MLlib 1.6 to 2.0
- Upgrading from MLlib 1.5 to 1.6
- Upgrading from MLlib 1.4 to 1.5
- Upgrading from MLlib 1.3 to 1.4
- Upgrading from MLlib 1.2 to 1.3
- Upgrading from MLlib 1.1 to 1.2
- Upgrading from MLlib 1.0 to 1.1
- Upgrading from MLlib 0.9 to 1.0
Note that this migration guide describes the items specific to MLlib. Many items of SQL migration can be applied when migrating MLlib to higher versions for DataFrame-based APIs. Please refer Migration Guide: SQL, Datasets and DataFrame.
Upgrading from MLlib 2.4 to 3.0
Breaking changes
OneHotEncoder
which is deprecated in 2.3, is removed in 3.0 andOneHotEncoderEstimator
is now renamed toOneHotEncoder
.org.apache.spark.ml.image.ImageSchema.readImages
which is deprecated in 2.3, is removed in 3.0, usespark.read.format('image')
instead.org.apache.spark.mllib.clustering.KMeans.train
with param Intruns
which is deprecated in 2.1, is removed in 3.0. Usetrain
method withoutruns
instead.org.apache.spark.mllib.classification.LogisticRegressionWithSGD
which is deprecated in 2.0, is removed in 3.0, useorg.apache.spark.ml.classification.LogisticRegression
orspark.mllib.classification.LogisticRegressionWithLBFGS
instead.org.apache.spark.mllib.feature.ChiSqSelectorModel.isSorted
which is deprecated in 2.1, is removed in 3.0, is not intended for subclasses to use.org.apache.spark.mllib.regression.RidgeRegressionWithSGD
which is deprecated in 2.0, is removed in 3.0, useorg.apache.spark.ml.regression.LinearRegression
withelasticNetParam
= 0.0. Note the defaultregParam
is 0.01 forRidgeRegressionWithSGD
, but is 0.0 forLinearRegression
.org.apache.spark.mllib.regression.LassoWithSGD
which is deprecated in 2.0, is removed in 3.0, useorg.apache.spark.ml.regression.LinearRegression
withelasticNetParam
= 1.0. Note the defaultregParam
is 0.01 forLassoWithSGD
, but is 0.0 forLinearRegression
.org.apache.spark.mllib.regression.LinearRegressionWithSGD
which is deprecated in 2.0, is removed in 3.0, useorg.apache.spark.ml.regression.LinearRegression
orLBFGS
instead.org.apache.spark.mllib.clustering.KMeans.getRuns
andsetRuns
which are deprecated in 2.1, are removed in 3.0, have no effect since Spark 2.0.0.org.apache.spark.ml.LinearSVCModel.setWeightCol
which is deprecated in 2.4, is removed in 3.0, is not intended for users.- From 3.0,
org.apache.spark.ml.classification.MultilayerPerceptronClassificationModel
extendsMultilayerPerceptronParams
to expose the training params. As a result,layers
inMultilayerPerceptronClassificationModel
has been changed fromArray[Int]
toIntArrayParam
. Users should useMultilayerPerceptronClassificationModel.getLayers
instead ofMultilayerPerceptronClassificationModel.layers
to retrieve the size of layers. org.apache.spark.ml.classification.GBTClassifier.numTrees
which is deprecated in 2.4.5, is removed in 3.0, usegetNumTrees
instead.org.apache.spark.ml.clustering.KMeansModel.computeCost
which is deprecated in 2.4, is removed in 3.0, useClusteringEvaluator
instead.- The member variable
precision
inorg.apache.spark.mllib.evaluation.MulticlassMetrics
which is deprecated in 2.0, is removed in 3.0. Useaccuracy
instead. - The member variable
recall
inorg.apache.spark.mllib.evaluation.MulticlassMetrics
which is deprecated in 2.0, is removed in 3.0. Useaccuracy
instead. - The member variable
fMeasure
inorg.apache.spark.mllib.evaluation.MulticlassMetrics
which is deprecated in 2.0, is removed in 3.0. Useaccuracy
instead. org.apache.spark.ml.util.GeneralMLWriter.context
which is deprecated in 2.0, is removed in 3.0, usesession
instead.org.apache.spark.ml.util.MLWriter.context
which is deprecated in 2.0, is removed in 3.0, usesession
instead.org.apache.spark.ml.util.MLReader.context
which is deprecated in 2.0, is removed in 3.0, usesession
instead.abstract class UnaryTransformer[IN, OUT, T <: UnaryTransformer[IN, OUT, T]]
is changed toabstract class UnaryTransformer[IN: TypeTag, OUT: TypeTag, T <: UnaryTransformer[IN, OUT, T]]
in 3.0.
Deprecations and changes of behavior
Deprecations
- SPARK-11215:
labels
inStringIndexerModel
is deprecated and will be removed in 3.1.0. UselabelsArray
instead. - SPARK-25758:
computeCost
inBisectingKMeansModel
is deprecated and will be removed in future versions. UseClusteringEvaluator
instead.
Changes of behavior
- SPARK-11215:
In Spark 2.4 and previous versions, when specifying
frequencyDesc
orfrequencyAsc
asstringOrderType
param inStringIndexer
, in case of equal frequency, the order of strings is undefined. Since Spark 3.0, the strings with equal frequency are further sorted by alphabet. And since Spark 3.0,StringIndexer
supports encoding multiple columns. - SPARK-20604:
In prior to 3.0 releases,
Imputer
requires input column to be Double or Float. In 3.0, this restriction is lifted soImputer
can handle all numeric types. - SPARK-23469:
In Spark 3.0, the
HashingTF
Transformer uses a corrected implementation of the murmur3 hash function to hash elements to vectors.HashingTF
in Spark 3.0 will map elements to different positions in vectors than in Spark 2. However,HashingTF
created with Spark 2.x and loaded with Spark 3.0 will still use the previous hash function and will not change behavior. - SPARK-28969:
The
setClassifier
method in PySpark’sOneVsRestModel
has been removed in 3.0 for parity with the Scala implementation. Callers should not need to set the classifier in the model after creation. - SPARK-25790: PCA adds the support for more than 65535 column matrix in Spark 3.0.
- SPARK-28927: When fitting ALS model on nondeterministic input data, previously if rerun happens, users would see ArrayIndexOutOfBoundsException caused by mismatch between In/Out user/item blocks. From 3.0, a SparkException with more clear message will be thrown, and original ArrayIndexOutOfBoundsException is wrapped.
- SPARK-29232:
In prior to 3.0 releases,
RandomForestRegressionModel
doesn’t update the parameter maps of the DecisionTreeRegressionModels underneath. This is fixed in 3.0.
Upgrading from MLlib 2.2 to 2.3
Breaking changes
- The class and trait hierarchy for logistic regression model summaries was changed to be cleaner
and better accommodate the addition of the multi-class summary. This is a breaking change for user
code that casts a
LogisticRegressionTrainingSummary
to aBinaryLogisticRegressionTrainingSummary
. Users should instead use themodel.binarySummary
method. See SPARK-17139 for more detail (note this is anExperimental
API). This does not affect the Pythonsummary
method, which will still work correctly for both multinomial and binary cases.
Deprecations and changes of behavior
Deprecations
OneHotEncoder
has been deprecated and will be removed in3.0
. It has been replaced by the newOneHotEncoderEstimator
(see SPARK-13030). Note thatOneHotEncoderEstimator
will be renamed toOneHotEncoder
in3.0
(butOneHotEncoderEstimator
will be kept as an alias).
Changes of behavior
- SPARK-21027:
The default parallelism used in
OneVsRest
is now set to 1 (i.e. serial). In2.2
and earlier versions, the level of parallelism was set to the default threadpool size in Scala. - SPARK-22156:
The learning rate update for
Word2Vec
was incorrect whennumIterations
was set greater than1
. This will cause training results to be different between2.3
and earlier versions. - SPARK-21681: Fixed an edge case bug in multinomial logistic regression that resulted in incorrect coefficients when some features had zero variance.
- SPARK-16957: Tree algorithms now use mid-points for split values. This may change results from model training.
- SPARK-14657:
Fixed an issue where the features generated by
RFormula
without an intercept were inconsistent with the output in R. This may change results from model training in this scenario.
Upgrading from MLlib 2.1 to 2.2
Breaking changes
There are no breaking changes.
Deprecations and changes of behavior
Deprecations
There are no deprecations.
Changes of behavior
- SPARK-19787:
Default value of
regParam
changed from1.0
to0.1
forALS.train
method (markedDeveloperApi
). Note this does not affect theALS
Estimator or Model, nor MLlib’sALS
class. - SPARK-14772:
Fixed inconsistency between Python and Scala APIs for
Param.copy
method. - SPARK-11569:
StringIndexer
now handlesNULL
values in the same way as unseen values. Previously an exception would always be thrown regardless of the setting of thehandleInvalid
parameter.
Upgrading from MLlib 2.0 to 2.1
Breaking changes
Deprecated methods removed
setLabelCol
infeature.ChiSqSelectorModel
numTrees
inclassification.RandomForestClassificationModel
(This now refers to the Param callednumTrees
)numTrees
inregression.RandomForestRegressionModel
(This now refers to the Param callednumTrees
)model
inregression.LinearRegressionSummary
validateParams
inPipelineStage
validateParams
inEvaluator
Deprecations and changes of behavior
Deprecations
- SPARK-18592:
Deprecate all Param setter methods except for input/output column Params for
DecisionTreeClassificationModel
,GBTClassificationModel
,RandomForestClassificationModel
,DecisionTreeRegressionModel
,GBTRegressionModel
andRandomForestRegressionModel
Changes of behavior
- SPARK-17870:
Fix a bug of
ChiSqSelector
which will likely change its result. NowChiSquareSelector
use pValue rather than raw statistic to select a fixed number of top features. - SPARK-3261:
KMeans
returns potentially fewer than k cluster centers in cases where k distinct centroids aren’t available or aren’t selected. - SPARK-17389:
KMeans
reduces the default number of steps from 5 to 2 for the k-means|| initialization mode.
Upgrading from MLlib 1.6 to 2.0
Breaking changes
There were several breaking changes in Spark 2.0, which are outlined below.
Linear algebra classes for DataFrame-based APIs
Spark’s linear algebra dependencies were moved to a new project, mllib-local
(see SPARK-13944).
As part of this change, the linear algebra classes were copied to a new package, spark.ml.linalg
.
The DataFrame-based APIs in spark.ml
now depend on the spark.ml.linalg
classes,
leading to a few breaking changes, predominantly in various model classes
(see SPARK-14810 for a full list).
Note: the RDD-based APIs in spark.mllib
continue to depend on the previous package spark.mllib.linalg
.
Converting vectors and matrices
While most pipeline components support backward compatibility for loading,
some existing DataFrames
and pipelines in Spark versions prior to 2.0, that contain vector or matrix
columns, may need to be migrated to the new spark.ml
vector and matrix types.
Utilities for converting DataFrame
columns from spark.mllib.linalg
to spark.ml.linalg
types
(and vice versa) can be found in spark.mllib.util.MLUtils
.
There are also utility methods available for converting single instances of
vectors and matrices. Use the asML
method on a mllib.linalg.Vector
/ mllib.linalg.Matrix
for converting to ml.linalg
types, and
mllib.linalg.Vectors.fromML
/ mllib.linalg.Matrices.fromML
for converting to mllib.linalg
types.
Refer to the MLUtils
Scala docs for further detail.
Refer to the MLUtils
Java docs for further detail.
Refer to the MLUtils
Python docs for further detail.
Deprecated methods removed
Several deprecated methods were removed in the spark.mllib
and spark.ml
packages:
setScoreCol
inml.evaluation.BinaryClassificationEvaluator
weights
inLinearRegression
andLogisticRegression
inspark.ml
setMaxNumIterations
inmllib.optimization.LBFGS
(marked asDeveloperApi
)treeReduce
andtreeAggregate
inmllib.rdd.RDDFunctions
(these functions are available onRDD
s directly, and were marked asDeveloperApi
)defaultStategy
inmllib.tree.configuration.Strategy
build
inmllib.tree.Node
- libsvm loaders for multiclass and load/save labeledData methods in
mllib.util.MLUtils
A full list of breaking changes can be found at SPARK-14810.
Deprecations and changes of behavior
Deprecations
Deprecations in the spark.mllib
and spark.ml
packages include:
- SPARK-14984:
In
spark.ml.regression.LinearRegressionSummary
, themodel
field has been deprecated. - SPARK-13784:
In
spark.ml.regression.RandomForestRegressionModel
andspark.ml.classification.RandomForestClassificationModel
, thenumTrees
parameter has been deprecated in favor ofgetNumTrees
method. - SPARK-13761:
In
spark.ml.param.Params
, thevalidateParams
method has been deprecated. We move all functionality in overridden methods to the correspondingtransformSchema
. - SPARK-14829:
In
spark.mllib
package,LinearRegressionWithSGD
,LassoWithSGD
,RidgeRegressionWithSGD
andLogisticRegressionWithSGD
have been deprecated. We encourage users to usespark.ml.regression.LinearRegression
andspark.ml.classification.LogisticRegression
. - SPARK-14900:
In
spark.mllib.evaluation.MulticlassMetrics
, the parametersprecision
,recall
andfMeasure
have been deprecated in favor ofaccuracy
. - SPARK-15644:
In
spark.ml.util.MLReader
andspark.ml.util.MLWriter
, thecontext
method has been deprecated in favor ofsession
. - In
spark.ml.feature.ChiSqSelectorModel
, thesetLabelCol
method has been deprecated since it was not used byChiSqSelectorModel
.
Changes of behavior
Changes of behavior in the spark.mllib
and spark.ml
packages include:
- SPARK-7780:
spark.mllib.classification.LogisticRegressionWithLBFGS
directly callsspark.ml.classification.LogisticRegression
for binary classification now. This will introduce the following behavior changes forspark.mllib.classification.LogisticRegressionWithLBFGS
:- The intercept will not be regularized when training binary classification model with L1/L2 Updater.
- If users set without regularization, training with or without feature scaling will return the same solution by the same convergence rate.
- SPARK-13429:
In order to provide better and consistent result with
spark.ml.classification.LogisticRegression
, the default value ofspark.mllib.classification.LogisticRegressionWithLBFGS
:convergenceTol
has been changed from 1E-4 to 1E-6. - SPARK-12363:
Fix a bug of
PowerIterationClustering
which will likely change its result. - SPARK-13048:
LDA
using theEM
optimizer will keep the last checkpoint by default, if checkpointing is being used. - SPARK-12153:
Word2Vec
now respects sentence boundaries. Previously, it did not handle them correctly. - SPARK-10574:
HashingTF
usesMurmurHash3
as default hash algorithm in bothspark.ml
andspark.mllib
. - SPARK-14768:
The
expectedType
argument for PySparkParam
was removed. - SPARK-14931:
Some default
Param
values, which were mismatched between pipelines in Scala and Python, have been changed. - SPARK-13600:
QuantileDiscretizer
now usesspark.sql.DataFrameStatFunctions.approxQuantile
to find splits (previously used custom sampling logic). The output buckets will differ for same input data and params.
Upgrading from MLlib 1.5 to 1.6
There are no breaking API changes in the spark.mllib
or spark.ml
packages, but there are
deprecations and changes of behavior.
Deprecations:
- SPARK-11358:
In
spark.mllib.clustering.KMeans
, theruns
parameter has been deprecated. - SPARK-10592:
In
spark.ml.classification.LogisticRegressionModel
andspark.ml.regression.LinearRegressionModel
, theweights
field has been deprecated in favor of the new namecoefficients
. This helps disambiguate from instance (row) “weights” given to algorithms.
Changes of behavior:
- SPARK-7770:
spark.mllib.tree.GradientBoostedTrees
:validationTol
has changed semantics in 1.6. Previously, it was a threshold for absolute change in error. Now, it resembles the behavior ofGradientDescent
’sconvergenceTol
: For large errors, it uses relative error (relative to the previous error); for small errors (< 0.01
), it uses absolute error. - SPARK-11069:
spark.ml.feature.RegexTokenizer
: Previously, it did not convert strings to lowercase before tokenizing. Now, it converts to lowercase by default, with an option not to. This matches the behavior of the simplerTokenizer
transformer.
Upgrading from MLlib 1.4 to 1.5
In the spark.mllib
package, there are no breaking API changes but several behavior changes:
- SPARK-9005:
RegressionMetrics.explainedVariance
returns the average regression sum of squares. - SPARK-8600:
NaiveBayesModel.labels
become sorted. - SPARK-3382:
GradientDescent
has a default convergence tolerance1e-3
, and hence iterations might end earlier than 1.4.
In the spark.ml
package, there exists one breaking API change and one behavior change:
- SPARK-9268: Java’s varargs support is removed
from
Params.setDefault
due to a Scala compiler bug. - SPARK-10097:
Evaluator.isLargerBetter
is added to indicate metric ordering. Metrics like RMSE no longer flip signs as in 1.4.
Upgrading from MLlib 1.3 to 1.4
In the spark.mllib
package, there were several breaking changes, but all in DeveloperApi
or Experimental
APIs:
- Gradient-Boosted Trees
- (Breaking change) The signature of the
Loss.gradient
method was changed. This is only an issues for users who wrote their own losses for GBTs. - (Breaking change) The
apply
andcopy
methods for the case classBoostingStrategy
have been changed because of a modification to the case class fields. This could be an issue for users who useBoostingStrategy
to set GBT parameters.
- (Breaking change) The signature of the
- (Breaking change) The return value of
LDA.run
has changed. It now returns an abstract classLDAModel
instead of the concrete classDistributedLDAModel
. The object of typeLDAModel
can still be cast to the appropriate concrete type, which depends on the optimization algorithm.
In the spark.ml
package, several major API changes occurred, including:
Param
and other APIs for specifying parametersuid
unique IDs for Pipeline components- Reorganization of certain classes
Since the spark.ml
API was an alpha component in Spark 1.3, we do not list all changes here.
However, since 1.4 spark.ml
is no longer an alpha component, we will provide details on any API
changes for future releases.
Upgrading from MLlib 1.2 to 1.3
In the spark.mllib
package, there were several breaking changes. The first change (in ALS
) is the only one in a component not marked as Alpha or Experimental.
- (Breaking change) In
ALS
, the extraneous methodsolveLeastSquares
has been removed. TheDeveloperApi
methodanalyzeBlocks
was also removed. - (Breaking change)
StandardScalerModel
remains an Alpha component. In it, thevariance
method has been replaced with thestd
method. To compute the column variance values returned by the originalvariance
method, simply square the standard deviation values returned bystd
. - (Breaking change)
StreamingLinearRegressionWithSGD
remains an Experimental component. In it, there were two changes:- The constructor taking arguments was removed in favor of a builder pattern using the default constructor plus parameter setter methods.
- Variable
model
is no longer public.
- (Breaking change)
DecisionTree
remains an Experimental component. In it and its associated classes, there were several changes:- In
DecisionTree
, the deprecated class methodtrain
has been removed. (The object/statictrain
methods remain.) - In
Strategy
, thecheckpointDir
parameter has been removed. Checkpointing is still supported, but the checkpoint directory must be set before calling tree and tree ensemble training.
- In
PythonMLlibAPI
(the interface between Scala/Java and Python for MLlib) was a public API but is now private, declaredprivate[python]
. This was never meant for external use.- In linear regression (including Lasso and ridge regression), the squared loss is now divided by 2. So in order to produce the same result as in 1.2, the regularization parameter needs to be divided by 2 and the step size needs to be multiplied by 2.
In the spark.ml
package, the main API changes are from Spark SQL. We list the most important changes here:
- The old SchemaRDD has been replaced with DataFrame with a somewhat modified API. All algorithms in
spark.ml
which used to use SchemaRDD now use DataFrame. - In Spark 1.2, we used implicit conversions from
RDD
s ofLabeledPoint
intoSchemaRDD
s by callingimport sqlContext._
wheresqlContext
was an instance ofSQLContext
. These implicits have been moved, so we now callimport sqlContext.implicits._
. - Java APIs for SQL have also changed accordingly. Please see the examples above and the Spark SQL Programming Guide for details.
Other changes were in LogisticRegression
:
- The
scoreCol
output column (with default value “score”) was renamed to beprobabilityCol
(with default value “probability”). The type was originallyDouble
(for the probability of class 1.0), but it is nowVector
(for the probability of each class, to support multiclass classification in the future). - In Spark 1.2,
LogisticRegressionModel
did not include an intercept. In Spark 1.3, it includes an intercept; however, it will always be 0.0 since it uses the default settings for spark.mllib.LogisticRegressionWithLBFGS. The option to use an intercept will be added in the future.
Upgrading from MLlib 1.1 to 1.2
The only API changes in MLlib v1.2 are in
DecisionTree
,
which continues to be an experimental API in MLlib 1.2:
-
(Breaking change) The Scala API for classification takes a named argument specifying the number of classes. In MLlib v1.1, this argument was called
numClasses
in Python andnumClassesForClassification
in Scala. In MLlib v1.2, the names are both set tonumClasses
. ThisnumClasses
parameter is specified either viaStrategy
or viaDecisionTree
statictrainClassifier
andtrainRegressor
methods. -
(Breaking change) The API for
Node
has changed. This should generally not affect user code, unless the user manually constructs decision trees (instead of using thetrainClassifier
ortrainRegressor
methods). The treeNode
now includes more information, including the probability of the predicted label (for classification). -
Printing methods’ output has changed. The
toString
(Scala/Java) and__repr__
(Python) methods used to print the full model; they now print a summary. For the full model, usetoDebugString
.
Examples in the Spark distribution and examples in the Decision Trees Guide have been updated accordingly.
Upgrading from MLlib 1.0 to 1.1
The only API changes in MLlib v1.1 are in
DecisionTree
,
which continues to be an experimental API in MLlib 1.1:
-
(Breaking change) The meaning of tree depth has been changed by 1 in order to match the implementations of trees in scikit-learn and in rpart. In MLlib v1.0, a depth-1 tree had 1 leaf node, and a depth-2 tree had 1 root node and 2 leaf nodes. In MLlib v1.1, a depth-0 tree has 1 leaf node, and a depth-1 tree has 1 root node and 2 leaf nodes. This depth is specified by the
maxDepth
parameter inStrategy
or viaDecisionTree
statictrainClassifier
andtrainRegressor
methods. -
(Non-breaking change) We recommend using the newly added
trainClassifier
andtrainRegressor
methods to build aDecisionTree
, rather than using the old parameter classStrategy
. These new training methods explicitly separate classification and regression, and they replace specialized parameter types with simpleString
types.
Examples of the new recommended trainClassifier
and trainRegressor
are given in the
Decision Trees Guide.
Upgrading from MLlib 0.9 to 1.0
In MLlib v1.0, we support both dense and sparse input in a unified way, which introduces a few breaking changes. If your data is sparse, please store it in a sparse format instead of dense to take advantage of sparsity in both storage and computation. Details are described below.