Mailman 3 July 2007 - SciPy-Dev

nan functions still broken
by Matthew Auger July 23, 2007

July 23, 2007

The stats nanstd() and nanmedian() functions are still broken. I submitted a patch associated with ticket 337, and an updated patch associated with ticket 368 (with an improved nanstd() algorithm, a fix for the gmean() function, some docstrings, and a new function sigmaclip()). Could someone please apply this patch. Thanks! Matt

2 1

[scikits] svm and pyem now available in scikits.learn
by David Cournapeau July 21, 2007

July 21, 2007

Hi, In short: for people who need to use svm (SVM) and pyem (Gaussian Mixture Models) from the scipy.sandbox, they are now available and working in scikits.learn package. - Getting the code : svn co http://svn.scipy.org/svn/scikits/trunk/learn - Importing the toolboxes : they both reside in the scikits.learn.machine namespace, that is "from scipy.sandbox import svm" becomes "from scikits.learn.machine import svm" and so on. Anything which does not work as before is a bug, and should be filled as such on the scikits trac system (http://projects.scipy.org/scipy/scikits/). For the curious, the learn namespace will soon contain some code to load/pre process/manipulate datasets and some basic learner based on the above algoritms, cheers, David

1 0

GSoC weekly report
by dmitrey July 20, 2007

July 20, 2007

hi all, this week I was asked to close some tickets. Unfortunately, I got to know that I should mention number of svn commits to scipy too late, when I had already done the commits. Next time I will wrote the svn commit numbers. So > http://projects.scipy.org/scipy/scipy/ticket/234 > implementing possibility to turn off warnings about failing to solve in leastsq and fsolve done, same way as proposed > http://projects.scipy.org/scipy/scipy/ticket/285 > (bracket parameters) this feature is still under discussion. Most of feedbacks lead to chapter 3 of Alan Isaac proposition: 3. Move the brent `code` into a `Brent` class. The `brent` function (with unchanged signature) remains as a convenient user interface to the `Brent` class, which can be used directly when more refined control is needed. The class will include bracketing parameters as a response to the ticket. > http://projects.scipy.org/scipy/scipy/ticket/296 > connecting tnc 1.3. In progress. > http://projects.scipy.org/scipy/scipy/ticket/377 > fmin_bfgs incorrectly expects a ZeroDivisonError had fixed > http://projects.scipy.org/scipy/scipy/ticket/390 > scipy.optimize.fmin_bfgs fails without warning I mentioned that the user func is incorrect for the solver - it's non-convex and has lots of local minima, so obtaining different results from different x0 is nothing special > http://projects.scipy.org/scipy/scipy/ticket/399 > connecting PSwarm. Since it requires connecting C code to python, Alan agreed to freeze the ticket (I have no chance to do the ticket in required time). > http://projects.scipy.org/scipy/scipy/ticket/416 > the ticket propose to modify something in C code (__minpack.h), something related to transpose. it says "MATRIXC2F transposing the wrong way in optimize.leastsq" here's the defenition of MATRIXC2F from minpack.h (line 97), it comes w/o any description & I can't understand what does it do: #define MATRIXC2F(jac,data,n,m) {double *p1=(double *)(jac), *p2, *p3=(double *)(data);\ int i,j;\ for (j=0;j<(m);p3++,j++) \ for (p2=p3,i=0;i<(n);p2+=(m),i++,p1++) \ *p1 = *p2; } AJennings (author of the ticket) proposed to change the line 152 MATRIXC2F(fjac, result_array->data, *n, *ldfjac) to MATRIXC2F(fjac, result_array->data, *ldfjac, *n) However, line 94 is the same as line 152, so, maybe, it also requires the replacement. Unfortunately, lots of other c-funcs are also undocumented, so I guess it will take a significant time to fix the ticket. > http://projects.scipy.org/scipy/scipy/ticket/344 ValueError: objects are not aligned in optimize.fmin_cg The problem here was in dot([[a,b],[c,d]], [e,f]) statement (a,b,c, etc are 1x1 ndarrays). It formed 2x2x1 matrix (at least array(...).shape yields the answer (2,2,1)), so multiplying it with [c,d] yields error. I found nothing better than arr = empty(2,2) arr[0,0]=a arr[0,1]=... So now it works. > http://projects.scipy.org/scipy/scipy/ticket/384 > "tnc requires Python list, not numpy.ndarray". I think it's better to connect tnc 1.3 before (i.e. close another ticket before doing this one) > http://projects.scipy.org/scipy/scipy/ticket/389 > initial guess to minimum is overwritten by the minimum So I just set copy(x0) instead of x0 in cobyla.py, now it works. > http://projects.scipy.org/scipy/scipy/ticket/388 > remove uses of apply in optimize.py done, apply(func, ...) is replaced by func(*(...)) Regards, D.

1 0

change scipy.optimize function signatures?? FEEDBACK NEEDED!! (ticket 285)
by Alan G Isaac July 20, 2007

July 20, 2007

Dmitrey has raised some questions about ticket 285: http://projects.scipy.org/scipy/scipy/ticket/285 The changes proposed are not complex to implement, but they increase the interface complexity. They also raise some design considerations. For example, according to the ticket, the signature for brent would change from def brent(func, args=(), brack=None, tol=1.48e-8, full_output=0, maxiter=500): to def brent(func, args=(), brack=None, tol=1.48e-8, full_output=0, maxiter=500, bracket_keywords = {} ): In my personal view, this signature is already messy. If it is to grow an argument, I suggest that the argument be a parameters object. This would work something like the following. (No deep thought on this yet.) class OptizationParams: def __init__(self, **kwargs): self.set_defaults() for k,v in kwargs.iteritems(): if hasattr(self, k): setattr(self, k, v) else: raise AttributeError def set_defaults(self): self.args = () self.tol = 1.48e-8 self.full_output=0 self.maxiter=500 self.bracket_interval = None self.bracket_grow_limit = None The signature for brent would become def brent(func, params=OptizationParams(), **kwargs): Any kwargs provided to `brent` would override the values in the OptParam instance. This should be fully backwards compatible and, looking forward, very flexible. Comments? Cheers, Alan Isaac

5 11

Where to add special functions ?
by David Cournapeau July 19, 2007

July 19, 2007

Hi, I would like to add a special function (implemented in python), but not sure where I should put it for a patch: it looks like scipy.special is implemented in C/Fortran except for orthgonal function. The function I would like to add is the generalized gamma: http://planetmath.org/encyclopedia/GammaFunctionMultivariateReal.html The function is useful to multi variate analysis (appears for example when one needs to compute the KL divergence of Wishart distributions, which appear quite often in Bayesian statistics). cheers, David

2 1

Scikits to be read-only 10am Central (7/19)
by Jarrod Millman July 19, 2007

July 19, 2007

Hello everyone, As part of David Cournapeau's SoC machine learning project, we will be moving a few packages out of scipy.sanbox to scikits.learn. As previously discussed on the list, the specific packages are: scipy.sandbox.ann scipy.sandbox.ga scipy.sandbox.pyem scipy.sandbox.svm I am going to try to make this as painless for everyone as possible, so we are going to make this transition in a few steps. The first step is that tomorrow, I am going to copy all 4 packages (with their histories) over to the machine learning scikit. To make sure I don't mess anything up, the scikit repository will be read-only starting at 10am central time. The changes won't take very long, so by 11am central the scikits repository will be read-write again. I will send out an email before and after the changes. Thanks, -- Jarrod Millman Computational Infrastructure for Research Labs 10 Giannini Hall, UC Berkeley phone: 510.643.4014 http://cirl.berkeley.edu/

1 1

Re: [SciPy-dev] [SciPy-user] changes in scipy.optimization.tnc (v 1.2 to 1.3)
by Jarrod Millman July 19, 2007

July 19, 2007

Hey Dmitrey, (I moved this discussion over to the developer's list.) Dmitrey is working on upgrading scipy.optimization.tnc from v1.2 to v1.3: http://projects.scipy.org/scipy/scipy/ticket/296 I started working on this awhile ago and never finished it. The code I worked on is in scipy.sandbox.newoptimize: http://projects.scipy.org/scipy/scipy/changeset/2530 Dmitrey has pointed out that the API has changed between 1.2 and 1.3; however, I don't think it will make much of a difference to anyone who is using scipy.optimize.tnc anyway. I was trying to change the external API of the tnc.py that Jean-Sebastien provides to match the scipy.optimize interface. Travis did the same thing when he originally added tnc version 1.2. I started making the changes in revision 2535: http://projects.scipy.org/scipy/scipy/changeset/2535 I assume this is still what we want to do. Is there something I am missing? Was there some specific change in the tnc v1.3 API that make it impossible to provide the same interface? I assume everyone using scipy should be using scipy.optimize.tnc.fmin_tnc if they want to use the tnc solver. And I think I was able to get the arguments to fmin_tnc to be same as before. Also make sure to double-check what I did in the sandbox.newoptimize code, because I was basing it off old code. There have been some changes since I worked on it, for instance: http://projects.scipy.org/scipy/scipy/ticket/423 Thanks, -- Jarrod Millman Computational Infrastructure for Research Labs 10 Giannini Hall, UC Berkeley phone: 510.643.4014 http://cirl.berkeley.edu/

1 0

Re: [SciPy-dev] ticket 390 ("scipy.optimize.fmin_bfgs fails without warning", Reported by: ycopin)
by Alan G Isaac July 18, 2007

July 18, 2007

On Wed, 18 Jul 2007, Yannick Copin apparently wrote: > I actually did not check formally if the solution found is > indeed a local minimum, but that would surprise me. So > indeed the problem is not so much that the algo could fall > on some secondary minima, but that it claims convergence > while the requirements for convergence are probably not > met (are they in this special case?) Hmm, I am a bit confused. This ticket suggested a convergence failure, but you are not sure there was one? What is your evidence of a failure? Could you please add a note to the ticket as well as explaining here. We cannot pursue this without a good reason to think there was a failure. Thank you for your help, Alan Isaac

2 1

ticket 344: bugfix commited to svn
by dmitrey July 18, 2007

July 18, 2007

hi all, I committed bugfix to ticket 344, as well as ticket 234, to svn so now it should be closed by someone. P.S. (to Alan Isaac) Yes, I meant exactly the same you have mentioned about tnc.

3 6

ticket 390 ("scipy.optimize.fmin_bfgs fails without warning", Reported by: ycopin)
by dmitrey July 18, 2007

July 18, 2007

hi all, the ticket 390 http://projects.scipy.org/scipy/scipy/ticket/390 says "fmin_bfgs fails without warning" however, I guess the user just had chose inappropriate solver. fmin_bfgs is intended for local minimum, convex funcs. the user's objfunc is chi2 (chi-square distribution estimation) from the func def model(pars, x): a,b,c = pars return a*S.sin(x/b)*S.exp(x/c) # M so a,b,c are variables to be estimated (i.e. required solution from fmin_bfgs) So, since user has sin(const/b), the func is 1) non-convex 2) has lots of local minimum So I guess it's nothing special that sometimes fmin_bfgs yields other solution (from other start points) than the one the user mentioned. So I think the ticket should be closed. However, I don't know how to close tickets. D.

3 4