Java heap space

[iliasku]

Java heap space

[KevinJMao]

I think this might be due to the size of the covariance matrix defined in MatrixHelpers.scala.

Assuming the bot trained on 30 images that are 200 x 200, the covariance matrix seems like it'd end up being a [40000 x 40000] matrix, which rounds out to just under 12GB of memory.

Turk and Pentland suggested in their original paper a shortcut based on the fact that only a handful of these eigenvectors are useful. Here's an implementation (I haven't had a chance to test/profile due to FB authentication giving me the finger.):

def computeEigenVectors(diffMatrix: Array[Array[Double]]): DoubleMatrix2D = {
  val factory = DoubleFactory2D.dense
  val diffDoubleMatrix = factory.make(diffMatrix.length, diffMatrix(0).length).assign(diffMatrix)

  val shortDoubleMatrix = diffDoubleMatrix.zMult(diffDoubleMatrix, null, 1.0, 0.0, false, true)
  val eigenVectorMatrix = new EigenvalueDecomposition(shortDoubleMatrix).getV
  diffDoubleMatrix.zMult(eigenVectorMatrix, null)
}

[aroyer]

might be related #74

[AlecZadikian9001]

I'm still having this problem. Tried in OS X and Ubuntu Server, both running Java 1.8.