dnAnalytics

New Post: Full Pivoting LU Decomposition

cuda — Sat, 04 Jul 2009 06:08:44 GMT

This discussion has been copied to a work item. Click here to go to the work item and continue the discussion.

Created Feature: Full Pivoting LU Decomposition

cuda — Sat, 04 Jul 2009 06:08:43 GMT

I have a suggestion for enhancing this excellent library. Most LU decompositions just handle the case involving row pivoting. However there is also full row and column pivoting which can be used to generate a unique decomposition for non-square matrices from which all sort of useful information can be extracted. The only library I know that has routines to do this is Eigen - http://eigen.tuxfamily.org. LAPACK can only do full pivoting on square matrices. Matlab (and clones) can't do without third-party support (http://www.mathworks.com/matlabcentral/fileexchange/authors/24880) and Mathematica is fairly primitive in this area.

New Post: Full Pivoting LU Decomposition

cuda — Sat, 04 Jul 2009 06:08:21 GMT

Hi,

I'll look in to it, but problably not for a month or so.

Thanks for the suggestion,

Marcus

New Post: Full Pivoting LU Decomposition

hsauro — Fri, 03 Jul 2009 15:03:00 GMT

New Post: Feature Suggestion: Apply

cuda — Thu, 02 Jul 2009 19:09:52 GMT

Hi Jullian,

We plan on putting out the beta the first week of August. Are you having a problem building the library from source?

Thanks,

Marcus

New Post: Feature Suggestion: Apply

jpulman — Thu, 02 Jul 2009 16:55:16 GMT

Hi Marcus,

Would it be possible to compile the dll with this new code in it? I'm looking at the 2009.4 build and don't seem to be able to find the Map functionality. Even if not an official release maybe a Beta?

I'm trying to implement a PointwiseDivide function, and was going to use Apply.

Kindest regards,

Julian

Commented Issue: System.OverflowException in SparseMatrix.GrowStorage()

pvandervelde — Sun, 28 Jun 2009 21:28:50 GMT

There is an issue with the SparseMatrix.GrowStorage() method for large Sparse Matrixes. I’m using a SparseMatrix with Row and Column sizes of ~105000 with a very small fill factor. The GrowStorage method calculates Row*Column to determine maximum array sizes. This calculation throws a System.OverflowException for large Row / Column values although the actual memory usage is way smaller. I changed the code (locally) to use long variables for the size calculations in the GrowStorage method and was able to use the large SparseMatrix without problems.
PS: This is a great library and a joy to use – thank you
Comments: ** Comment from web user: pvandervelde **

Ah yes good one. That will have to be fixed at some point. Thanks for reporting the bug :)

New Post: Weibull distribution

cuda — Thu, 25 Jun 2009 10:26:24 GMT

Hi Arun,

I've created a ticket for the request. Not sure when it will be implemented though.

Regards,

Marcus

New Post: Weibull distribution

cuda — Thu, 25 Jun 2009 10:25:32 GMT

This discussion has been copied to a work item. Click here to go to the work item and continue the discussion.

Created Feature: Weibull distribution

cuda — Thu, 25 Jun 2009 10:25:31 GMT

Kindly add Weibull to dnAnalytics.Statistics.Distributions
Thanks,
Arun

Created Issue: System.OverflowException in SparseMatrix.GrowStorage()

sigristu — Thu, 25 Jun 2009 07:10:10 GMT

New Post: Weibull distribution

arunveeramany — Wed, 24 Jun 2009 00:58:42 GMT

Kindly add Weibull to dnAnalytics.Statistics.Distributions

Thanks,

Arun

Reviewed: 2009.4 (Jun 19, 2009)

jpulman — Fri, 19 Jun 2009 13:52:05 GMT

Rated 5 Stars (out of 5) - Excellent open source library allowing C# programmers to get their hands into the speed of MKL without having to fork out thousands of dollars for commercial solutions. Team response is very rapid, and turnaround regarding suggestions very speedy. This project has the potential to become the de-facto analytics and lin alg lib for C#. Get into that standard now.

Updated Wiki: Home

cuda — Thu, 11 Jun 2009 11:59:08 GMT

News | Documentation | FAQ | Contributing | Build Status

dnAnalytics is an open source numerical library for the .NET Framework and Mono. The library is written in C# and is available as a fully managed library, or with a wrapper around the Intel® Math Kernel Library (MKL). The MKL wrapped version provides significantly better performance when working with large data sets.

dnAnalytics is compatible with .NET 2.0 or later, and Mono. The managed version will run on a Windows XP or newer, and any platform that supports Mono. The MKL wrapped version supports 32bit and 64bit versions of Windows XP or newer, and 32bit and 64bit versions of Linux.

2009.4 Features
* Linear algebra classes with support for sparse matrices and vectors (with a F# friendly interface).
* Dense and sparse solvers.
* Probability distributions.
* Random number generation (including Mersenne Twister MT19937).
* QR, LU, SVD, and Cholesky decomposition classes.
* Matrix IO classes that read and write matrices form/to Matlab, Matrix Market, and delimited files.
* Complex and “special” math routines.
* Descriptive Statistics, Histogram, and Pearson Correlation Coefficient.
* Overload mathematical operators to simplify complex expressions.
* Visual Studio visual debuggers for matrices and vectors
* Runs under Microsoft® Windows and platforms that support Mono.
* Optional support for Intel®Math Kernel Library (Microsoft® Windows and Linux)
Download 2009.4

dnAnalytics Social Networking
Twitter
Ohloh
Linkedin

Source Control
The source code is now hosted in a Subversion repository provided by CVSDude. We also provide a Git mirror at GitHub.

Subversion Repository URL: http://dnanalytics-svn.cvsdude.com/dnanalytics/
Subversion Commit RSS Feed: http://feeds.feedburner.com/dnanalytics_commmits
Git Mirror: http://github.com/cuda/dnanalytics

Commented Issue: Parallel Random Number Generators

2bike4 — Thu, 04 Jun 2009 14:58:17 GMT

When one writes a multithreaded application which uses random number generators, we can do two things:
1) lock the random number generator
2) make a random number generator for each thread

The first option is best from the point of view of getting random samples; but it might come at a serious performance cost. The second option is fast, but has potential issues. When I call new System.Random() on two threads at the same time, they might initialize using the same seed and hence the two threads will generate exactly the same numbers.

It would be good if we could add a feature that addresses this issue. Some references:
http://sprng.cs.fsu.edu/
http://www.netlib.org/utk/lsi/pcwLSI/text/node141.html
Comments: ** Comment from web user: 2bike4 **

Here is the reference to matlab's random number section. They just recently added the ability to have independent random streams which is very needed in parallel work.

http://www.mathworks.com/access/helpdesk/help/techdoc/math/brnuahp.html

New Post: Native version slower than managed one ?

cuda — Sat, 30 May 2009 08:00:57 GMT

Hi,

There is a manged/native transition performance hit. So for small matrices, the performance will be roughly the same (and perhaps slower for very small matrices). I'll do some benchmarking on Monday to see where the crossover point is for SVD.

Regards,

Marcus

New Post: Native version slower than managed one ?

pleasereset — Fri, 29 May 2009 20:54:57 GMT

Oh sorry I forgot to precise that, so the source Matrix M is varying between [8,3] and [160,3], I'd say an average size of [50,3];

The snippet shown above is called 100000 times within the processing time indicated.

New Post: Native version slower than managed one ?

cuda — Fri, 29 May 2009 20:32:37 GMT

Hi,

What are the dimensions of your test matrix?

Thanks,

Marcus

New Post: Native version slower than managed one ?

pleasereset — Fri, 29 May 2009 19:37:14 GMT

Hi folks,

First, thanks to every contributor for this useful library !
I post here because I'm wondering about a performance issue on a project involving dnAnalytics.

To identify the problem, I'm running a loop computing normal at a point in an unordered point cloud 5000 times.
This code mainly rely on this snippet :

            dnAnalytics.LinearAlgebra.Matrix A = M.Transpose().Multiply(M);
            dnAnalytics.LinearAlgebra.Decomposition.Svd SVD = new Svd(A, true);

            // The eigen vector of A corresponding to the smallest eigenvalue of A is the normal of the best fit plane
            double[] EigenValues = SVD.S().ToArray();
            dnAnalytics.LinearAlgebra.Matrix EigenVectors = SVD.U();

            int minIndex = 0;
            double minValue = EigenValues[0];
            for (int i = 1; i < EigenValues.Length; i++)
            {
                if (EigenValues[i] < minValue)
                {
                    minValue = EigenValues[i];
                    minIndex = i;
                }
            }

            pV.X = (float)EigenVectors[0, minIndex];
            pV.Y = (float)EigenVectors[1, minIndex];
            pV.Z = (float)EigenVectors[2, minIndex];

Running on a Core 2 Duo, I obtain the following results :

- Managed version : 48s

- Native version : 51s

I'm really wondering about this poor performance using the native version, even tho intel's library is loaded at runtime.

Created Issue: Includeing a FloatDenseMatrix Class

jkshen — Wed, 27 May 2009 10:54:02 GMT

Create a Dense Matrix using "float" type to save the memory.