Dear all, I would just like to note that we also develop SVM implementation. It is intended for large-scale datasets and makes use of PETSc parallel linear algebra. Currently, it supports only linear kernels - Hessian is, in fact, MATNORMAL with arbitrary underlying data matrix - it is, e.g. possible to use MATDENSE or MATAIJ depending on the problem. For the solution of the arising quadratic program (QP), it uses solvers from our PermonQP package. Both PermonSVM and PermonQP are libraries depending on PETSc. They are written in the PETSc coding style, pretty much like SLEPc. http://permon.it4i.cz/permonqp.htm http://permon.it4i.cz/permonsvm.htm https://github.com/it4innovations/permon https://github.com/it4innovations/permonsvm So far, PermonQP only implements an Augmented Lagrangian type algorithm which can be combined with any solver for box-constrained QP. In PermonQP, there are some concrete ones and also TAO wrapper. However, adding an Interior Point implementation is interesting for us as well. PermonSVM is so far a proof-of-concept thing, but it already scales pretty well (almost proportionally to the application of the data matrix to a vector). See, e.g. our PASC poster https://www.researchgate.net/publication/318317204_PERMON_PASC17_Poster We'll be grateful for any feedback on this. Jakub On 22.9.2017 06:06, Richard Tran Mills wrote:
Thanks for sharing this, Barry. I haven't had time to read their paper, but it looks worth a read.
Hong, since many machine-learning or data-mining problems can be cast as linear algebra problems (several examples involving eigenproblems come to mind), I'm guessing that there must be several people using PETSc (with SLEPc, likely) in this this area, but I don't think I've come across any published examples. What have others seen?
Most of the machine learning and data-mining papers I read seem employ sequential algorithms or, at most, algorithms targeted at on-node parallelism only. With available data sets getting as large and easily available as they are, I'm surprised that there isn't more focus on doing things with distributed parallelism. One of my cited papers is on a distributed parallel k-means implementation I worked on some years ago: we didn't do anything especially clever with it, but today it is still one of the *only* parallel clustering publications I've seen.
I'd love to 1) hear about what other machine-learning or data-mining applications using PETSc that others have come across and 2) hear about applications in this area where people aren't using PETSc but it looks like they should!
Cheers, Richard
On Thu, Sep 21, 2017 at 12:51 PM, Zhang, Hong <[email protected] <mailto:[email protected]>> wrote:
Great news! According to their papers, MLSVM works only in serial. I am not sure what is stopping them using PETSc in parallel.
Btw, are there any other cases that use PETSc for machine learning?
Hong (Mr.)
> On Sep 21, 2017, at 1:02 PM, Barry Smith <[email protected] <mailto:[email protected]>> wrote: > > > From: Ilya Safro [email protected] <mailto:[email protected]> > Date: September 17, 2017 > Subject: MLSVM 1.0, Multilevel Support Vector Machines > > We are pleased to announce the release of MLSVM 1.0, a library of fast > multilevel algorithms for training nonlinear support vector machine > models on large-scale datasets. The library is developed as an > extension of PETSc to support, among other applications, the analysis > of datasets in scientific computing. > > Highlights: > - The best quality/performance trade-off is achieved with algebraic > multigrid coarsening > - Tested on academic, industrial, and healthcare datasets > - Generates multiple models for each training > - Effective on imbalanced datasets > > Download MLSVM at https://github.com/esadr/mlsvm > > Corresponding paper: Sadrfaridpour, Razzaghi and Safro "Engineering > multilevel support vector machines", 2017, > https://arxiv.org/pdf/1707.07657.pdf <https://arxiv.org/pdf/1707.07657.pdf> >