Mailman 3 August 2007 - SciPy-Dev

Bandwidth selection
by Pierre GM Aug. 15, 2007

Aug. 15, 2007

All, I'm slowly cleaning some code I had written a few months ago, porting SiZer to python. <educational break> SiZer, or Significant Zero crossing, is a graphical tool to estimate what features in a distribution are significant. It is based on gaussian kernel smoothing at different bandwidths: depending on the first derivative of the smooth, features are judged significant or not at a given confidence level. </educational break> The initial code was written in Matlab w/ pieces of Fortran, and I received the permission from the main author to port it to scipy and release it as BSD. The package introduces some gaussian-kernel smoothing functions, as well as some automatic bandwidth selectors. In a first step, I want to limit myself to 1D datasets, w/ gaussian kernels. * What has already been done in scipy concerning bandwith selectors ? There's a scipy.stats.kde, but the documentation is scarce... * What would be the best way to release the code ? Scikit ? Thanks a lot in advance for your inputs P.

1 0

SLSQP optimizer
by Rob Falck Aug. 15, 2007

Aug. 15, 2007

Hello, I've been tinkering with the SLSQP (Sequential Least SQuares Programming) optimizer on netlib, and have had success in wrapping it using f2py. I'd like to put this optimizer in Scipy because it offers bounds on the independent variables as well as both equality and inequality constraints. My first concern is, is the license compatible with SciPy? SLSQP is apparently under the ACM license. The following URL's show the Fortran source and some documentation: - http://www.netlib.org/toms/733 - http://portal.acm.org/citation.cfm?doid=192115.192124 - Rob Falck

3 4

Re: [SciPy-dev] about numpy.max() and numpy.min() - isn't it a bug?
by Dmitrey Kroshko Aug. 14, 2007

Aug. 14, 2007

I use the min and max from numpy (from numpy import *) Ok, maybe they are same as python buildins But ordinary users usually use min and max, not amin and amax, so for to prevent the bug mentioned either they somehow should be warned, or the behavior of the numpy.min and max should be set to always produce numpy.array (for more safety). Also, I guess it's possible to contact python developers that wrote python min and max and ask them to yield numpy.array (when any of min/max args is numpy.array) instead of python type. Also, I think the behavior of the +/- operators (to yield python type) is not optimal. Suppose I have x = (4.4*t+myNumpyArr).tolist() So I should care does myNumpyArr is of size > 1 or not. Of course, there are funcs like atleast1d or something like that but users begin to use that ones only when some time have been elapsed to find and fix the bug. Same to amax/amin vs max/min. //my 2 cents, D. Darren Dale wrote: On Tuesday 14 August 2007 11:41:51 am Dmitrey Kroshko wrote: Hi all, as for me I think the behavior of the func is not optimal from numpy import * max(array(2.), 0) array(2.0) max(array(-2.), 0) 0 You are not using numpy's amax and amin, but the python builtins. import numpy numpy.amin(numpy.array(-2.), 0) -2.0

2 1

scipy svn fails to build on Debian
by Ondrej Certik Aug. 14, 2007

Aug. 14, 2007

Hi, I have a fresh Debian unstable and scipy svn fails to build: $ ./setup.py build [...] adding 'build/src.linux-i686-2.4/fortranobject.c' to sources. adding 'build/src.linux-i686-2.4' to include_dirs. building extension "scipy.linsolve._zsuperlu" sources building extension "scipy.linsolve._dsuperlu" sources building extension "scipy.linsolve._csuperlu" sources building extension "scipy.linsolve._ssuperlu" sources building extension "scipy.linsolve.umfpack.__umfpack" sources creating build/src.linux-i686-2.4/scipy/linsolve creating build/src.linux-i686-2.4/scipy/linsolve/umfpack adding 'Lib/linsolve/umfpack/umfpack.i' to sources. creating build/src.linux-i686-2.4/Lib/linsolve creating build/src.linux-i686-2.4/Lib/linsolve/umfpack swig: Lib/linsolve/umfpack/umfpack.i swig -python -I/usr/include -o build/src.linux-i686-2.4/Lib/linsolve/umfpack/_umfpack_wrap.c -outdir build/src.linux-i686-2.4/Lib/linsolve/umfpack Lib/linsolve/umfpack/umfpack.i Lib/linsolve/umfpack/umfpack.i:192: Error: Unable to find 'umfpack.h' Lib/linsolve/umfpack/umfpack.i:193: Error: Unable to find 'umfpack_solve.h' Lib/linsolve/umfpack/umfpack.i:194: Error: Unable to find 'umfpack_defaults.h' Lib/linsolve/umfpack/umfpack.i:195: Error: Unable to find 'umfpack_triplet_to_col.h' Lib/linsolve/umfpack/umfpack.i:196: Error: Unable to find 'umfpack_col_to_triplet.h' Lib/linsolve/umfpack/umfpack.i:197: Error: Unable to find 'umfpack_transpose.h' Lib/linsolve/umfpack/umfpack.i:198: Error: Unable to find 'umfpack_scale.h' Lib/linsolve/umfpack/umfpack.i:200: Error: Unable to find 'umfpack_report_symbolic.h' Lib/linsolve/umfpack/umfpack.i:201: Error: Unable to find 'umfpack_report_numeric.h' Lib/linsolve/umfpack/umfpack.i:202: Error: Unable to find 'umfpack_report_info.h' Lib/linsolve/umfpack/umfpack.i:203: Error: Unable to find 'umfpack_report_control.h' Lib/linsolve/umfpack/umfpack.i:215: Error: Unable to find 'umfpack_symbolic.h' Lib/linsolve/umfpack/umfpack.i:216: Error: Unable to find 'umfpack_numeric.h' Lib/linsolve/umfpack/umfpack.i:225: Error: Unable to find 'umfpack_free_symbolic.h' Lib/linsolve/umfpack/umfpack.i:226: Error: Unable to find 'umfpack_free_numeric.h' Lib/linsolve/umfpack/umfpack.i:248: Error: Unable to find 'umfpack_get_lunz.h' Lib/linsolve/umfpack/umfpack.i:272: Error: Unable to find 'umfpack_get_numeric.h' error: command 'swig' failed with exit status 1 Ondrej

1 1

about numpy.max() and numpy.min() - isn't it a bug?
by Dmitrey Kroshko Aug. 14, 2007

Aug. 14, 2007

Hi all, as for me I think the behavior of the func is not optimal >>> from numpy import * >>> max(array(2.), 0) array(2.0) >>> max(array(-2.), 0) 0 So suppose usually I have x values > 0 and 1 time from 100000 - less than zero. Then sometimes I just get the error "int object has no attribute T" (provided I have code r = max(x,0).T, also, here could be other funcs - max(x,0).tolist(), max(x,0).sum() etc) the same bug if I have 0. (float) instead of 0. Maybe, the same should be fixed for numpy.min() : >>> min(array(-2.), 0) array(-2.0) >>> min(array(2.), 0) 0 Regards, D.

2 1

OS X universal SciPy: success!
by Josh Marshall Aug. 14, 2007

Aug. 14, 2007

I've been trying to easily build an OS X universal SciPy for some time now. [1] My prior efforts consisted of lipo-ing together the PPC and x86 'Superpacks' put together by Chris Fonnesbeck [2], which worked for distributing my image processing app locally. However, this isn't really a good way to do it, specially not for putting up for general use. I came across a successful Universal build of gfortran [4,5], and then shortly found this message [5] claiming it could possibly be used to build a universal SciPy. So, I gave it a shot and it works! (with some tricks...) The build has both ppc and i386 architectures in every .so in the scipy install. The tests run fine on my G4, but I haven't yet had the chance to try it on an Intel Mac. If anyone is keen to do so, please let me know. There will need to be some modifications to numpy distutils, since it presumes that there isn't a universal Fortran. The patch below is just a hack to get it working, and will break any non-universal gfortran. Essentially, all that needs to happen is to have '-arch ppc -arch i386' added to any call to gfortran (both compile and link) , and the '-march' flags removed. What's the best way to add this functionality to numpy distutils? I couldn't think of any way to test for a universal compiler, other than trying to compile a test file and seeing if it dies with the multiple arch flags. Regards, Josh Marshall [1] http://mail.python.org/pipermail/pythonmac-sig/2006-December/ 018556.html [2] http://trichech.us/?page_id=5 [3] http://r.research.att.com/tools/ [4] http://r.research.att.com/gfortran-4.2.1.dmg [5] http://mail.python.org/pipermail/pythonmac-sig/2007-May/018975.html isengard:~/Development/Python/numpy-svn/numpy/distutils/fcompiler Josh $ svn diff gnu.py Index: gnu.py =================================================================== --- gnu.py (revision 3964) +++ gnu.py (working copy) @@ -102,6 +102,7 @@ minor) opt.extend(['-undefined', 'dynamic_lookup', '-bundle']) + opt.extend(['-arch ppc -arch i386']) else: opt.append("-shared") if sys.platform.startswith('sunos'): @@ -183,12 +184,13 @@ # Since Apple doesn't distribute a GNU Fortran compiler, we # can't add -arch ppc or -arch i386, as only their version # of the GNU compilers accepts those. - for a in '601 602 603 603e 604 604e 620 630 740 7400 7450 750'\ - '403 505 801 821 823 860'.split(): - if getattr(cpu,'is_ppc%s'%a)(): - opt.append('-mcpu='+a) - opt.append('-mtune='+a) - break + #for a in '601 602 603 603e 604 604e 620 630 740 7400 7450 750'\ + # '403 505 801 821 823 860'.split(): + # if getattr(cpu,'is_ppc%s'%a)(): + # opt.append('-mcpu='+a) + # opt.append('-mtune='+a) + # break + opt.append('-arch ppc -arch i386') return opt

1 0

problems with login/register at scikits page
by Dmitrey Kroshko Aug. 14, 2007

Aug. 14, 2007

hi all, I constantly get problems with login/register at scikits page https://projects.scipy.org/scipy/scikits/register My old registered login names don't work. If I try to register a new one - after I press "register account" I am asked (in popup window) to enter login name and password and those (just created 5 sec ago, as well as old ones) are not working. if it's important I'm using Mozilla Firefox 2.0.0.6 (& Ubuntu 7.04) (but I had encountered same problems with earlier Mozilla versions as well). Can anyone fix the problem? Regards, D.

2 1

Re: [SciPy-dev] Fwd: from GSoC student Dmitrey Kroshko
by Dmitrey Kroshko Aug. 13, 2007

Aug. 13, 2007

Hi Matthieu, I just implemented very simple algorithm for line-search subproblem, described in the book B. Pshenichniy "Linearisation method", page 46, and lincher is almost the same as the NLP solver from same page 46. You can check that one in lincher.py, line 296 (as you see, it didn't use gradient info vs yours). But since the book is dated 1983, I guess there are better algorithms for now. I will be happy if you'll provide any one (that can be used with non-smooth funcs). So I have updated svn, and if you are interested here's results of mine and yours funcs: #my Armijo implementation itn 66 : Fk= 85.4300044557 maxResidual= 8.77891444873e-07 istop: 4 (|| F[k] - F[k-1] || < funtol) Solver: Time elapsed = 7.41 CPU Time Elapsed = 7.28 objFunValue: 85.4300044557 (feasible) #Matthieu: state = {'direction' : direction, 'gradient': lsF.gradient(x0)} mylinesearch = line_search.StrongWolfePowellRule(sigma=5) destination = mylinesearch(function = lsF, origin = x0, step = direction, state = state) itn 78 : Fk= 85.4300278074 maxResidual= 6.97178904829e-07 istop: 4 (|| F[k] - F[k-1] || < funtol) Solver: Time elapsed = 8.58 CPU Time Elapsed = 8.46 objFunValue: 85.4300278074 (feasible) if I use line_search.StrongWolfePowellRule() (i.e. with default param sigma) it yields (the example from lincher.py head) itn 0: Fk= 1428.11851019 maxResidual= 2242631.78131 itn 10 : Fk= 86.6072664467 maxResidual= 0.466521056114 N= 5336.61507377 and then CPU hang up (at least, I didn't observe anything till ~2 min and then stopped; my iprint = 10). Matthieu Brucher wrote: Hi, Some precisions :) this week I wrote line-search optimizer based on Armijo rule. It seems not to be committed yet. I've coded it myself thursday or friday : http://projects.scipy.org/scipy/scikits/browser/trunk/openopt/scikits/openo… Too sad dmitrey did not code it with my framework. It's very simple, but now lincher didn't require other line-search optimizers from other software - I mean Matthieu's one (unfortunately it very often makes CPU hang up) Precision here so that there are no misunderstanding : the one dmitrey is talking anout is one of many, it's the one with Strong Wolfe-Powell rule, and dmitrey's cost function is not smooth thus the line search has trouble dealing with it. For those who are intersted : http://projects.scipy.org/scipy/scikits/wiki/Optimization Matthieu _______________________________________________ Scipy-dev mailing list Scipy-dev(a)scipy.org http://projects.scipy.org/mailman/listinfo/scipy-dev

2 1

Using C++ in non optional code ?
by David Cournapeau Aug. 13, 2007

Aug. 13, 2007

Hi, I wanted to now if the possibility to use C++ code in scipy was discussed before ? The only C++ code lying around are for optional packages. Several people have expressed the wish to use C++, and I would like to know if there is any official policy regarding its usage. cheers, David

3 3

Fwd: from GSoC student Dmitrey Kroshko
by Alan G Isaac Aug. 13, 2007

Aug. 13, 2007

Dmitrey's ISP in the Ukraine is experiencing problems, so he asked from someone else's computer that I post this brief weekly update. He'll be back in communication as soon as possible. Cheers, Alan Isaac ------ Forwarded message ------ Date: Sat, 11 Aug 2007 12:58:30 +0400 Subject: from GSoC student Dmitrey Kroshko this week I wrote line-search optimizer based on Armijo rule. It's very simple, but now lincher didn't require other line-search optimizers from other software - I mean Matthieu's one (unfortunately it very often makes CPU hang up), scipy.optimize.line_search or it's fortran version (they can't solve the problem as well). Maybe, these solvers just can't handle non-smooth func like f + max(constraints) is. On the other hand, I want to use these line-search solvers when no constraints are present (however, UC solvers like bfgs etc usually will be better in the case), or if the user's problem has just 1 constraint (in the case max(constraint)=constraint=smooth func). Also, I made some changes to docstrings (both openopt and scipy.optimize), but it still requires more work. -------- End of message -------

2 1