Re: Memory exhausted when using \

[Mathieu Dubois]

Le 16/07/2012 15:43, Martin Helm a écrit :
> Am 16.07.2012 15:13, schrieb Mathieu Dubois:
>    
> > Le 16/07/2012 15:04, marco atzeri a écrit :
> >      
> > > On 7/16/2012 2:55 PM, Markus Bergholz wrote:
> > >        
> > > > like that?
> > > >
> > > > octave:1>  A=[246 710, 49 260]
> > > > A =
> > > >
> > > >      246   710    49   260
> > > >
> > > > octave:2>  B=[49 260, 1]
> > > > B =
> > > >
> > > >       49   260     1
> > > >
> > > > octave:3>  A \ B
> > > > ans =
> > > >
> > > >      1.8994e-02   1.0079e-01   3.8764e-04
> > > >      5.4820e-02   2.9088e-01   1.1188e-03
> > > >      3.7834e-03   2.0075e-02   7.7212e-05
> > > >      2.0075e-02   1.0652e-01   4.0970e-04
> > > >
> > > > octave:4>
> > > >
> > > > or am i wrong?
> > > > but updating octave and use an up to date version is always a good
> > > > idea!
> > > >
> > > > regards
> > > > markus
> > > >
> > > >
> > > > On Mon, Jul 16, 2012 at 2:43 PM, Mathieu Dubois
> > > >   wrote:
> > > >          
> > > > > Hello,
> > > > >
> > > > > I was porting some code from Matlab to Octave but I'm facing a "memory
> > > > > exhausted" error. I'm using Octave 3.2 on a i386 system (I may have
> > > > > access
> > > > > to more powerful systems but I would like to be able to run the
> > > > > code on this
> > > > > machine).
> > > > >
> > > > > This happens when using "\" to solve a linear system. The matrix A
> > > > > is (246
> > > > > 710, 49 260) and b is (49 260, 1).
> > > > >
> > > > > I have seen that Octave have other solvers for linear system. Do
> > > > > you think
> > > > > that may help?
> > > > > Do you have any idea if upgrading to 3.6 could help?
> > > > >
> > > > > Thanks in advance,
> > > > > Mathieu
> > > > >            
> > > I suspect he means a matrix with dimension 246710 x 49260
> > > a vector with 49260 x 1
> > >
> > > As the matrix is almost 10 Gb no surprise of memory exhausted
> > >
> > > Regards
> > > Marco
> > >        
> > You're right Marco I mean a 246710 x 49260 matrix (I wrote the size of
> > the matrix).
> >
> > The construction of A and b works because they are sparse. It's the
> > division that don't work.
> >      
> Try an iterative solver like pcg and check if it is fast enough it uses
> less memory than the direct solvers (there are other iterative solvers
> as well if pcg does not fit your needs)
> http://www.gnu.org/software/octave/doc/interpreter/Iterative-Techniques.html
>    
Naively replacing:
X = A\b
by:
X = pcg(A, b)
gives:
error: operator *: nonconformant arguments (op1 is 246710x49260, op2 is 
246710x1)
error: called from:
error:   /usr/share/octave/3.2.3/m/sparse/pcg.m at line 262, column 3

Any idea?