Re: [Pysparse-users] Full matrix conversion to spmatrix.ll_mat

[Dominique O. <dom...@gm...>]

Hi Rodrigo,

If you look at the jdsym documentation you will see that the input
matrix (and preconditioners) need not be in linked-list format, or any
other Pysparse format for that matter. The matrix object need only
have the shape, matvec and matvec_transp attributes. If you can
generate irow, jcol and val arrays efficiently, then it is easy to
wrap those into a simple sparse matrix class that provides the
required attributes. What kind of preconditioner are you using?

Regards,
Dominique

On Thu, Nov 19, 2009 at 3:44 AM, iza...@t-...
<iza...@t-...> wrote:
>
> Hello Dominique,
>
> Thank you very much for answering. Conversion to scipy sparse formats is not
> as expensive as using put. I can generate the three vectors irow, icol, val
> in a short time. The problem is to put the values in the pysparse ll format.
> I can even generate a scipy linked list format efficiently, but the problem
> is this is not the same as the pysparse linked list format and I can not use
> it with jdsym.
>
> The option to use put is unfortunately very expensive ( I have tried it).  I
> will try to generate the pysparse matrix directly but I am afraid all the
> operations I used are not supported. For example I do a very expensive
> double loop operation in C with Pyrex and I am able to generate a n*n vector
> that I reshape inside python to do further processing in matrix form. Can I
> reshape a vector into pysparse matrix format?
>
> Best regards
>
> Rodrigo
>
> --
> Hi Rodrigo,
>
> Thanks for using PySparse! All comments can help us improve the
> library. A (rough) design decision in PySparse is that matrix
> operations should be "cheap", i.e., O(nnz), for matrices that are
> indeed sparse. If your matrix is dense, I'm afraid any constructor of
> the form spmatrix.ll_mat(K) will require O(n^2) operations. Please
> correct me if I'm wrong but the Matrix Market format lists all the
> nonzero elements of your matrix, and we *do* need all that information
> to construct the matrix. I don't see how to lower this cost. Using a
> different sparse format won't help either.
>
> What could help is detect any exploitable structure in your matrix.
> Yes it is dense but it probably isn't random and may have some pattern
> to it. For instance, it could be (anti-)symmetric, (block-)circulant,
> (block-)Toeplitz, or whatever. I feel that is where savings might be
> found.
>
> Alternatively, is there any chance to bypass the 2-dimensional array
> that you process before building the PySparse matrix, construct the
> PySparse matrix directly and operate on it instead? Or operate on
> arrays of the form (irow,jcol,val) and then use put()?
>
> I hope this helps. Good luck.
>
> --
> Dominique
>
> ---Original-Nachricht-----
> Subject: Re: [Pysparse-users] Full matrix conversion to spmatrix.ll_mat
> Date: Wed, 18 Nov 2009 23:30:24 +0100
> From: Dominique Orban <dom...@gm...>
> To: "iza...@t-..." <iza...@t-...>
>
> On Wed, Nov 18, 2009 at 11:47 AM, iza...@t-...
> <iza...@t-...> wrote:
>> Dear Sirs,
>>
>> Im using pysparse with the eigenvalue solver jdsym. It works great !!.
>> The
>> only problem I have at the moment is the conversion time from the original
>> format of my matrix. Here is a description of my process:
>>
>> - Read a file wit around 20000 points
>> - Process this points and get a very big array (20000*20000) (this part is
>> done with pyrex)
>> - I reshape the array into a matrix (20000,20000)
>> - I do some math ( matrix transpose, matrix scaling, dot product,
>> transpose,
>> etc.)
>> - I get a matrix from which I would like to obtain a couple of
>> eigenvectors
>>
>> In order to use jdsm I need the matrix in linked list format. I have tried
>> several things:
>>
>> - write the matrix in matrix market format (using scipy) and the read with
>> spmatrix.ll_mat_from_mtx
>>   (it takes very very long)
>> - convert directly to sparse coo format and then into  ll format
>>  (I get an error that this format is not supported for conversion)
>> - use the internal routine put(V,index1,index2)
>> (this is faster but still takes a long time compared with the eigenvalue
>> problem)
>>  - I also use
>>    for ii in range(N):
>>     for jj in range(N):
>>         Al[ii,jj]=K[ii,jj]
>> (this is almost the same as using put)
>>
>>
>> Is there any way to have it faster in the right format? This will improve
>> a
>> lot the all process of obtaining the eigenvalues. If would be great  to
>> have
>> something like:
>>
>> spmatrix.ll_mat(K)
>>
>> where K is a full matrix.
>>
>> I appreciate any help with this problem.
>
>



-- 
Dominique