Github user srowen commented on a diff in the pull request:
https://github.com/apache/spark/pull/23126#discussion_r237520303
--- Diff:
mllib/src/main/scala/org/apache/spark/mllib/linalg/distributed/RowMatrix.scala
---
@@ -128,6 +128,82 @@ class RowMatrix @Since("1.0.0") (
RowMatrix.triuToFull(n, GU.data)
}
+ private def computeDenseVectorCovariance(mean: Vector, n: Int, m: Long):
Matrix = {
+
+ val bc = rows.context.broadcast(mean)
+
+ // Computes n*(n+1)/2, avoiding overflow in the multiplication.
+ // This succeeds when n <= 65535, which is checked above
+ val nt = if (n % 2 == 0) ((n / 2) * (n + 1)) else (n * ((n + 1) / 2))
+
+ val MU = rows.treeAggregate(new BDV[Double](nt))(
+ seqOp = (U, v) => {
+
+ val n = v.size
+ val na = Array.ofDim[Double](n)
+ val means = bc.value
+ if (v.isInstanceOf[DenseVector]) {
+ v.foreachActive{(index, value) =>
--- End diff --
Nit (this might fail scalastyle): space around the braces here. But we
don't need foreachActive here I think; they're all 'active' in a dense vector
and every element needs the mean subtracted. Even for a sparse vector you have
to do this. Do you really need a separate case here? we're assuming the vectors
are (mostly) dense in this method.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
