implemented QR factorization

this is a first sketch, the algorithm is not yet initialized and relies
on knowledge of DMat internals. A next step would be to implement this
algorithm in a more generic manner.
This commit is contained in:
Vincent Barrielle 2014-05-09 18:59:26 +02:00
parent 9badebf24c
commit 2fd880a62d
4 changed files with 95 additions and 1 deletions

View File

@ -44,6 +44,7 @@ pub use traits::{
pub use structs::{ pub use structs::{
Identity, Identity,
decomp_qr,
DMat, DVec, DMat, DVec,
Iso2, Iso3, Iso4, Iso2, Iso3, Iso4,
Mat1, Mat2, Mat3, Mat4, Mat1, Mat2, Mat3, Mat4,

View File

@ -4,12 +4,15 @@
use rand::Rand; use rand::Rand;
use rand; use rand;
use std::num::{One, Zero}; use std::num::{One, Zero, Float};
use traits::operations::ApproxEq; use traits::operations::ApproxEq;
use std::mem; use std::mem;
use structs::dvec::{DVec, DVecMulRhs}; use structs::dvec::{DVec, DVecMulRhs};
use traits::operations::{Inv, Transpose, Mean, Cov}; use traits::operations::{Inv, Transpose, Mean, Cov};
use traits::structure::Cast; use traits::structure::Cast;
use traits::geometry::Norm;
use std::cmp::min;
use std::fmt::{Show, Formatter, Result};
#[doc(hidden)] #[doc(hidden)]
mod metal; mod metal;
@ -494,6 +497,62 @@ impl<N: Clone + Num + Cast<f32> + DMatDivRhs<N, DMat<N>> + ToStr > Cov<DMat<N>>
} }
} }
/// QR decomposition using Householder reflections
/// # Arguments
/// * `m` matrix to decompose
pub fn decomp_qr<N: Clone + Num + Float + Show>(m: &DMat<N>) -> (DMat<N>, DMat<N>) {
let rows = m.nrows();
let cols = m.ncols();
assert!(rows >= cols);
let mut q : DMat<N> = DMat::new_identity(rows);
let mut r = m.clone();
let subtract_reflection = |vec: DVec<N>| -> DMat<N> {
// FIXME: we don't handle the complex case here
let mut qk : DMat<N> = DMat::new_identity(rows);
let start = rows - vec.at.len();
for j in range(start, rows) {
for i in range(start, rows) {
unsafe {
let vv = vec.at_fast(i-start)*vec.at_fast(j-start);
let qkij = qk.at_fast(i,j);
qk.set_fast(i, j, qkij - vv - vv);
}
}
}
qk
};
let iterations = min(rows-1, cols);
for ite in range(0u, iterations) {
// we get the ite-th column truncated from its first ite elements,
// this is fast thanks to the matrix being column major
let start= m.offset(ite, ite);
let stop = m.offset(rows, ite);
let mut v = DVec::from_vec(rows - ite, r.mij.slice(start, stop));
let alpha =
if unsafe { v.at_fast(ite) } >= Zero::zero() {
-Norm::norm(&v)
}
else {
Norm::norm(&v)
};
unsafe {
let x = v.at_fast(0);
v.set_fast(0, x - alpha);
}
let _ = v.normalize();
let qk = subtract_reflection(v);
r = qk * r;
q = q * Transpose::transpose_cpy(&qk);
}
(q, r)
}
impl<N: ApproxEq<N>> ApproxEq<N> for DMat<N> { impl<N: ApproxEq<N>> ApproxEq<N> for DMat<N> {
#[inline] #[inline]
fn approx_epsilon(_: Option<DMat<N>>) -> N { fn approx_epsilon(_: Option<DMat<N>>) -> N {
@ -515,6 +574,18 @@ impl<N: ApproxEq<N>> ApproxEq<N> for DMat<N> {
} }
} }
impl<N: Show + Clone> Show for DMat<N> {
fn fmt(&self, form:&mut Formatter) -> Result {
for i in range(0u, self.nrows()) {
for j in range(0u, self.ncols()) {
write!(form.buf, "{} ", self.at(i, j));
}
write!(form.buf, "\n");
}
write!(form.buf, "\n")
}
}
macro_rules! scalar_mul_impl ( macro_rules! scalar_mul_impl (
($n: ident) => ( ($n: ident) => (
impl DMatMulRhs<$n, DMat<$n>> for $n { impl DMatMulRhs<$n, DMat<$n>> for $n {

View File

@ -1,6 +1,7 @@
//! Data structures and implementations. //! Data structures and implementations.
pub use self::dmat::DMat; pub use self::dmat::DMat;
pub use self::dmat::decomp_qr;
pub use self::dvec::DVec; pub use self::dvec::DVec;
pub use self::vec::{Vec0, Vec1, Vec2, Vec3, Vec4, Vec5, Vec6}; pub use self::vec::{Vec0, Vec1, Vec2, Vec3, Vec4, Vec5, Vec6};
pub use self::mat::{Identity, Mat1, Mat2, Mat3, Mat4, Mat5, Mat6}; pub use self::mat::{Identity, Mat1, Mat2, Mat3, Mat4, Mat5, Mat6};

View File

@ -2,6 +2,7 @@ use std::num::{Float, abs};
use rand::random; use rand::random;
use na::{Vec1, Vec3, Mat1, Mat2, Mat3, Mat4, Mat5, Mat6, Rot3, DMat, DVec, Indexable}; use na::{Vec1, Vec3, Mat1, Mat2, Mat3, Mat4, Mat5, Mat6, Rot3, DMat, DVec, Indexable};
use na; use na;
use na::decomp_qr;
macro_rules! test_inv_mat_impl( macro_rules! test_inv_mat_impl(
($t: ty) => ( ($t: ty) => (
@ -206,3 +207,23 @@ fn test_dmat_from_vec() {
assert!(mat1 == mat2); assert!(mat1 == mat2);
} }
#[test]
fn test_decomp_qr() {
let mat = DMat::from_row_vec(
5,
3,
[
4.0, 2.0, 0.60,
4.2, 2.1, 0.59,
3.9, 2.0, 0.58,
4.3, 2.1, 0.62,
4.1, 2.2, 0.63
]
);
let (q, r) = decomp_qr(&mat);
let mat_ = q * r;
assert!(na::approx_eq(&mat_, &mat));
}