Github user dbtsai commented on a diff in the pull request:
https://github.com/apache/spark/pull/13796#discussion_r75407380
--- Diff:
mllib/src/main/scala/org/apache/spark/ml/classification/MultinomialLogisticRegression.scala
---
@@ -0,0 +1,611 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.ml.classification
+
+import scala.collection.mutable
+
+import breeze.linalg.{DenseVector => BDV}
+import breeze.optimize.{CachedDiffFunction, LBFGS => BreezeLBFGS, OWLQN =>
BreezeOWLQN}
+import org.apache.hadoop.fs.Path
+
+import org.apache.spark.SparkException
+import org.apache.spark.annotation.{Experimental, Since}
+import org.apache.spark.internal.Logging
+import org.apache.spark.ml.feature.Instance
+import org.apache.spark.ml.linalg._
+import org.apache.spark.ml.param._
+import org.apache.spark.ml.param.shared._
+import org.apache.spark.ml.util._
+import org.apache.spark.mllib.linalg.VectorImplicits._
+import org.apache.spark.mllib.stat.MultivariateOnlineSummarizer
+import org.apache.spark.rdd.RDD
+import org.apache.spark.sql.{Dataset, Row}
+import org.apache.spark.sql.functions.{col, lit}
+import org.apache.spark.sql.types.DoubleType
+import org.apache.spark.storage.StorageLevel
+
+/**
+ * Params for multinomial logistic (softmax) regression.
+ */
+private[classification] trait MultinomialLogisticRegressionParams
+ extends ProbabilisticClassifierParams with HasRegParam with
HasElasticNetParam with HasMaxIter
+ with HasFitIntercept with HasTol with HasStandardization with
HasWeightCol {
+
+ /**
+ * Set thresholds in multiclass (or binary) classification to adjust the
probability of
+ * predicting each class. Array must have length equal to the number of
classes, with values >= 0.
+ * The class with largest value p/t is predicted, where p is the
original probability of that
+ * class and t is the class' threshold.
+ *
+ * @group setParam
+ */
+ def setThresholds(value: Array[Double]): this.type = {
+ set(thresholds, value)
+ }
+
+ /**
+ * Get thresholds for binary or multiclass classification.
+ *
+ * @group getParam
+ */
+ override def getThresholds: Array[Double] = {
+ $(thresholds)
+ }
+}
+
+/**
+ * :: Experimental ::
+ * Multinomial Logistic (softmax) regression.
+ */
+@Since("2.1.0")
+@Experimental
+class MultinomialLogisticRegression @Since("2.1.0") (
+ @Since("2.1.0") override val uid: String)
+ extends ProbabilisticClassifier[Vector,
+ MultinomialLogisticRegression, MultinomialLogisticRegressionModel]
+ with MultinomialLogisticRegressionParams with DefaultParamsWritable
with Logging {
+
+ @Since("2.1.0")
+ def this() = this(Identifiable.randomUID("mlogreg"))
+
+ /**
+ * Set the regularization parameter.
+ * Default is 0.0.
+ *
+ * @group setParam
+ */
+ @Since("2.1.0")
+ def setRegParam(value: Double): this.type = set(regParam, value)
+ setDefault(regParam -> 0.0)
+
+ /**
+ * Set the ElasticNet mixing parameter.
+ * For alpha = 0, the penalty is an L2 penalty. For alpha = 1, it is an
L1 penalty.
+ * For 0 < alpha < 1, the penalty is a combination of L1 and L2.
+ * Default is 0.0 which is an L2 penalty.
+ *
+ * @group setParam
+ */
+ @Since("2.1.0")
+ def setElasticNetParam(value: Double): this.type = set(elasticNetParam,
value)
+ setDefault(elasticNetParam -> 0.0)
+
+ /**
+ * Set the maximum number of iterations.
+ * Default is 100.
+ *
+ * @group setParam
+ */
+ @Since("2.1.0")
+ def setMaxIter(value: Int): this.type = set(maxIter, value)
+ setDefault(maxIter -> 100)
+
+ /**
+ * Set the convergence tolerance of iterations.
+ * Smaller value will lead to higher accuracy with the cost of more
iterations.
+ * Default is 1E-6.
+ *
+ * @group setParam
+ */
+ @Since("2.1.0")
+ def setTol(value: Double): this.type = set(tol, value)
+ setDefault(tol -> 1E-6)
+
+ /**
+ * Whether to fit an intercept term.
+ * Default is true.
+ *
+ * @group setParam
+ */
+ @Since("2.1.0")
+ def setFitIntercept(value: Boolean): this.type = set(fitIntercept, value)
+ setDefault(fitIntercept -> true)
+
+ /**
+ * Whether to standardize the training features before fitting the model.
+ * The coefficients of models will be always returned on the original
scale,
+ * so it will be transparent for users. Note that with/without
standardization,
+ * the models should always converge to the same solution when no
regularization
+ * is applied. In R's GLMNET package, the default behavior is true as
well.
+ * Default is true.
+ *
+ * @group setParam
+ */
+ @Since("2.1.0")
+ def setStandardization(value: Boolean): this.type = set(standardization,
value)
+ setDefault(standardization -> true)
+
+ /**
+ * Sets the value of param [[weightCol]].
+ * If this is not set or empty, we treat all instance weights as 1.0.
+ * Default is not set, so all instances have weight one.
+ *
+ * @group setParam
+ */
+ @Since("2.1.0")
+ def setWeightCol(value: String): this.type = set(weightCol, value)
+
+ @Since("2.1.0")
+ override def setThresholds(value: Array[Double]): this.type =
super.setThresholds(value)
+
+ override protected[spark] def train(dataset: Dataset[_]):
MultinomialLogisticRegressionModel = {
+ val w = if (!isDefined(weightCol) || $(weightCol).isEmpty) lit(1.0)
else col($(weightCol))
+ val instances: RDD[Instance] =
+ dataset.select(col($(labelCol)).cast(DoubleType), w,
col($(featuresCol))).rdd.map {
+ case Row(label: Double, weight: Double, features: Vector) =>
+ Instance(label, weight, features)
+ }
+
+ val handlePersistence = dataset.rdd.getStorageLevel ==
StorageLevel.NONE
+ if (handlePersistence) instances.persist(StorageLevel.MEMORY_AND_DISK)
+
+ val instr = Instrumentation.create(this, instances)
+ instr.logParams(regParam, elasticNetParam, standardization, thresholds,
+ maxIter, tol, fitIntercept)
+
+ val (summarizer, labelSummarizer) = {
+ val seqOp = (c: (MultivariateOnlineSummarizer, MultiClassSummarizer),
+ instance: Instance) =>
+ (c._1.add(instance.features, instance.weight),
c._2.add(instance.label, instance.weight))
+
+ val combOp = (c1: (MultivariateOnlineSummarizer,
MultiClassSummarizer),
+ c2: (MultivariateOnlineSummarizer, MultiClassSummarizer)) =>
+ (c1._1.merge(c2._1), c1._2.merge(c2._2))
+
+ instances.treeAggregate(
+ new MultivariateOnlineSummarizer, new MultiClassSummarizer)(seqOp,
combOp)
+ }
+
+ val histogram = labelSummarizer.histogram
+ val numInvalid = labelSummarizer.countInvalid
+ val numFeatures = summarizer.mean.size
+ val numFeaturesPlusIntercept = if (getFitIntercept) numFeatures + 1
else numFeatures
+
+ val numClasses =
MetadataUtils.getNumClasses(dataset.schema($(labelCol))) match {
+ case Some(n: Int) =>
+ require(n >= histogram.length, s"Specified number of classes $n
was " +
+ s"less than the number of unique labels ${histogram.length}")
+ n
+ case None => histogram.length
+ }
--- End diff --
This comes down to how we handle metadata I think. The problem here could
be that a training data only has label 3 and label 4, but in the current
design, we treat it as 5 classes problem resulting a coefficient matrix with
dimensions `5 * featureSizePlusIntercept`. This will not only introduce
negative infinity as intercepts in label 0, 1, and 2, but also we need to do
smoothing in computing the initial coefficients. This will make the entire
optimization less stable as well since classes 0, 1, and 2 are no-seen with
negative infinity intercepts.
Now I strongly think we either have a better metadata handling to map each
class from 0 to K -1 such that all the labels have its own samples, or just
fail the training like R. Users need to make sure that all the labels from 0 to
K - 1 have at least one sample, and they have the responsibility of making sure
all the labels in testing or validation should be in training.
However, for cross validation, there is chance that training data doesn't
contain certain label when sampling is introduced. Thus, I think the first
approach can be more desirable, which knows the real numClasses from metadata,
and only trains for those labels in the training resulting smaller coefficient
matrix. In prediction, we just put probability as zeros for those label not in
training. Thus, we don't need to do smoothing and the training will be more
stable without having negative infinity as intercepts.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
