[
https://issues.apache.org/jira/browse/SPARK-5016?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14312269#comment-14312269
]
Travis Galoppo commented on SPARK-5016:
---------------------------------------
The k Gaussians are updated with code that right now looks like
{code}
var i = 0
while (i < k) {
val mu = sums.means(i) / sums.weights(i)
BLAS.syr(-sums.weights(i),
Vectors.fromBreeze(mu).asInstanceOf[DenseVector],
Matrices.fromBreeze(sums.sigmas(i)).asInstanceOf[DenseMatrix])
weights(i) = sums.weights(i) / sumWeights
gaussians(i) = new MultivariateGaussian(mu, sums.sigmas(i) /
sums.weights(i))
i = i + 1
}
{code}
... the matrix inversion (or, in reality, partial inversion... the inverse is
not explicitly calculated) occurs during the creation of the
MultivariateGaussian objects... this code could be parallelized something like:
{code}
val (ws, gs) = sc.parallelize(0 until k).map{ i =>
val mu = sums.means(i) / sums.weights(i)
BLAS.syr(-sums.weights(i),
Vectors.fromBreeze(mu).asInstanceOf[DenseVector],
Matrices.fromBreeze(sums.sigmas(i)).asInstanceOf[DenseMatrix])
val weight = sums.weights(i) / sumWeights
val gaussian = new MultivariateGaussian(mu, sums.sigmas(i) /
sums.weights(i))
(weight, gaussian)
}.collect.unzip
(0 until k).foreach{ i =>
weights(i) = ws(i)
gaussians(i) = gs(i)
}
{code}
... effectively distributing the computation of the k MutlivariateGaussians
(and their weights).
As for the threshold values for k / numFeatures... this is probably a function
of cluster size and interconnect speed. These thresholds should probably be
optional parameters to GaussianMixture. Personally, I would vote for the
default behavior to not perform this parallelization, and let the user decide
when the time is right to allow it.
> GaussianMixtureEM should distribute matrix inverse for large numFeatures, k
> ---------------------------------------------------------------------------
>
> Key: SPARK-5016
> URL: https://issues.apache.org/jira/browse/SPARK-5016
> Project: Spark
> Issue Type: Improvement
> Components: MLlib
> Affects Versions: 1.2.0
> Reporter: Joseph K. Bradley
>
> If numFeatures or k are large, GMM EM should distribute the matrix inverse
> computation for Gaussian initialization.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
