426 lines
12 KiB
Go
426 lines
12 KiB
Go
// Copyright ©2013 The Gonum Authors. All rights reserved.
|
||
// Use of this source code is governed by a BSD-style
|
||
// license that can be found in the LICENSE file.
|
||
|
||
package mat
|
||
|
||
import (
|
||
"gonum.org/v1/gonum/blas/blas64"
|
||
"gonum.org/v1/gonum/lapack"
|
||
"gonum.org/v1/gonum/lapack/lapack64"
|
||
)
|
||
|
||
const badRcond = "mat: invalid rcond value"
|
||
|
||
// SVD is a type for creating and using the Singular Value Decomposition
|
||
// of a matrix.
|
||
type SVD struct {
|
||
kind SVDKind
|
||
|
||
s []float64
|
||
u blas64.General
|
||
vt blas64.General
|
||
}
|
||
|
||
// SVDKind specifies the treatment of singular vectors during an SVD
|
||
// factorization.
|
||
type SVDKind int
|
||
|
||
const (
|
||
// SVDNone specifies that no singular vectors should be computed during
|
||
// the decomposition.
|
||
SVDNone SVDKind = 0
|
||
|
||
// SVDThinU specifies the thin decomposition for U should be computed.
|
||
SVDThinU SVDKind = 1 << (iota - 1)
|
||
// SVDFullU specifies the full decomposition for U should be computed.
|
||
SVDFullU
|
||
// SVDThinV specifies the thin decomposition for V should be computed.
|
||
SVDThinV
|
||
// SVDFullV specifies the full decomposition for V should be computed.
|
||
SVDFullV
|
||
|
||
// SVDThin is a convenience value for computing both thin vectors.
|
||
SVDThin SVDKind = SVDThinU | SVDThinV
|
||
// SVDFull is a convenience value for computing both full vectors.
|
||
SVDFull SVDKind = SVDFullU | SVDFullV
|
||
)
|
||
|
||
// succFact returns whether the receiver contains a successful factorization.
|
||
func (svd *SVD) succFact() bool {
|
||
return len(svd.s) != 0
|
||
}
|
||
|
||
// Factorize computes the singular value decomposition (SVD) of the input matrix A.
|
||
// The singular values of A are computed in all cases, while the singular
|
||
// vectors are optionally computed depending on the input kind.
|
||
//
|
||
// The full singular value decomposition (kind == SVDFull) is a factorization
|
||
// of an m×n matrix A of the form
|
||
//
|
||
// A = U * Σ * Vᵀ
|
||
//
|
||
// where Σ is an m×n diagonal matrix, U is an m×m orthogonal matrix, and V is an
|
||
// n×n orthogonal matrix. The diagonal elements of Σ are the singular values of A.
|
||
// The first min(m,n) columns of U and V are, respectively, the left and right
|
||
// singular vectors of A.
|
||
//
|
||
// Significant storage space can be saved by using the thin representation of
|
||
// the SVD (kind == SVDThin) instead of the full SVD, especially if
|
||
// m >> n or m << n. The thin SVD finds
|
||
//
|
||
// A = U~ * Σ * V~ᵀ
|
||
//
|
||
// where U~ is of size m×min(m,n), Σ is a diagonal matrix of size min(m,n)×min(m,n)
|
||
// and V~ is of size n×min(m,n).
|
||
//
|
||
// Factorize returns whether the decomposition succeeded. If the decomposition
|
||
// failed, routines that require a successful factorization will panic.
|
||
func (svd *SVD) Factorize(a Matrix, kind SVDKind) (ok bool) {
|
||
// kill previous factorization
|
||
svd.s = svd.s[:0]
|
||
svd.kind = kind
|
||
|
||
m, n := a.Dims()
|
||
var jobU, jobVT lapack.SVDJob
|
||
|
||
// TODO(btracey): This code should be modified to have the smaller
|
||
// matrix written in-place into aCopy when the lapack/native/dgesvd
|
||
// implementation is complete.
|
||
switch {
|
||
case kind&SVDFullU != 0:
|
||
jobU = lapack.SVDAll
|
||
svd.u = blas64.General{
|
||
Rows: m,
|
||
Cols: m,
|
||
Stride: m,
|
||
Data: use(svd.u.Data, m*m),
|
||
}
|
||
case kind&SVDThinU != 0:
|
||
jobU = lapack.SVDStore
|
||
svd.u = blas64.General{
|
||
Rows: m,
|
||
Cols: min(m, n),
|
||
Stride: min(m, n),
|
||
Data: use(svd.u.Data, m*min(m, n)),
|
||
}
|
||
default:
|
||
jobU = lapack.SVDNone
|
||
}
|
||
switch {
|
||
case kind&SVDFullV != 0:
|
||
svd.vt = blas64.General{
|
||
Rows: n,
|
||
Cols: n,
|
||
Stride: n,
|
||
Data: use(svd.vt.Data, n*n),
|
||
}
|
||
jobVT = lapack.SVDAll
|
||
case kind&SVDThinV != 0:
|
||
svd.vt = blas64.General{
|
||
Rows: min(m, n),
|
||
Cols: n,
|
||
Stride: n,
|
||
Data: use(svd.vt.Data, min(m, n)*n),
|
||
}
|
||
jobVT = lapack.SVDStore
|
||
default:
|
||
jobVT = lapack.SVDNone
|
||
}
|
||
|
||
// A is destroyed on call, so copy the matrix.
|
||
aCopy := DenseCopyOf(a)
|
||
svd.kind = kind
|
||
svd.s = use(svd.s, min(m, n))
|
||
|
||
work := []float64{0}
|
||
lapack64.Gesvd(jobU, jobVT, aCopy.mat, svd.u, svd.vt, svd.s, work, -1)
|
||
work = getFloat64s(int(work[0]), false)
|
||
ok = lapack64.Gesvd(jobU, jobVT, aCopy.mat, svd.u, svd.vt, svd.s, work, len(work))
|
||
putFloat64s(work)
|
||
if !ok {
|
||
svd.kind = 0
|
||
}
|
||
return ok
|
||
}
|
||
|
||
// Kind returns the SVDKind of the decomposition. If no decomposition has been
|
||
// computed, Kind returns -1.
|
||
func (svd *SVD) Kind() SVDKind {
|
||
if !svd.succFact() {
|
||
return -1
|
||
}
|
||
return svd.kind
|
||
}
|
||
|
||
// Rank returns the rank of A based on the count of singular values greater than
|
||
// rcond scaled by the largest singular value.
|
||
// Rank will panic if the receiver does not contain a successful factorization or
|
||
// rcond is negative.
|
||
func (svd *SVD) Rank(rcond float64) int {
|
||
if rcond < 0 {
|
||
panic(badRcond)
|
||
}
|
||
if !svd.succFact() {
|
||
panic(badFact)
|
||
}
|
||
s0 := svd.s[0]
|
||
for i, v := range svd.s {
|
||
if v <= rcond*s0 {
|
||
return i
|
||
}
|
||
}
|
||
return len(svd.s)
|
||
}
|
||
|
||
// Cond returns the 2-norm condition number for the factorized matrix. Cond will
|
||
// panic if the receiver does not contain a successful factorization.
|
||
func (svd *SVD) Cond() float64 {
|
||
if !svd.succFact() {
|
||
panic(badFact)
|
||
}
|
||
return svd.s[0] / svd.s[len(svd.s)-1]
|
||
}
|
||
|
||
// Values returns the singular values of the factorized matrix in descending order.
|
||
//
|
||
// If the input slice is non-nil, the values will be stored in-place into
|
||
// the slice. In this case, the slice must have length min(m,n), and Values will
|
||
// panic with ErrSliceLengthMismatch otherwise. If the input slice is nil, a new
|
||
// slice of the appropriate length will be allocated and returned.
|
||
//
|
||
// Values will panic if the receiver does not contain a successful factorization.
|
||
func (svd *SVD) Values(s []float64) []float64 {
|
||
if !svd.succFact() {
|
||
panic(badFact)
|
||
}
|
||
if s == nil {
|
||
s = make([]float64, len(svd.s))
|
||
}
|
||
if len(s) != len(svd.s) {
|
||
panic(ErrSliceLengthMismatch)
|
||
}
|
||
copy(s, svd.s)
|
||
return s
|
||
}
|
||
|
||
// UTo extracts the matrix U from the singular value decomposition. The first
|
||
// min(m,n) columns are the left singular vectors and correspond to the singular
|
||
// values as returned from SVD.Values.
|
||
//
|
||
// If dst is empty, UTo will resize dst to be m×m if the full U was computed
|
||
// and size m×min(m,n) if the thin U was computed. When dst is non-empty, then
|
||
// UTo will panic if dst is not the appropriate size. UTo will also panic if
|
||
// the receiver does not contain a successful factorization, or if U was
|
||
// not computed during factorization.
|
||
func (svd *SVD) UTo(dst *Dense) {
|
||
if !svd.succFact() {
|
||
panic(badFact)
|
||
}
|
||
kind := svd.kind
|
||
if kind&SVDThinU == 0 && kind&SVDFullU == 0 {
|
||
panic("svd: u not computed during factorization")
|
||
}
|
||
r := svd.u.Rows
|
||
c := svd.u.Cols
|
||
if dst.IsEmpty() {
|
||
dst.ReuseAs(r, c)
|
||
} else {
|
||
r2, c2 := dst.Dims()
|
||
if r != r2 || c != c2 {
|
||
panic(ErrShape)
|
||
}
|
||
}
|
||
|
||
tmp := &Dense{
|
||
mat: svd.u,
|
||
capRows: r,
|
||
capCols: c,
|
||
}
|
||
dst.Copy(tmp)
|
||
}
|
||
|
||
// VTo extracts the matrix V from the singular value decomposition. The first
|
||
// min(m,n) columns are the right singular vectors and correspond to the singular
|
||
// values as returned from SVD.Values.
|
||
//
|
||
// If dst is empty, VTo will resize dst to be n×n if the full V was computed
|
||
// and size n×min(m,n) if the thin V was computed. When dst is non-empty, then
|
||
// VTo will panic if dst is not the appropriate size. VTo will also panic if
|
||
// the receiver does not contain a successful factorization, or if V was
|
||
// not computed during factorization.
|
||
func (svd *SVD) VTo(dst *Dense) {
|
||
if !svd.succFact() {
|
||
panic(badFact)
|
||
}
|
||
kind := svd.kind
|
||
if kind&SVDThinV == 0 && kind&SVDFullV == 0 {
|
||
panic("svd: v not computed during factorization")
|
||
}
|
||
r := svd.vt.Rows
|
||
c := svd.vt.Cols
|
||
if dst.IsEmpty() {
|
||
dst.ReuseAs(c, r)
|
||
} else {
|
||
r2, c2 := dst.Dims()
|
||
if c != r2 || r != c2 {
|
||
panic(ErrShape)
|
||
}
|
||
}
|
||
|
||
tmp := &Dense{
|
||
mat: svd.vt,
|
||
capRows: r,
|
||
capCols: c,
|
||
}
|
||
dst.Copy(tmp.T())
|
||
}
|
||
|
||
// SolveTo calculates the minimum-norm solution to a linear least squares problem
|
||
//
|
||
// minimize over n-element vectors x: |b - A*x|_2 and |x|_2
|
||
//
|
||
// where b is a given m-element vector, using the SVD of m×n matrix A stored in
|
||
// the receiver. A may be rank-deficient, that is, the given effective rank can be
|
||
//
|
||
// rank ≤ min(m,n)
|
||
//
|
||
// The rank can be computed using SVD.Rank.
|
||
//
|
||
// Several right-hand side vectors b and solution vectors x can be handled in a
|
||
// single call. Vectors b are stored in the columns of the m×k matrix B and the
|
||
// resulting vectors x will be stored in the columns of dst. dst must be either
|
||
// empty or have the size equal to n×k.
|
||
//
|
||
// The decomposition must have been factorized computing both the U and V
|
||
// singular vectors.
|
||
//
|
||
// SolveTo returns the residuals calculated from the complete SVD. For this
|
||
// value to be valid the factorization must have been performed with at least
|
||
// SVDFullU.
|
||
func (svd *SVD) SolveTo(dst *Dense, b Matrix, rank int) []float64 {
|
||
if !svd.succFact() {
|
||
panic(badFact)
|
||
}
|
||
if rank < 1 || len(svd.s) < rank {
|
||
panic("svd: rank out of range")
|
||
}
|
||
kind := svd.kind
|
||
if kind&SVDThinU == 0 && kind&SVDFullU == 0 {
|
||
panic("svd: u not computed during factorization")
|
||
}
|
||
if kind&SVDThinV == 0 && kind&SVDFullV == 0 {
|
||
panic("svd: v not computed during factorization")
|
||
}
|
||
|
||
u := Dense{
|
||
mat: svd.u,
|
||
capRows: svd.u.Rows,
|
||
capCols: svd.u.Cols,
|
||
}
|
||
vt := Dense{
|
||
mat: svd.vt,
|
||
capRows: svd.vt.Rows,
|
||
capCols: svd.vt.Cols,
|
||
}
|
||
s := svd.s[:rank]
|
||
|
||
_, bc := b.Dims()
|
||
c := getDenseWorkspace(svd.u.Cols, bc, false)
|
||
defer putDenseWorkspace(c)
|
||
c.Mul(u.T(), b)
|
||
|
||
y := getDenseWorkspace(rank, bc, false)
|
||
defer putDenseWorkspace(y)
|
||
y.DivElem(c.slice(0, rank, 0, bc), repVector{vec: s, cols: bc})
|
||
dst.Mul(vt.slice(0, rank, 0, svd.vt.Cols).T(), y)
|
||
|
||
res := make([]float64, bc)
|
||
if rank < svd.u.Cols {
|
||
c = c.slice(len(s), svd.u.Cols, 0, bc)
|
||
for j := range res {
|
||
col := c.ColView(j)
|
||
res[j] = Dot(col, col)
|
||
}
|
||
}
|
||
return res
|
||
}
|
||
|
||
type repVector struct {
|
||
vec []float64
|
||
cols int
|
||
}
|
||
|
||
func (m repVector) Dims() (r, c int) { return len(m.vec), m.cols }
|
||
func (m repVector) At(i, j int) float64 {
|
||
if i < 0 || len(m.vec) <= i || j < 0 || m.cols <= j {
|
||
panic(ErrIndexOutOfRange.string) // Panic with string to prevent mat.Error recovery.
|
||
}
|
||
return m.vec[i]
|
||
}
|
||
func (m repVector) T() Matrix { return Transpose{m} }
|
||
|
||
// SolveVecTo calculates the minimum-norm solution to a linear least squares problem
|
||
//
|
||
// minimize over n-element vectors x: |b - A*x|_2 and |x|_2
|
||
//
|
||
// where b is a given m-element vector, using the SVD of m×n matrix A stored in
|
||
// the receiver. A may be rank-deficient, that is, the given effective rank can be
|
||
//
|
||
// rank ≤ min(m,n)
|
||
//
|
||
// The rank can be computed using SVD.Rank.
|
||
//
|
||
// The resulting vector x will be stored in dst. dst must be either empty or
|
||
// have length equal to n.
|
||
//
|
||
// The decomposition must have been factorized computing both the U and V
|
||
// singular vectors.
|
||
//
|
||
// SolveVecTo returns the residuals calculated from the complete SVD. For this
|
||
// value to be valid the factorization must have been performed with at least
|
||
// SVDFullU.
|
||
func (svd *SVD) SolveVecTo(dst *VecDense, b Vector, rank int) float64 {
|
||
if !svd.succFact() {
|
||
panic(badFact)
|
||
}
|
||
if rank < 1 || len(svd.s) < rank {
|
||
panic("svd: rank out of range")
|
||
}
|
||
kind := svd.kind
|
||
if kind&SVDThinU == 0 && kind&SVDFullU == 0 {
|
||
panic("svd: u not computed during factorization")
|
||
}
|
||
if kind&SVDThinV == 0 && kind&SVDFullV == 0 {
|
||
panic("svd: v not computed during factorization")
|
||
}
|
||
|
||
u := Dense{
|
||
mat: svd.u,
|
||
capRows: svd.u.Rows,
|
||
capCols: svd.u.Cols,
|
||
}
|
||
vt := Dense{
|
||
mat: svd.vt,
|
||
capRows: svd.vt.Rows,
|
||
capCols: svd.vt.Cols,
|
||
}
|
||
s := svd.s[:rank]
|
||
|
||
c := getVecDenseWorkspace(svd.u.Cols, false)
|
||
defer putVecDenseWorkspace(c)
|
||
c.MulVec(u.T(), b)
|
||
|
||
y := getVecDenseWorkspace(rank, false)
|
||
defer putVecDenseWorkspace(y)
|
||
y.DivElemVec(c.sliceVec(0, rank), NewVecDense(rank, s))
|
||
dst.MulVec(vt.slice(0, rank, 0, svd.vt.Cols).T(), y)
|
||
|
||
var res float64
|
||
if rank < c.Len() {
|
||
c = c.sliceVec(rank, c.Len())
|
||
res = Dot(c, c)
|
||
}
|
||
return res
|
||
}
|