hyper/hypercubes.go

118 lines
3.2 KiB
Go

package hyper
// Params returns discretization parameters.
// numBuckets represents number of discretization buckets into
// which all values will fall. Ids of those buckets will be used
// to create hashes.
// min and max are minimum and maximum possible values
// of discretized variable.
// bucketWidth is width of the discretization bucket.
// bucketPct is percentage of bucketWidth to allow for an error
// of discretized variable (a specific value of a discretized
// variable may fall into 2 buckets simultaneosly).
// eps is actual width corresponding to the bucketWidth bucketPct
// on the discretized variable axis.
func Params(
numBuckets int, min, max, bucketPct float64) (bucketWidth, eps float64) {
if bucketPct >= 0.5 {
panic(`Error: bucketPct must be less than 50%.
Recommendation: decrease numBuckets instead.`)
}
bucketWidth = (max - min) / float64(numBuckets)
eps = bucketPct * bucketWidth
return bucketWidth, eps
}
// Hypercubes returns a set of hypercubes, which represent
// fuzzy discretization of one n-dimensional vector, as described in
// https://vitali-fedulov.github.io/algorithm-for-hashing-high-dimensional-float-vectors.html
// One hupercube is defined by bucket numbers in each dimension.
// The function also returns the central hypercube (in which
// the vector end is located).
// min and max are minimum and maximum possible values of
// the vector components. The assumption is that min and max
// are the same for all dimensions.
// bucketWidth and eps are defined in the Params function.
func Hypercubes(
vector []float64, min, max, bucketWidth, eps float64) (
set [][]int, central []int) {
var (
bC, bS int // Central and side bucket ids.
setCopy [][]int // Set copy.
length int
branching bool // Branching flag.
)
// For each component of the vector.
for _, val := range vector {
bC = int(val / bucketWidth)
central = append(central, bC)
branching = false
// Value is in the lower uncertainty interval.
if val-float64(bC)*bucketWidth < eps {
bS = bC - 1
if val-eps >= min {
branching = true
}
// Value is in the upper uncertainty interval.
} else if float64(bC+1)*bucketWidth-val < eps {
bS = bC + 1
if val+eps <= max {
branching = true
}
}
if branching {
setCopy = make([][]int, len(set))
copy(setCopy, set)
if len(set) == 0 {
set = append(set, []int{bC})
} else {
length = len(set)
for i := 0; i < length; i++ {
set[i] = append(set[i], bC)
}
}
if len(setCopy) == 0 {
setCopy = append(setCopy, []int{bS})
} else {
length = len(setCopy)
for i := 0; i < length; i++ {
setCopy[i] = append(setCopy[i], bS)
}
}
set = append(set, setCopy...)
} else {
if len(set) == 0 {
set = append(set, []int{bC})
} else {
length = len(set)
for i := 0; i < length; i++ {
set[i] = append(set[i], bC)
}
}
}
}
// Real use case verification that branching works correctly
// and no buckets are lost for a very large number of vectors.
// TODO: Remove once tested.
length = len(vector)
for i := 0; i < len(set); i++ {
if len(set[i]) != length {
panic(`Number of hypercube coordinates must equal to len(vector).`)
}
}
return set, central
}