Golub-Kahan SVD algorithm for KMP tensors #499
@ -8,7 +8,7 @@ import kotlinx.benchmark.Benchmark
|
||||
import kotlinx.benchmark.Blackhole
|
||||
import kotlinx.benchmark.Scope
|
||||
import kotlinx.benchmark.State
|
||||
import space.kscience.kmath.tensors.core.BroadcastDoubleTensorAlgebra.svd
|
||||
import space.kscience.kmath.tensors.core.BroadcastDoubleTensorAlgebra.svdGolubKahan
|
||||
import space.kscience.kmath.tensors.core.DoubleTensorAlgebra
|
||||
import space.kscience.kmath.tensors.core.DoubleTensorAlgebra.Companion.svdPowerMethod
|
||||
|
||||
@ -26,9 +26,9 @@ class SVDBenchmark {
|
||||
}
|
||||
|
||||
@Benchmark
|
||||
fun svdGolabKahan(blackhole: Blackhole) {
|
||||
fun svdGolubKahan(blackhole: Blackhole) {
|
||||
blackhole.consume(
|
||||
tensor.svd()
|
||||
tensor.svdGolubKahan()
|
||||
)
|
||||
}
|
||||
}
|
@ -834,6 +834,10 @@ public open class DoubleTensorAlgebra :
|
||||
}
|
||||
|
||||
override fun StructureND<Double>.svd(): Triple<DoubleTensor, DoubleTensor, DoubleTensor> {
|
||||
return this.svdGolubKahan()
|
||||
}
|
||||
|
||||
public fun StructureND<Double>.svdGolubKahan(): Triple<DoubleTensor, DoubleTensor, DoubleTensor> {
|
||||
val size = tensor.dimension
|
||||
val commonShape = tensor.shape.sliceArray(0 until size - 2)
|
||||
val (n, m) = tensor.shape.sliceArray(size - 2 until size)
|
||||
@ -855,7 +859,7 @@ public open class DoubleTensorAlgebra :
|
||||
.slice(matrix.bufferStart until matrix.bufferStart + matrixSize)
|
||||
.toDoubleArray()
|
||||
)
|
||||
curMatrix.as2D().svdHelper(uTensors[index].as2D(), sTensorVectors[index], vTensors[index].as2D())
|
||||
curMatrix.as2D().svdGolubKahanHelper(uTensors[index].as2D(), sTensorVectors[index], vTensors[index].as2D())
|
||||
}
|
||||
|
||||
return Triple(uTensor.transpose(), sTensor, vTensor)
|
||||
|
@ -372,7 +372,7 @@ private fun SIGN(a: Double, b: Double): Double {
|
||||
else
|
||||
return -abs(a)
|
||||
}
|
||||
internal fun MutableStructure2D<Double>.svdHelper(u: MutableStructure2D<Double>, w: BufferedTensor<Double>, v: MutableStructure2D<Double>) {
|
||||
internal fun MutableStructure2D<Double>.svdGolubKahanHelper(u: MutableStructure2D<Double>, w: BufferedTensor<Double>, v: MutableStructure2D<Double>) {
|
||||
val shape = this.shape
|
||||
val m = shape.component1()
|
||||
val n = shape.component2()
|
||||
|
@ -179,7 +179,7 @@ internal class TestDoubleLinearOpsTensorAlgebra {
|
||||
2.000000, 3.000000, 4.000000,
|
||||
3.000000, 4.000000, 5.000000,
|
||||
4.000000, 5.000000, 6.000000,
|
||||
5.000000, 6.000000, 7.000000
|
||||
5.000000, 6.000000, 9.000000
|
||||
)
|
||||
testSVDFor(fromArray(intArrayOf(5, 3), buffer1))
|
||||
val buffer2 = doubleArrayOf(
|
||||
@ -198,10 +198,52 @@ internal class TestDoubleLinearOpsTensorAlgebra {
|
||||
testSVDFor(tensor2)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun testSVDGolubKahan() = DoubleTensorAlgebra{
|
||||
testSVDGolubKahanFor(fromArray(intArrayOf(2, 3), doubleArrayOf(1.0, 2.0, 3.0, 4.0, 5.0, 6.0)))
|
||||
testSVDGolubKahanFor(fromArray(intArrayOf(2, 2), doubleArrayOf(-1.0, 0.0, 239.0, 238.0)))
|
||||
val buffer1 = doubleArrayOf(
|
||||
1.000000, 2.000000, 3.000000,
|
||||
2.000000, 3.000000, 4.000000,
|
||||
3.000000, 4.000000, 5.000000,
|
||||
4.000000, 5.000000, 6.000000,
|
||||
5.000000, 6.000000, 9.000000
|
||||
)
|
||||
testSVDGolubKahanFor(fromArray(intArrayOf(5, 3), buffer1))
|
||||
val buffer2 = doubleArrayOf(
|
||||
1.0, 2.0, 3.0, 2.0, 3.0,
|
||||
4.0, 3.0, 4.0, 5.0, 4.0,
|
||||
5.0, 6.0, 5.0, 6.0, 7.0
|
||||
)
|
||||
testSVDGolubKahanFor(fromArray(intArrayOf(3, 5), buffer2))
|
||||
}
|
||||
|
||||
@Test
|
||||
fun testBatchedSVDGolubKahan() = DoubleTensorAlgebra{
|
||||
val tensor1 = randomNormal(intArrayOf(2, 5, 3), 0)
|
||||
testSVDGolubKahanFor(tensor1)
|
||||
val tensor2 = DoubleTensorAlgebra.randomNormal(intArrayOf(30, 30, 30), 0)
|
||||
testSVDGolubKahanFor(tensor2)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun testSVDPowerMethod() = DoubleTensorAlgebra{
|
||||
testSVDPowerMethodFor(fromArray(intArrayOf(2, 3), doubleArrayOf(1.0, 2.0, 3.0, 4.0, 5.0, 6.0)))
|
||||
testSVDPowerMethodFor(fromArray(intArrayOf(2, 2), doubleArrayOf(-1.0, 0.0, 239.0, 238.0)))
|
||||
val buffer1 = doubleArrayOf(
|
||||
1.000000, 2.000000, 3.000000,
|
||||
2.000000, 3.000000, 4.000000,
|
||||
3.000000, 4.000000, 5.000000,
|
||||
4.000000, 5.000000, 6.000000,
|
||||
5.000000, 6.000000, 9.000000
|
||||
)
|
||||
testSVDPowerMethodFor(fromArray(intArrayOf(5, 3), buffer1))
|
||||
val buffer2 = doubleArrayOf(
|
||||
1.0, 2.0, 3.0, 2.0, 3.0,
|
||||
4.0, 3.0, 4.0, 5.0, 4.0,
|
||||
5.0, 6.0, 5.0, 6.0, 7.0
|
||||
)
|
||||
testSVDPowerMethodFor(fromArray(intArrayOf(3, 5), buffer2))
|
||||
}
|
||||
|
||||
@Test
|
||||
@ -237,6 +279,18 @@ private fun DoubleTensorAlgebra.testSVDFor(tensor: DoubleTensor) {
|
||||
assertTrue(tensor.eq(tensorSVD))
|
||||
}
|
||||
|
||||
private fun DoubleTensorAlgebra.testSVDGolubKahanFor(tensor: DoubleTensor) {
|
||||
val svd = tensor.svdGolubKahan()
|
||||
|
||||
val tensorSVD = svd.first
|
||||
.dot(
|
||||
diagonalEmbedding(svd.second)
|
||||
.dot(svd.third.transpose())
|
||||
)
|
||||
|
||||
assertTrue(tensor.eq(tensorSVD))
|
||||
}
|
||||
|
||||
private fun DoubleTensorAlgebra.testSVDPowerMethodFor(tensor: DoubleTensor, epsilon: Double = 1e-10) {
|
||||
val svd = tensor.svdPowerMethod()
|
||||
|
||||
|
Loading…
Reference in New Issue
Block a user