Skip to content

[SPARK-2163] class LBFGS optimize with Double tolerance instead of Int #1104

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
BaiGang wants to merge 1 commit into from
Closed

[SPARK-2163] class LBFGS optimize with Double tolerance instead of Int #1104

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,7 @@ class LBFGS(private var gradient: Gradient, private var updater: Updater)
* Set the convergence tolerance of iterations for L-BFGS. Default 1E-4.
* Smaller value will lead to higher accuracy with the cost of more iterations.
*/
def setConvergenceTol(tolerance: Int): this.type = {
def setConvergenceTol(tolerance: Double): this.type = {
this.convergenceTol = tolerance
this
}
Expand Down
34 changes: 34 additions & 0 deletions mllib/src/test/scala/org/apache/spark/mllib/optimization/LBFGSSuite.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -195,4 +195,38 @@ class LBFGSSuite extends FunSuite with LocalSparkContext with Matchers {
assert(lossLBFGS3.length == 6)
assert((lossLBFGS3(4) - lossLBFGS3(5)) / lossLBFGS3(4) < convergenceTol)
}

Copy link
Member

@dbtsai dbtsai Jun 18, 2014

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The bug isn't found because we only test the static runLBFGS method instead of the class. We probably can change all the existing tests to use the one in class, so we don't need to add another test.

@mengxr what do you think?

Copy link
Contributor

@mengxr mengxr Jun 18, 2014

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry I didn't see this error either ... Let's keep both in tests for better test coverage.

Copy link
Member

@dbtsai dbtsai Jun 18, 2014

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We may add the same test to SGD as well. My bad. Our internal one is right. Probably when I copy and paste, I don't do thing right.

test("Optimize via class LBFGS.") {
val regParam = 0.2

// Prepare another non-zero weights to compare the loss in the first iteration.
val initialWeightsWithIntercept = Vectors.dense(0.3, 0.12)
val convergenceTol = 1e-12
val maxNumIterations = 10

val lbfgsOptimizer = new LBFGS(gradient, squaredL2Updater)
.setNumCorrections(numCorrections)
.setConvergenceTol(convergenceTol)
.setMaxNumIterations(maxNumIterations)
.setRegParam(regParam)

val weightLBFGS = lbfgsOptimizer.optimize(dataRDD, initialWeightsWithIntercept)

val numGDIterations = 50
val stepSize = 1.0
val (weightGD, _) = GradientDescent.runMiniBatchSGD(
dataRDD,
gradient,
squaredL2Updater,
stepSize,
numGDIterations,
regParam,
miniBatchFrac,
initialWeightsWithIntercept)

// for class LBFGS and the optimize method, we only look at the weights
assert(compareDouble(weightLBFGS(0), weightGD(0), 0.02) &&
compareDouble(weightLBFGS(1), weightGD(1), 0.02),
"The weight differences between LBFGS and GD should be within 2%.")
}
}