Skip to content

Commit

Permalink
Simplified/optimized Vec.zeros[N].
Browse files Browse the repository at this point in the history 
Added rankVariableSort: Ordering[(Double, Int)]

Removed:
 - inline modifiers from vector extension methods that loop over vec data
 - references to zip, map, and sum from covariance method
 - references to ListBuffer, Vector, tabulate, fill, unzip, zip, etc from elementRanks method
  • Loading branch information
c committed Oct 3, 2023
1 parent 0350f1d commit 91feb55
Showing 1 changed file with 48 additions and 46 deletions.
94 changes: 48 additions & 46 deletions slash/shared/src/main/scala/ai/dragonfly/math/vector/package.scala
View file
Original file line number Diff line number Diff line change
Expand Up @@ -21,21 +21,23 @@ import ai.dragonfly.math.vector.Vec
import unicode.*
import narr.*

import scala.language.implicitConversions

import scala.compiletime.ops.any.==
import scala.compiletime.ops.int.*

import scala.collection.mutable.ListBuffer // for fast append in elementRanks

package object vector {

opaque type Vec[N <: Int] = NArray[Double]
object Vec {

export narr.Extensions.given
inline def apply[N <: Int](a: NArray[Double]): Vec[N] = { // cast a NArray[Double] as Vec[N]
dimensionCheck(a, valueOf[N])
a
}

inline def zeros[N <: Int](using ValueOf[N]): Vec[N] = fill[N](0.0)
inline def zeros[N <: Int](using ValueOf[N]): Vec[N] = new DoubleArray(valueOf[N]).asInstanceOf[Vec[N]]

inline def ones[N <: Int](using ValueOf[N]): Vec[N] = fill[N](1.0)

Expand Down Expand Up @@ -94,6 +96,8 @@ package object vector {
def fromTuple(t: (Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double)):Vec[21] = Vec.tabulate(i => DBL(t(i)))
def fromTuple(t: (Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double, Double)):Vec[22] = Vec.tabulate(i => DBL(t(i)))

val rankVariableSort: Ordering[(Double, Int)] = Ordering.by(_._1)

extension[N <: Int] (thisVector: Vec[N])(using ValueOf[N], N >= 1 =:= true) {
inline def x: Double = thisVector(0)
}
Expand Down Expand Up @@ -168,7 +172,7 @@ package object vector {
copyOfThisVector
}

inline def sum = {
def sum = {
var sum = 0.0
var i = 0
while (i < thisVector.dimension) {
Expand All @@ -178,31 +182,35 @@ package object vector {
sum
}

inline def mean: Double = {
thisVector.sum / thisVector.dimension
}
inline def mean: Double = thisVector.sum / thisVector.dimension

//It is assumed, that we consider a sample rather than a complete population
inline def variance: Double = {
def variance: Double = {
// https://www.cuemath.com/sample-variance-formula/
val μ = thisVector.mean
thisVector.map(i => squareInPlace(i - μ)).sum / (thisVector.size - 1)
}

// It is assumed, that we consider a sample rather than a complete population
inline def stdDev: Double = {
def stdDev: Double = {
// https://www.cuemath.com/data/standard-deviation/
val mu = thisVector.mean
val diffs_2 = thisVector.map( num => squareInPlace(num - mu) )
Math.sqrt( diffs_2.sum / (thisVector.size - 1 ) )
}

def covariance(thatVector : Vec[N] ) = {
def covariance(thatVector : Vec[N] ):Double = {
val μThis = thisVector.mean
val μThat = thatVector.mean
thisVector.zip(thatVector).map{ case (thisV, thatV) => (thisV - μThis) * (thatV - μThat) }.sum / (thisVector.size -1)
var cv:Double = 0
var i:Int = 0; while (i < thisVector.dimension) {
cv += (thisVector(i) - μThis) * (thatVector(i) - μThat)
i += 1
}
cv / (thisVector.dimension -1)
}

inline def pearsonCorrelationCoefficient(thatVector: Vec[N]): Double = {
def pearsonCorrelationCoefficient(thatVector: Vec[N]): Double = {
val n = thisVector.size
var i = 0

Expand All @@ -223,48 +231,42 @@ package object vector {
(n * sum_xy - (sum_x * sum_y)) / Math.sqrt( (sum_x2 * n - squareInPlace(sum_x)) * (sum_y2 * n - squareInPlace(sum_y)) )
}

inline def spearmansRankCorrelation(thatVector: Vec[N]) : Double =
def spearmansRankCorrelation(thatVector: Vec[N]) : Double = {
val theseRanks = thisVector.elementRanks
val thoseRanks = thatVector.elementRanks
theseRanks.pearsonCorrelationCoefficient(thoseRanks)
}


// An alias - pearson is the most commonly requested type of correlation
inline def corr(thatVector: Vec[N]): Double = pearsonCorrelationCoefficient(thatVector)

inline def elementRanks: Vec[N] = {
val (sorted, originalPosition) = thisVector.zipWithIndex.toVector.sortBy(_._1).unzip
val ranks : Vec[N] = NArray.tabulate[Double](thisVector.dimension)(i => (i+1).toDouble)

var currentValue = sorted(0)
var i = 0
var currentSum = 0.0
var currentCount = 0
val resultList = ListBuffer[Double]()


for (value <- sorted) {
if (value == currentValue) {
currentSum += ranks(i)
currentCount += 1
} else {
resultList.appendAll(ListBuffer.fill(currentCount)(currentSum / currentCount))
currentValue = value
currentCount = 1
currentSum = ranks(i)
def elementRanks: Vec[N] = {
val indexed:NArray[(Double, Int)] = thisVector.zipWithIndex
indexed.sort(rankVariableSort)

val ranks : Vec[N] = new DoubleArray(thisVector.dimension) // faster than zeros.
ranks(indexed.last._2) = thisVector.dimension
var currentValue:Double = indexed(0)._1
var r0:Int = 0
var rank:Int = 1
while (rank < thisVector.dimension) {
val temp: Double = indexed(rank)._1
if (temp != currentValue) {
val avg: Double = (1.0 + (rank + r0)) / 2.0
var i:Int = r0; while (i < rank) {
ranks(indexed(i)._2) = avg
i += 1
}
r0 = rank
currentValue = temp
}
i = i + 1
}
resultList.appendAll(List.fill(currentCount)(currentSum / currentCount))

val rankResult : Vec[N] = new NArray[Double](thisVector.dimension)
for( (idx, r) <- originalPosition.zip(resultList)) {
rankResult(idx) = r
rank += 1
}
rankResult
ranks
}

inline def normSquared: Double = {
def normSquared: Double = {
var mag2 = 0.0
var i = 0
while (i < dimension) {
Expand All @@ -280,7 +282,7 @@ package object vector {

inline def magnitudeSquared: Double = normSquared

inline def euclideanDistanceSquaredTo(v0: Vec[N]): Double = {
def euclideanDistanceSquaredTo(v0: Vec[N]): Double = {
var distance = 0.0
var i = 0
while (i < dimension) {
Expand Down Expand Up @@ -333,7 +335,7 @@ package object vector {
inline def *(scalar: Double): Vec[N] = copy.scale(scalar)

inline def *= (scalar: Double): Vec[N] = scale(scalar)
inline def scale(scalar: Double): Vec[N] = {
def scale(scalar: Double): Vec[N] = {
var i = 0
while (i < dimension) {
thisVector(i) = thisVector(i) * scalar
Expand All @@ -346,7 +348,7 @@ package object vector {

inline def divide(divisor: Double): Vec[N] = scale(1.0 / divisor)

inline def round(): Unit = {
def round(): Unit = {
var i = 0
while (i < dimension) {
thisVector(i) = Math.round(thisVector(i)).toDouble
Expand All @@ -356,7 +358,7 @@ package object vector {

inline def discretize(): Unit = round()

inline def discretize(r: Double): Unit = {
def discretize(r: Double): Unit = {
var i = 0
while (i < dimension) {
thisVector(i) = r * Math.round(thisVector(i) / r).toDouble
Expand Down

0 comments on commit 91feb55

Please sign in to comment.