numpy-discussion

[Numpy-discussion] Sparse matrices in NumPy?

[Paul G. <get...@mi...>]

I have a problem where I need to do matrix multiplication of a 7731x7731
matrix; storing this thing would take over 250 MB of memory, which is more
than my machine has. :(  However, the matrix is empty except for the main
diagonal.  Ideally, all that needs to be stored is a single vector 7731
elements long, and then tweak matrix multiplication algorithms to account for
this.  Are there any facilities in NumPy to do this sort of thing, or do I
have to roll my own?  Is there a way to effeciently store a very sparse
matrix and do standard matrix multiplies?  Thanks.

	-Paul Gettings
	 Dep't of Geology & Geophysics
	 University of Utah
--
  But Your Honor, they needed killin'.

Re: [Numpy-discussion] Sparse matrices in NumPy?

[Herbert L. R. <roi...@ha...>]

Unless I am misunderstanding something, you have a vector, which is the
diagonal of a square matrix, and thus contains 7731 elements.  You want to
multiply it by another vector, which has 7731 columns.  Element by element
by element multiplication will, I think, give you the result that you want.

In matlab you would use the diag command and then use the .* operator to get
element by element multiplication.

In NumPY you would simply use the * operator to multiply the two vectors.

Travis Oliphant has sparse matrix classes that you might want to check out,
but I don't see why you would need them for this task.

HLR

----- Original Message -----
From: "Paul Gettings" <get...@mi...>
To: <num...@li...>
Sent: Monday, June 05, 2000 6:49 AM
Subject: [Numpy-discussion] Sparse matrices in NumPy?


> I have a problem where I need to do matrix multiplication of a 7731x7731
> matrix; storing this thing would take over 250 MB of memory, which is more
> than my machine has. :(  However, the matrix is empty except for the main
> diagonal.  Ideally, all that needs to be stored is a single vector 7731
> elements long, and then tweak matrix multiplication algorithms to account
for
> this.  Are there any facilities in NumPy to do this sort of thing, or do I
> have to roll my own?  Is there a way to effeciently store a very sparse
> matrix and do standard matrix multiplies?  Thanks.
>
> -Paul Gettings
> Dep't of Geology & Geophysics
> University of Utah
> --
>   But Your Honor, they needed killin'.
>
>
> _______________________________________________
> Numpy-discussion mailing list
> Num...@li...
> http://lists.sourceforge.net/mailman/listinfo/numpy-discussion
>

Re: [Numpy-discussion] Sparse matrices in NumPy?

[Paul G. <get...@mi...>]

> But memory is so cheap these days! ;-)
I am a grad student, and have no money. :(

>  > However, the matrix is empty except for the main diagonal. 
> Multiplying by a diagonalized matrix is just vectorized multiplication:
> a 0 0 
> 0 b 0   .  <x, y, z> = <az, by, cz>
> 0 0 c
My mistake - I need to multiply the 7731x7731 by a 7731x220 element matrix -
square matrix times rectangular matrix, not just 2 diagonal matrices.
Otherwise, the problem wouldn't be so hard. :)

--
  That which does not kill you, didn't try hard enough.

Re: [Numpy-discussion] Sparse matrices in NumPy?

[Charles G W. <cg...@fn...>]

Paul Gettings writes:
 > 
 > >  > However, the matrix is empty except for the main diagonal. 
 > > Multiplying by a diagonalized matrix is just vectorized multiplication:
 > > a 0 0 
 > > 0 b 0   .  <x, y, z> = <az, by, cz>
 > > 0 0 c
 > My mistake - I need to multiply the 7731x7731 by a 7731x220 element matrix -
 > square matrix times rectangular matrix, not just 2 diagonal matrices.
 > Otherwise, the problem wouldn't be so hard. :)

Still it's not so bad, you just need to break the 7731x220 matrix into
220 vectors of length 7731 and multiply each of them by the diagonal
"matrix", one at a time, and glue the results back together.  The
7731x220 matrix should weigh in at about 6MB, hopefully you have
enough memory for this...

Re: [Numpy-discussion] Sparse matrices in NumPy?

[Jon S. <js...@wm...>]

I don't know if I am missing something, but:

Let's suppose you have A so that A.shape is (7731,). It is diagonal, so,
obviously, you don't need to save it all. Just a vector.

You also have B, shaped like this: (7731,220)

And you want to multiply A*B (the matrix way).

I would dare to say that what you really need is
C=A[:,NewAxis]*B
C will be shaped as (7731,220), which is what you probably need.

Jon Saenz.				| Tfno: +34 946012470
Depto. Fisica Aplicada II               | Fax:  +34 944648500
Facultad de Ciencias.   \\ Universidad del Pais Vasco \\
Apdo. 644   \\ 48080 - Bilbao  \\ SPAIN

On Mon, 5 Jun 2000, Paul Gettings wrote:

> > But memory is so cheap these days! ;-)
> I am a grad student, and have no money. :(
> 
> >  > However, the matrix is empty except for the main diagonal. 
> > Multiplying by a diagonalized matrix is just vectorized multiplication:
> > a 0 0 
> > 0 b 0   .  <x, y, z> = <az, by, cz>
> > 0 0 c
> My mistake - I need to multiply the 7731x7731 by a 7731x220 element matrix -
> square matrix times rectangular matrix, not just 2 diagonal matrices.
> Otherwise, the problem wouldn't be so hard. :)
> 
> --
>   That which does not kill you, didn't try hard enough.
> 
> 
> _______________________________________________
> Numpy-discussion mailing list
> Num...@li...
> http://lists.sourceforge.net/mailman/listinfo/numpy-discussion
> 

Re: [Numpy-discussion] Sparse matrices in NumPy?

[Paul G. <get...@mi...>]

> I would dare to say that what you really need is
> C=A[:,NewAxis]*B
> C will be shaped as (7731,220), which is what you probably need.
Based on a simple (small) test, this is exactly what I want; the result of
the NewAxis computation is identical to a matrix multiply of the full-sized
matrices.  Thanks.  Now I just have to read the docs until I understand
precisely how and why this works. :)

--
  101 USES FOR A DEAD MICROPROCESSOR
        (62)  Fungus trellis