Github user sethah commented on a diff in the pull request:
https://github.com/apache/spark/pull/18305#discussion_r128041731
--- Diff:
mllib/src/test/scala/org/apache/spark/ml/optim/aggregator/LogisticAggregatorSuite.scala
---
@@ -0,0 +1,254 @@
+/*
+ * 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.optim.aggregator
+
+import org.apache.spark.SparkFunSuite
+import org.apache.spark.ml.feature.Instance
+import org.apache.spark.ml.linalg.{BLAS, Matrices, Vector, Vectors}
+import org.apache.spark.ml.util.TestingUtils._
+import org.apache.spark.mllib.util.MLlibTestSparkContext
+
+class LogisticAggregatorSuite extends SparkFunSuite with
MLlibTestSparkContext {
+
+ import DifferentiableLossAggregatorSuite.getClassificationSummarizers
+
+ @transient var instances: Array[Instance] = _
+ @transient var instancesConstantFeature: Array[Instance] = _
+
+ override def beforeAll(): Unit = {
+ super.beforeAll()
+ instances = Array(
+ Instance(0.0, 0.1, Vectors.dense(1.0, 2.0)),
+ Instance(1.0, 0.5, Vectors.dense(1.5, 1.0)),
+ Instance(2.0, 0.3, Vectors.dense(4.0, 0.5))
+ )
+ instancesConstantFeature = Array(
+ Instance(0.0, 0.1, Vectors.dense(1.0, 2.0)),
+ Instance(1.0, 0.5, Vectors.dense(1.0, 1.0)),
+ Instance(2.0, 0.3, Vectors.dense(1.0, 0.5))
+ )
+ }
+
+
+ /** Get summary statistics for some data and create a new
LogisticAggregator. */
+ private def getNewAggregator(
+ instances: Array[Instance],
+ coefficients: Vector,
+ fitIntercept: Boolean,
+ isMultinomial: Boolean): LogisticAggregator = {
+ val (featuresSummarizer, ySummarizer) =
+
DifferentiableLossAggregatorSuite.getClassificationSummarizers(instances)
+ val numClasses = ySummarizer.histogram.length
+ val featuresStd = featuresSummarizer.variance.toArray.map(math.sqrt)
+ val bcFeaturesStd = spark.sparkContext.broadcast(featuresStd)
+ val bcCoefficients = spark.sparkContext.broadcast(coefficients)
+ new LogisticAggregator(bcFeaturesStd, numClasses, fitIntercept,
isMultinomial)(bcCoefficients)
+ }
+
+ test("aggregator add method input size") {
+ val coefArray = Array(1.0, 2.0, -2.0, 3.0, 0.0, -1.0)
+ val interceptArray = Array(4.0, 2.0, -3.0)
+ val agg = getNewAggregator(instances, Vectors.dense(coefArray ++
interceptArray),
+ fitIntercept = true, isMultinomial = true)
+ withClue("LogisticAggregator features dimension must match
coefficients dimension") {
+ intercept[IllegalArgumentException] {
+ agg.add(Instance(1.0, 1.0, Vectors.dense(2.0)))
+ }
+ }
+ }
+
+ test("negative weight") {
+ val coefArray = Array(1.0, 2.0, -2.0, 3.0, 0.0, -1.0)
+ val interceptArray = Array(4.0, 2.0, -3.0)
+ val agg = getNewAggregator(instances, Vectors.dense(coefArray ++
interceptArray),
+ fitIntercept = true, isMultinomial = true)
+ withClue("LogisticAggregator does not support negative instance
weights") {
+ intercept[IllegalArgumentException] {
+ agg.add(Instance(1.0, -1.0, Vectors.dense(2.0, 1.0)))
+ }
+ }
+ }
+
+ test("check sizes multinomial") {
+ val rng = new scala.util.Random
+ val numFeatures = instances.head.features.size
+ val numClasses = instances.map(_.label).toSet.size
+ val coefWithIntercept = Vectors.dense(
+ Array.fill(numClasses * (numFeatures + 1))(rng.nextDouble))
+ val coefWithoutIntercept = Vectors.dense(
+ Array.fill(numClasses * numFeatures)(rng.nextDouble))
+ val aggIntercept = getNewAggregator(instances, coefWithIntercept,
fitIntercept = true,
+ isMultinomial = true)
+ val aggNoIntercept = getNewAggregator(instances, coefWithoutIntercept,
fitIntercept = false,
+ isMultinomial = true)
+ instances.foreach(aggIntercept.add)
+ instances.foreach(aggNoIntercept.add)
+
+ assert(aggIntercept.gradient.size === (numFeatures + 1) * numClasses)
+ assert(aggNoIntercept.gradient.size === numFeatures * numClasses)
+ }
+
+ test("check sizes binomial") {
+ val rng = new scala.util.Random
+ val binaryInstances = instances.filter(_.label < 2.0)
+ val numFeatures = binaryInstances.head.features.size
+ val coefWithIntercept = Vectors.dense(Array.fill(numFeatures +
1)(rng.nextDouble))
+ val coefWithoutIntercept =
Vectors.dense(Array.fill(numFeatures)(rng.nextDouble))
+ val aggIntercept = getNewAggregator(binaryInstances,
coefWithIntercept, fitIntercept = true,
+ isMultinomial = false)
+ val aggNoIntercept = getNewAggregator(binaryInstances,
coefWithoutIntercept,
+ fitIntercept = false, isMultinomial = false)
+ binaryInstances.foreach(aggIntercept.add)
+ binaryInstances.foreach(aggNoIntercept.add)
+
+ assert(aggIntercept.gradient.size === numFeatures + 1)
+ assert(aggNoIntercept.gradient.size === numFeatures)
+ }
+
+ test("check correctness multinomial") {
+ /*
+ Check that the aggregator computes loss/gradient for:
+ -sum_i w_i * (beta_y dot x_i - log(sum_k e^(beta_k dot x_i)))
+ */
+ val coefArray = Array(1.0, 2.0, -2.0, 3.0, 0.0, -1.0)
+ val interceptArray = Array(4.0, 2.0, -3.0)
+ val numFeatures = instances.head.features.size
+ val numClasses = instances.map(_.label).toSet.size
+ val intercepts = Vectors.dense(interceptArray)
+ val (featuresSummarizer, ySummarizer) =
getClassificationSummarizers(instances)
+ val featuresStd = featuresSummarizer.variance.toArray.map(math.sqrt)
+ val weightSum = instances.map(_.weight).sum
+
+ val agg = getNewAggregator(instances, Vectors.dense(coefArray ++
interceptArray),
+ fitIntercept = true, isMultinomial = true)
+ instances.foreach(agg.add)
+
+ // compute the loss
+ val stdCoef = coefArray.indices.map(i => coefArray(i) / featuresStd(i
/ numClasses)).toArray
--- End diff --
What users interpret `standardization` to mean may be different than what
it means in Spark, sure. I wasn't around when that design choice was made.
Standardization is **always** done internally because it improves
convergence. When no regularization is applied, you can verify that the two
scenarios:
* training on scaled features and then converting the coefficients to the
unscaled space
* training on un-scaled features
produce _the same coefficients_. So, instead of literally not standardizing
the features when standardization is false, we just do it anyway because it's
better for convergence, and it doesn't affect the results.
However, when regularization _is_ applied, the results are not the same.
The l2 loss, for example, is `sum beta_j^2`. But the coefficients that we are
using during training are scaled coefficients (since we always scale), which
are effectively `beta_{j, scaled} = beta_j * sigma_j`. We need to
"unstandardize" the coefficients for the regularization part of the loss
because their scales do matter. If we didn't unstandardize them, we'd change
the regularization loss to `sum (beta_j * sigma_j)^2`, which would not be
correct if the user wished not to train in the scaled space.
Do I think this is all a bit confusing, especially for users? Yes. I
believe this was done to match glmnet. I think it's fine the way it is, but
that discussion is for another JIRA anyway.
---
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]
