func TestLSHStream(t *testing.T) {
var seed int64 = 0
var d, k, l int = 64, 6, 4
data := []float64{1.0, 0.0, 2.0, 7.0, 4.0, 0.0, 8.0, 3.0, 2.0, 1.0}
var inDimensions, outDimentions int = 10, 2
hash := hash.NewMurmur(1<<63 - 1)
decoder := decoder.NewSpherical(d, k, l)
projector := projector.NewDBFriendly(inDimensions, outDimentions, seed)
lsh := lsh.NewLSH(hash, decoder, projector)
lsh.LSHHashStream(data, 1)
t.Log("√ LSH Stream test complete")
}
func BenchmarkStreamLSH(b *testing.B) {
var seed int64 = 0
var d, k, l int = 64, 6, 4
data := []float64{1.0, 0.0, 2.0, 7.0, 4.0, 0.0, 8.0, 3.0, 2.0, 1.0}
var inDimensions, outDimentions int = 10, 2
hash := hash.NewMurmur(1<<63 - 1)
decoder := decoder.NewSpherical(d, k, l)
projector := projector.NewDBFriendly(inDimensions, outDimentions, seed)
for i := 0; i < b.N; i++ {
lsh := lsh.NewLSH(hash, decoder, projector)
b.StopTimer()
lsh.LSHHashStream(data, 1)
b.StartTimer()
}
}
func BenchmarkDBFriendlyProjection(b *testing.B) {
var inDimensions, outDimentions int = 10, 2
for i := 0; i < b.N; i++ {
b.StopTimer()
var randomGen = rand.New(rand.NewSource(int64(time.Now().Nanosecond())))
data := make([]float64, inDimensions)
for i := 0; i < inDimensions; i++ {
data[i] = randomGen.Float64()
}
b.StartTimer()
var seed int64 = int64(time.Now().Nanosecond())
RP := projector.NewDBFriendly(inDimensions, outDimentions, seed)
RP.Project(data)
}
}
func (this *KMeans) Run() {
//This is a condition to avoid infinite Run..
maxiters := 10000
swaps := 3
fulldata := this.data
data := make([][]float64, 0)
var p types.Projector = nil
if this.projectionDimension != 0 {
p = projector.NewDBFriendly(len(fulldata[0]), this.projectionDimension, rand.Int63())
}
for _, v := range fulldata {
if p != nil {
data = append(data, p.Project(v))
} else {
data = append(data, v)
}
}
this.n = len(data)
this.means = make([][]float64, this.k)
for i := 0; i < this.k; i++ {
this.means[i] = data[i*(this.n/this.k)]
}
this.clusters = make([][]int, this.k)
//initilize cluster lists to be evenly diveded sequentailly
for i := 0; i < this.k; i++ {
cluster := make([]int, this.n/this.k)
clusterStart := i * (this.n / this.k)
for j := 0; j < this.n/this.k; j++ {
cluster[j] = j + clusterStart
}
this.clusters[i] = cluster
}
for swaps > 2 && maxiters > 0 {
maxiters--
this.UpdateMeans(data)
swaps = this.AssignClusters(data)
}
if maxiters == 0 {
fmt.Println("Warning: Max Iterations Reached")
}
data = fulldata
this.UpdateMeans(data)
}
func TestDBFriendly(t *testing.T) {
//There is probably a better way to test this than hard coding.
data := []float64{1.0, 0.0, 2.0, 7.0, 4.0, 0.0, 8.0, 3.0, 2.0, 1.0}
expectedResult := []float64{1.224744871391589, 13.472193585307478}
var inDimensions, outDimentions int = 10, 2
//Use a uniform seed for testing
var seed int64 = 0
RP := projector.NewDBFriendly(inDimensions, outDimentions, seed)
result := RP.Project(data)
if len(result) != len(expectedResult) {
t.Error("The result and expected result are not the same length.")
}
for i := 0; i < len(result); i++ {
if result[i] != expectedResult[i] {
t.Error(fmt.Sprintf("The result at index %d: %f did not match the expected result: %f", i, result[i], expectedResult[i]))
}
}
t.Log("√ DBFriendly Projector test complete")
}
// The datapoints are seeded in so that the first two data points are near eachother in euclidian geometery and the 3rd and 4th datapoint are
// near eachother in euclidian geometery. So the result1Cluster1 and result2Cluster1 should be closer together than the other two points.
// The same is true for the points in cluster two vs either point in cluster one.
func TestLSHSimple(t *testing.T) {
var seed int64 = 0
// We want to limit the dimension reduction because it causes a lot of noise.
var inDimensions, outDimentions, numberOfClusters, numberOfSearches int = 10, 5, 3, 1
dataPoint1Cluster1 := []float64{1.0, 0.0, 2.0, 7.0, 4.0, 0.0, 8.0, 3.0, 2.0, 1.0}
dataPoint2Cluster1 := []float64{2.0, 3.0, 2.0, 6.0, 5.5, 2.0, 8.0, 3.1, 2.0, 0.0}
dataPoint1Cluster2 := []float64{100.0, -120.0, 6.0, 18.0, 209.0, 0.0, -2.0, 1036.0, 15.0, 123.0}
dataPoint2Cluster2 := []float64{99.0, -119.0, 2.0, 18.0, 208.5, 0.0, -3.0, 1048.0, 13.0, 122.0}
hash := hash.NewMurmur(1<<63 - 1)
decoder := decoder.NewSpherical(inDimensions, numberOfClusters, numberOfSearches)
projector := projector.NewDBFriendly(inDimensions, outDimentions, seed)
lsh := lsh.NewLSH(hash, decoder, projector)
result1Cluster1 := lsh.LSHHashSimple(dataPoint1Cluster1)
result2Cluster1 := lsh.LSHHashSimple(dataPoint2Cluster1)
result1Cluster2 := lsh.LSHHashSimple(dataPoint1Cluster2)
result2Cluster2 := lsh.LSHHashSimple(dataPoint2Cluster2)
// Assert that results are still localy sensetive based on the original euclidian geometry
if math.Abs(float64(result1Cluster1-result2Cluster1)) > math.Abs(float64(result1Cluster1-result1Cluster2)) {
t.Errorf("\nThe first datapoint in cluster two is closer to the first data point in cluster one than the second data point in cluster one"+
"\ndatapoint cluster one datapoint one: %d, \ndatapoint cluster one datapoint two: %d, \ndatapoint cluster two datapoint one: %d",
result1Cluster1, result2Cluster1, result1Cluster2)
}
if math.Abs(float64(result1Cluster1-result2Cluster1)) > math.Abs(float64(result1Cluster1-result2Cluster2)) {
t.Errorf("\nThe second datapoint in cluster two is closer to the first data point in cluster one than the second data point in cluster one"+
"\nCluster one datapoint one: %d, \nCluster one datapoint two: %d, \nCluster two datapoint two: %d",
result1Cluster1, result2Cluster1, result2Cluster2)
}
if math.Abs(float64(result1Cluster2-result2Cluster2)) > math.Abs(float64(result1Cluster1-result1Cluster2)) {
t.Errorf("\nThe first datapoint in cluster one is closer to the first data point in cluster two than the second data point in cluster two"+
"\nCluster one datapoint one: %d, \nCluster two datapoint one: %d, \nCluster two datapoint two: %d",
result1Cluster1, result1Cluster2, result2Cluster2)
}
t.Log("√ LSH Simple test complete")
}
func NewProjector(n, t int, randomseed int64) types.Projector {
return projector.NewDBFriendly(n, t, randomseed)
}