Re: [Pysparse-users] ambiguous error using solvers: TypeError: matvec() takes exactly 2 arguments (3 given)

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

On Tue, Jun 14, 2011 at 7:11 AM, Oz Nahum Tiram <na...@gm...> wrote:

> On Tue, Jun 14, 2011 at 5:12 AM, Dominique Orban
> <dom...@gm...> wrote:
> >
> > Hi Oz,
> > I just pushed improved wrappers for the iterative solvers to the
> 'suitesparse' branch of Pysparse (which is meant to become trunk in the near
> future). The wrappers are now more streamlined. For some reason, Bicgstab
> and Gmres were left out when we wrote those wrappers. I just added them. If
> you look in itsolvers_utils.py, you will see an example towards the bottom
> of the file. The wrappers accept both ll_mat and PysparseMatrix input. This
> is the better way to call the iterative solvers. There shouldn't be much of
> a difference in speed.
> > What do you mean by "why is your code not there"?
> >
> > --
> > Dominique
>
> Hi Dominique,
> I looked at the code in itsolvers_utils.py, which uses the classes and
> objects concept.
> It is not unfamiliar  for me to write code like that, but imho it adds
> a layer of complexity I prefer to avoid.
>

Complexity? In my view such classes offer more flexibility (e.g., you can
subclass them) and are designed to make the user's life easier. For
instance, now, they accept both ll_mat and PysparseMatrix input. I also
moved arguments such as tol, itmax and K (the preconditioner) to the solve()
method as that is where they belong. The old-style calling sequence is more
C-like and less Python-like.

In what way are they more complex to you?


> What I meant in my question is from the two ways I used the matrices,
> is there some which gives better performance ?
>

No, there should not be any performance hit when calling the solver. There
may be a slight hit when you construct the matrix, but I would guess it is
negligible and I've never bothered to measure it.

>What do you mean by "why is your code not there"?
> About code in scipy: it is good to have both projects, I just thought
> if you merge your code into scipy, it will reach bigger audiences.
>

As you may know, Pysparse used to be part of Scipy several years ago. For
some reason, the Scipy folks decided to write their own sparse matrix
library.

-- 
Dominique