[Developers] Gaussian mixture model with mixed effects for the mean aond proportions

[Mark Maunder]

Anyone developed a Gaussian mixture model (e.g. several normal distributions representing the length distribution for each age) in ADMB that has a linear mixed effect for the mean (e.g. the mean length at age) and a linear mixed effect for the proportion that each Gaussian makes of the total (e.g. the proportion that age makes of the total population)?

I would like to know if this is computationally feasible in ADMB (32,000 data points, 2500 realizations of one of the random effects, 240 realizations for another random effect) or if TMB would be better.

Thanks,

Mark

From: developers-bounces at admb-project.org [mailto:developers-bounces at admb-project.org] On Behalf Of Jim Ianelli - NOAA Federal
Sent: Monday, August 04, 2014 2:47 PM
To: dave fournier
Cc: developers at admb-project.org
Subject: Re: [Developers] interfacing with openblas for large problems

Seems like a worthwhile enterprise.

Looking at the site there are a number of interesting aspects of OpenBlas that could help with ideas for the Foundation, not least of which is their balance sheet and funding model:

https://docs.google.com/spreadsheet/ccc?key=0AghkTjXe2lDndE1UZml0dGpaUzJmZGhvenBZd1F2R1E&usp=sharing#gid=0

also hadn't heard of:
https://www.bountysource.com/faq
before, looks like something of interest.

also an aside but a useful git-branching model...: http://nvie.com/files/Git-branching-model.pdf

so back to the question, how can we as a community help with the interface work? I suppose steps might be something like:

  1.  pick some example function sets
  2.  map the variable types appropriately (so they will work with ADMB
  3.  write adjoint code for dvar functions (?)
  4.  test against current ADMB implmentations
  5.  ...
thanks for the interest dave.

jim


On Mon, Aug 4, 2014 at 12:31 PM, dave fournier <davef at otter-rsch.com<mailto:davef at otter-rsch.com>> wrote:
I wonder if there is any interest in doing a bit of work to interface some of the common
function like ludecomp and choleski decomp with the openblas versions.

this is not too difficult once one figures out the boring stuff like the fact that matrices are strored
by column in contiguous arrays.

we are now looking at ever larger problems.  I compared the admb lu decomp of a 2,000x2,000 matrix
from numerical recipes with the openblas version.

Time for openblas               lu decomp 172.05 ms
Time for numerical recipes lu decomp 22102.9     ms

which is about 125 times faster
_______________________________________________
Developers mailing list
Developers at admb-project.org<mailto:Developers at admb-project.org>
http://lists.admb-project.org/mailman/listinfo/developers



--

James Ianelli
NMFS/NOAA Building 4
7600 Sand Pt Way NE
Seattle WA 98115

 206 526 6510

Visit the ADMB project http://admb-project.org/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20140804/3089a6ab/attachment-0001.html>