Re: [eigen] Student contribution proposal

[ Thread Index | Date Index | More Archives ]

On Wed, Mar 20, 2013 at 8:49 AM, Gael Guennebaud <gael.guennebaud@xxxxxxxxx> wrote:
Hi Simon,

I always welcome contributions :) especially if they are consolidating
existing modules.

I'm afraid the Todo-list, it is pretty out-of-date. Regarding the SVD
module, there is still a need for a faster SVD than the current
two-sided Jacobi algorithm. There are many discussions on this topic
in this mailing-list. I could gather them for you. In very short, the
divide-and-conquer algorithm seems to be the best approach.

+1, a good implementation of e.g. HouseholderBidiagonalizedSVD or other fast SVD would be  nice.
The QR module is pretty complete with three versions of Householder QR
(no, column, and full pivoting). The column-pivoting version still
need to be accelerated by working on blocks of vertical panels.

I'd like to throw in my support for blocking the column-pivoting QR implementation; we are using col-piv QR now in Ceres Solver's dense mode.
It would also be nice to have methods to compute the kernel for all
rank-revealing factorizations (SVD, ColPivotingQR, FullPivotingQR) as
already done in FullPivLU.

In libmv (and by extension Blender) computes the kernel of various wide and tall matrices all over the place. Fast and stable methods for this is useful in many computer vision applications, since computer vision tends to give rise to various homogeneous equations.

Currently we use the very stable, but also very slow, JacobiSVD. I would love to switch to something faster for some of these uses.
If you had other ideas in mind, feel free to propose them.

Again going back to what would be useful for Ceres: Currently Ceres depends on CHOLMOD for non-iterative general sparse solving. Depending on CHOLMOD makes Ceres have LGPL dependencies, and that is unacceptable for many applications. Eigen already has a sparse Cholesky factorizer (I think?) but is lacking a clean-room AMD implementation. If sparse solving interests you, we'd love to have that as well.

Let me know if you have further questions,

On Wed, Mar 20, 2013 at 10:36 AM, Simon Courtemanche
<simon.courtemanche@xxxxxxxxx> wrote:
> Hello Eigen,
> I'm teaching at an engineering school of computer science and applied
> mathematics, and I would like to propose to some students to contribute to
> Eigen.
> I was using Eigen for a couple of years, but I've never developed the code.
> Is the Todo list up-to-date ?
> Can I propose a contribution to the SVD module or to the QR modules ?
> Best regards,
> Simon Courtemanche
> ---
> PhD student
> Ensimag, Grenoble INP
> INRIA Rhone-Alpes, France

Mail converted by MHonArc 2.6.19+