GoRedSVD

GoRedSVD is some Swig-generated Go glue wrapping around Nicolas Tessore's RedSVD-h, which is based on Daisuke Okanohara's redsvd implementation.

What is RedSVD?

redsvd is a library for solving several matrix decompositions including singular value decomposition (SVD), principal component analysis (PCA), and eigen value decomposition. redsvd can handle very large matrix efficiently, and optimized for a truncated SVD of sparse matrices.

(quoted from redsvd project page)

GoRedSVD implements only the SVD part, but it should be reasonably easy to adapt SymEigen and PCA parts also.

Build requirements

• Swig 3.0.6 or newer (tested on 3.0.7)
• Eigen 3 library (tested on 3.2.6)

It should also be possible to link in Intel®'s Math Kernel Library, which should enable multicore processing for the SVD.

Building

Build process is roughly this:

1. Build files using swig
2. Copy the resulting files into gopath
3. Provide cgo with the location of Eigen's include path
4. Build with go build.
5. You can run tests with go test.

swig -go -intgosize 64 -cgo -c++ src/RedSVD.i
mkdir -p $GOPATH/src/redsvd
cp RedSVD-h RedSVD.h RedSVD_wrap.cxx redsvd.go redsvd_test.go $GOPATH/src/redsvd

cd $GOPATH/src/redsvd
export CGO_CPPFLAGS='-I/usr/local/Cellar/eigen/3.2.6/include/eigen3/'
go build

Or you can use the included Makefile.

API

• NewGoRedSVD() GoRedSVD: creates new GoRedSVD instance
• SetMatrix32(cols int, rows int, matrix map[int]map[int]float32): sets a sparse matrix
• SetMatrix64(cols int, rows int, matrix map[int]map[int]float64): float64 version of previous (simply typecasted to float32 currently)
• SetDenseMatrix32(matrix [][]float32): sets a dense matrix
• RedSVD() bool: performs SVD
• RedSVD(rank int) bool: performs SVD with specified rank
• MatrixU() [][]float32: returns U
• MatrixV() [][]float32: returns V
• SetUnnormalized(unnormalized bool): whether to also calculate un-normalized versions of the U and V matrices
  • MatrixUNotNormalized() [][]float32: returns U, not normalized
  • MatrixVNotNormalized() [][]float32: returns V, not normalized
• SingularValues() []float32: returns singular values
• DeleteGoRedSVD(arg GoRedSVD): frees the GoRedSVD instance

License

Copyright (c) 2016 Rocket Internet

Please see LICENSE file in repository (MIT). The RedSVD-h header is licensed separately using BSD license (see RedVSD-h file for license details).

Sample code

package main

import (
       "redsvd"
       "fmt"
       "math/rand"
)

func main() {
     svd := redsvd.NewGoRedSVD()

     fmt.Printf("Generating random matrix\n")

     sparseMatrix32 := make(map[int]map[int]float32)
     sparseMatrix64 := make(map[int]map[int]float64)
     for row := 0; row < 10; row++ {
         sparseMatrix32[row] = make(map[int]float32)
		 sparseMatrix64[row] = make(map[int]float64)
		 for col := 0; col < 10; col++ {
			 sparseMatrix32[row][col] = rand.Float32()
			 sparseMatrix64[row][col] = rand.Float64()
		 }
     }

     fmt.Printf("sparseMatrix32: %v\n", sparseMatrix32)
     fmt.Printf("sparseMatrix64: %v\n", sparseMatrix64)

     svd.SetMatrix32(10, 10, sparseMatrix32)

     svd.RedSVD()

     u := svd.MatrixU()
     fmt.Printf("float32 matrixU %v\n", u)

     v := svd.MatrixV()
     fmt.Printf("float32 matrixV %v\n", v)

     singularValues := svd.SingularValues()
     fmt.Printf("float 32 singularValues %v\n", singularValues)

     svd.SetMatrix64(10, 10, sparseMatrix64)
     svd.RedSVD()

     u = svd.MatrixU()
     fmt.Printf("float64 matrixU %v\n", u)

     v = svd.MatrixV()
     fmt.Printf("float64 matrixV %v\n", v)

     singularValues = svd.SingularValues()
     fmt.Printf("float64 singularValues %v\n", singularValues)

     redsvd.DeleteGoRedSVD(svd)
}