pymc-user — PyMC users discussion

[Pymc-user] PyMC user's guide revision

[Christopher F. <fon...@ma...>]

Hi everyone,

I am in the midst of revising the PyMC user's manual in anticipation  
of the official 1.0 release. If any users have suggestions or errata  
to report, now would be an ideal time. I am planning on adding a few  
more examples for the new document, and generally fleshing out some  
of the methodology.

Thanks,
C.

--
Christopher Fonnesbeck
+ Atlanta, GA
+ fonnesbeck at mac.com
+ Contact me on AOL IM using email address

Re: [Pymc-user] poisson for negative values

[Christopher F. <ch...@tr...>]

On May 15, 2006, at 7:20 PM, Seth de l'Isle wrote:

> Hello,
>
> I was expecting a poisson prior to limit a parameter to possitive  
> values
> much like using the exponential prior would.  Instead it halted
> processing.
>
>>>> from PyMC.MCMC import *
>>>> class foo(MetropolisHastings): pass
>     ...
>>>> f = foo()
>>>> f.exponential_prior(-1, 5)
>     nan
>>>> f.poisson_prior(-1, 5)
>     PAUSE  negative factorial in factln statement executed
>     To resume execution, type go.  Other input will terminate the job.
>
>     STOP
>
> Shouldn't poisson_prior and poisson_like return nan as well?

Hi Seth,

Thats a good point. At the moment, the likelihoods only enforce the  
parameters of the distribution, and not the data. I essentially leave  
it up to the user not to pass negative data values. Part of the  
reason for this is to avoid as much checking as possible (i.e. "if"  
statements), as this slows down the model.

C.

--
Christopher J. Fonnesbeck, Ph.D.

Population Ecologist, Marine Mammal Section
Fish & Wildlife Research Institute (FWC)
St. Petersburg, FL

Adjunct Assistant Professor
Warnell School of Forestry and Natural Resources
University of Georgia
Athens, GA

T: 727.235.5570
E: chris at trichech.us

[Pymc-user] poisson for negative values

[Seth de l'I. <sz...@ub...>]

Hello,

I was expecting a poisson prior to limit a parameter to possitive values
much like using the exponential prior would.  Instead it halted
processing.

    >>> from PyMC.MCMC import *
    >>> class foo(MetropolisHastings): pass
    ...
    >>> f = foo()
    >>> f.exponential_prior(-1, 5)
    nan
    >>> f.poisson_prior(-1, 5)
    PAUSE  negative factorial in factln statement executed
    To resume execution, type go.  Other input will terminate the job.

    STOP

Shouldn't poisson_prior and poisson_like return nan as well?

Thanks,
  Seth

-- 
       Listen    and    be     Heard
         http://PopularVoice.Org
Media Reform : Participatory Journalism : Populism 

[Pymc-user] Scaling of parameters

[David H. <dav...@gm...>]

Hi,

I'm using PyMC to sample from the distribution of two parameters, one whose
domain is  about [0,2] and the other [300,600]. Since they are correlated,
I'm block updating the two together, ie. self.parameters('AB', init_val =3D
[1, 500])

It appears that doing this is not efficient, in the sense that the jumps fo=
r
B are of the same order than for A. Hence, it takes a very long time to
converge to the stationary distribution. Taking the log of B has allowed me
to cut the number of simulations dramatically. Is this normal ?

I thought that each parameter would get its own jump variance, with a
correlation coefficient linking the two.

Thanks,

David

Re: [Pymc-user] use TkAgg?

[Christopher F. <ch...@tr...>]

On Apr 21, 2006, at 9:02 AM, Hanno Klemm wrote:
>
> I understand that the parameters to be estimated are defined via the
> sample.parameter method. There I can define a proposed distribution
> via dist='some_dist'. However, 'some_dist' needs to be zero-centered,
> if I understand the comment in the code correctly. What is the best
> way of dealing with a case where you want to have your parameters
> greater-equal than zero? Or am I missing the point completely?

Well, in theory the proposal distribution is arbitrary -- it can  
literally be any distribution over the support of the parameter. In  
practice, however, some distributions work better than others. A  
noncentral proposal distribution is fine, but it may not converge as  
efficiently as a central one. I can't think of a reason a priori that  
you would want to have a nonzero expected value, but perhaps you have  
one.

At the moment, 3 proposal distributions are built-in to PyMC: normal  
(the default), uniform, and double exponential. The double  
exponential proposal is interesting because it will occasionally  
propose extreme values, due to its long tails, which can help avoid  
metastable behaviour.

I plan on releasing, with the official 1.0 version, a more  
comprehensive user's guide that will rather explicitly lay all this out.

Regards,
Chris

--
Christopher J. Fonnesbeck

Population Ecologist, Marine Mammal Section
Fish & Wildlife Research Institute (FWC)
St. Petersburg, FL

Adjunct Assistant Professor
Warnell School of Forest Resources
University of Georgia
Athens, GA

T: 727.235.5570
E: chris at trichech.us

Re: [Pymc-user] use TkAgg?

[Hanno K. <kl...@ph...>]

Hi Christopher,

thanks for the quick reply. 

Another question is bothering me at the moment (this time maybe a
little less technical):

I understand that the parameters to be estimated are defined via the
sample.parameter method. There I can define a proposed distribution
via dist='some_dist'. However, 'some_dist' needs to be zero-centered,
if I understand the comment in the code correctly. What is the best
way of dealing with a case where you want to have your parameters
greater-equal than zero? Or am I missing the point completely?

Best regards,
Hanno



Christopher Fonnesbeck <ch...@tr...> said:

> On Apr 21, 2006, at 5:24 AM, Hanno Klemm wrote:
> 
> > Is there a deeper reason for "hard-wiring" TkAgg usage into PyMC and
> > are there further problems down the road I have to anticipate with my
> > approach?
> 
> Hello Hanno,
> 
> The only reason TkAgg is hard-wired into PyMC is that it was a better  
> default than WxAgg, since Tkinter is easy to install on most systems  
> (unlike WXpython), and even comes built in to some python  
> distributions, like ActiveState. There should be no problem in  
> choosing another backend.
> 
> Good luck,
> Chris
> 
> --
> Christopher J. Fonnesbeck
> 
> Population Ecologist, Marine Mammal Section
> Fish & Wildlife Research Institute (FWC)
> St. Petersburg, FL
> 
> Adjunct Assistant Professor
> Warnell School of Forest Resources
> University of Georgia
> Athens, GA
> 
> T: 727.235.5570
> E: chris at trichech.us
> 
> 
> 



-- 
Hanno Klemm
kl...@ph...

Re: [Pymc-user] use TkAgg?

[Christopher F. <ch...@tr...>]

On Apr 21, 2006, at 5:24 AM, Hanno Klemm wrote:

> Is there a deeper reason for "hard-wiring" TkAgg usage into PyMC and
> are there further problems down the road I have to anticipate with my
> approach?

Hello Hanno,

The only reason TkAgg is hard-wired into PyMC is that it was a better  
default than WxAgg, since Tkinter is easy to install on most systems  
(unlike WXpython), and even comes built in to some python  
distributions, like ActiveState. There should be no problem in  
choosing another backend.

Good luck,
Chris

--
Christopher J. Fonnesbeck

Population Ecologist, Marine Mammal Section
Fish & Wildlife Research Institute (FWC)
St. Petersburg, FL

Adjunct Assistant Professor
Warnell School of Forest Resources
University of Georgia
Athens, GA

T: 727.235.5570
E: chris at trichech.us

[Pymc-user] use TkAgg?

[Hanno K. <kl...@ph...>]

Hi,

I have just started toying around with PyMC. However, I ran directly
in the beginning into a catch: I don't have Tkinter installed for
various reasons.

This throws up an error when importing PyMC, because it wants the
TkAgg backend of matplotlib to be used. However, after changing 

use('TkAgg') 

into 

use('GTK')

in Matplot.py 

and setting interactive=True in matplotlibrc, everything works well.
Is there a deeper reason for "hard-wiring" TkAgg usage into PyMC and
are there further problems down the road I have to anticipate with my
approach?

Thanks,
Hanno


-- 
Hanno Klemm
kl...@ph...

[Pymc-user] PyMC 1.0 call for bugs

[Christopher F. <ch...@tr...>]

Hi all,

I'm preparing the official 1.0 release of PyMC, scheduled hopefully  
for next week. The aim of this release is not to introduce a whole  
crop of new features, but rather, to make the existing feature set  
relatively solid and devoid of bugs. So, if anyone has any bugs to  
report, or usability concerns, either major or minor, now is the time  
to pass them along. This includes any likelihood distributions that  
are missing which you feel should be included.

One thing that will accompany the release is an updated user's guide,  
as well as a step-by-step online tutorial for building models in PyMC.

Thanks,
Chris

--
Christopher J. Fonnesbeck

Population Ecologist, Marine Mammal Section
Fish & Wildlife Research Institute (FWC)
St. Petersburg, FL

Adjunct Assistant Professor
Warnell School of Forest Resources
University of Georgia
Athens, GA

T: 727.235.5570
E: chris at trichech.us

Re: [Pymc-user] Correlation between parameters

[David H. <dav...@gm...>]

Thanks Chris,

Just so I understand correctly, does specyfing p =3D [A,B] means that
A,B are updated using bivariate normal jumps with a correlation
coefficient that is tuned along with the covariances ?


Also, I'm planning to use PyMC to do mixed Gibbs-Metropolis sampling.=20
That is, some parameters would be sampled directly from a distribution
and others would be chosen via MH jumping rules. If I ever get there,
I could write a little how-to for inclusion in the PyMC tutorial.
Would that be of any interest to you ?


Thanks,

David



2006/2/24, Chris Fonnesbeck <fon...@gm...>:
> On 2/15/06, David Huard <dav...@et...> wrote:
> >  I have a model with two parameters A and B. When I compute their
> > distribution using PyMC, I get a 'cigar' distribution, in other words,
> > parameters A and B are strongly correlated. I figure this must put a st=
rain
> > on the routine, forcing it to decrease the jump radius to maintain an
> > acceptable acceptance ratio.  Is there a way to indicate that those
> > parameters are correlated and to use a multivariate distribution for th=
e
> > jumps ?
> >
>
> Sorry for the late reply. The best way to deal with highly correlated
> parameters is to block update -- i.e. sample them together. Thus,
> rather than have A and B, you can just specify a vector-valued
> parameter A with init_val =3D [x, y]. Of course, this assumes A and B
> both use the same likelihood function. Is this the case? I should
> probably add the ability to block update among arbitrary parameters.
>
> >  P.S. Adding the generalized extreme distribution to flib would sure be
> > appreciated.
>
> Look for it in the next release.
>
> C.
>
> --
> Chris Fonnesbeck + Atlanta, GA + http://trichech.us
>
>
> -------------------------------------------------------
> This SF.Net email is sponsored by xPML, a groundbreaking scripting langua=
ge
> that extends applications into web and mobile media. Attend the live webc=
ast
> and join the prime developer group breaking into this new coding territor=
y!
> http://sel.as-us.falkag.net/sel?cmdlnk&kid=110944&bid$1720&dat=121642
> _______________________________________________
> Pymc-user mailing list
> Pym...@li...
> https://lists.sourceforge.net/lists/listinfo/pymc-user
>

Re: [Pymc-user] Correlation between parameters

[Chris F. <fon...@gm...>]

On 2/15/06, David Huard <dav...@et...> wrote:
>  I have a model with two parameters A and B. When I compute their
> distribution using PyMC, I get a 'cigar' distribution, in other words,
> parameters A and B are strongly correlated. I figure this must put a stra=
in
> on the routine, forcing it to decrease the jump radius to maintain an
> acceptable acceptance ratio.  Is there a way to indicate that those
> parameters are correlated and to use a multivariate distribution for the
> jumps ?
>

Sorry for the late reply. The best way to deal with highly correlated
parameters is to block update -- i.e. sample them together. Thus,
rather than have A and B, you can just specify a vector-valued
parameter A with init_val =3D [x, y]. Of course, this assumes A and B
both use the same likelihood function. Is this the case? I should
probably add the ability to block update among arbitrary parameters.

>  P.S. Adding the generalized extreme distribution to flib would sure be
> appreciated.

Look for it in the next release.

C.

--
Chris Fonnesbeck + Atlanta, GA + http://trichech.us

Re: [Pymc-user] generating toy data with PyMC???

[David H. <dav...@gm...>]

I doubt this is the easy way to do what you want. I'd rather build a
model with known statistical distribution and generate random
variables.

For instante :
1. Generate random gender (0 or 1) with a Bernouilli distribution with
parameter p, the proportion of men.
2. Generate age, according to gender. Try to find data for a given
population and simply draw randomly from it.
3. Generate income (dependent on gender and age) This is the tricky
part, there are many distributions that can model correlation,
(bivariate gumbel, bivariate normal, copulas), but you'll have to
specify a model for the correlation anyway. I doubt a linear
correlation would do...
4. Same problem for outcome.

You'll find a wide array of distributions to generate randomly from in
scipy.stats and in the random module.

Cheers,

David



2006/2/24, Hugo Koopmans <hug...@gm...>:
> Hi there,
>
> I have done some experiments with PyMC. Has been working very well so far=
,
> keep up the good work!!!
>
> Now, I want to use PyMC to generate data to do experiments with missing
> values. Therefore I need to generate toy data first.
> This toydata for example could consist of the following variables:
> age, income and gender and an outcome (e.g. change of buying product X)
> now I would like to have an underlying model to generate data from in whi=
ch
> for instance age and income are correlated in some way and females like t=
he
> product more then males. Also the correlation between income and age is
> stronger for females then for males.
>
> Would this be possible using PyMC? Did anyone do something like this? Sam=
ple
> code would be appriciated very much!
>
> Regards,
>
> hugo

[Pymc-user] generating toy data with PyMC???

[Hugo K. <hug...@gm...>]

Hi there,

I have done some experiments with PyMC. Has been working very well so far,
keep up the good work!!!

Now, I want to use PyMC to *generate* data to do experiments with missing
values. Therefore I need to generate toy data first.
This toydata for example could consist of the following variables:
age, income and gender and an outcome (e.g. change of buying product X)
now I would like to have an underlying model to generate data from in which
for instance age and income are correlated in some way and females like the
product more then males. Also the correlation between income and age is
stronger for females then for males.

Would this be possible using PyMC? Did anyone do something like this? Sampl=
e
code would be appriciated very much!

Regards,

hugo

[Pymc-user] Correlation between parameters

[David H. <dav...@et...>]

I have a model with two parameters A and B. When I compute their
distribution using PyMC, I get a 'cigar' distribution, in other words,
parameters A and B are strongly correlated. I figure this must put a
strain on the routine, forcing it to decrease the jump radius to
maintain an acceptable acceptance ratio.  Is there a way to indicate
that those parameters are correlated and to use a multivariate
distribution for the jumps ?


Thanks, 


David


P.S. Adding the generalized extreme distribution to flib would sure be
appreciated. 

[Pymc-user] Floats 64 and 32

[David H. <dav...@gm...>]

Hi,  does someone know how to use PyMC with the new scipy core ?
It seems that arrays are float64 by default and the functions in Numeric
only take float32... The effect is that in the method summary, the mean and
variance cannot be computed since they take only 32 bit floats...

Thanks,

David

[Pymc-user] RE: Problem installing PyMC on Win32 with Python 2.3

[Chris F. <fon...@gm...>]

Sorry for the extreme lateness of this reply -- as you can see, there
is little traffic on these lists.

There are now windows installers for PyMC, so the easiest route is
simply to install that. If you prefer to build it yourself, the
appropriate command is:

python setup.py build --compiler=mingw32 install

Hope that helps.

Chris

[Pymc-user] Problem installing PyMC on Win32 with Python 2.3

[Razafindrakoto J. C. <Jea...@ax...>]

=20

When trying to install PyMC, I get the following message error.

Does anyone can help  ?

=20

Regards,

=20

JCR

=20

=20

D:\Razafindrakoto\PythonEnthought\Lib\site-packages\PyMC-0.2.6>python
setup.py install

running install

running build

running config_fc

running build_src

building extension "PyMC.flib" sources

f2py:> build\src\flibmodule.c

creating build

creating build\src

Reading fortran codes...

        Reading file 'PyMC/flib.f'

Post-processing...

        Block: flib

                        Block: poisson

                        Block: binomial

                        Block: normal

                        Block: wishart

                        Block: mvnorm

                        Block: trace

                        Block: gamma

                        Block: trans

                        Block: matmult

                        Block: chol

                        Block: dtrm

                        Block: elgs

Building modules...

        Building module "flib"...

                Constructing wrapper function "poisson"...

                  like =3D poisson(x,mu)

                Constructing wrapper function "binomial"...

                  like =3D binomial(x,n,p)

                Constructing wrapper function "normal"...

                  like =3D normal(x,mu,tau)

                Constructing wrapper function "wishart"...

                  like =3D wishart(x,alpha,beta)

                Constructing wrapper function "mvnorm"...

                  like =3D mvnorm(x,mu,tau)

                Constructing wrapper function "trace"...

                  trace(mat,tr,[k])

                Constructing wrapper function "gamma"...

                  gx =3D gamma(xx)

                Constructing wrapper function "trans"...

                  tmat =3D trans(mat)

                Constructing wrapper function "matmult"...

                  prod =3D matmult(mat1,mat2)

                Constructing wrapper function "chol"...

                  c =3D chol(a)

                Constructing wrapper function "dtrm"...

                  d =3D dtrm(a)

                Constructing wrapper function "elgs"...

                  elgs(a,indx,[n])

        Wrote C/API module "flib" to file "build\src/flibmodule.c"

  adding 'build\src\fortranobject.c' to sources.

  adding 'build\src' to include_dirs.

copying D:\RAZAFI~1\PYTHON~1\lib\site-packages\f2py2e\src\fortranobje=
ct.c ->
bui

ld\src

copying D:\RAZAFI~1\PYTHON~1\lib\site-packages\f2py2e\src\fortranobje=
ct.h ->
bui

ld\src

running build_py

creating build\lib.win32-2.3

creating build\lib.win32-2.3\PyMC

copying PyMC\Approximation.py -> build\lib.win32-2.3\PyMC

copying PyMC\KalmanFiltering.py -> build\lib.win32-2.3\PyMC

copying PyMC\MCMC.py -> build\lib.win32-2.3\PyMC

copying PyMC\Plot.py -> build\lib.win32-2.3\PyMC

copying PyMC\Plot_matplot.py -> build\lib.win32-2.3\PyMC

copying PyMC\Plot_scipy.py -> build\lib.win32-2.3\PyMC

copying PyMC\ReinforcementLearning.py -> build\lib.win32-2.3\PyMC

copying PyMC\__init__.py -> build\lib.win32-2.3\PyMC

running build_ext

No module named msvccompiler in scipy_distutils, trying from distutil=
s..

error: Python was built with version 6 of Visual Studio, and extensio=
ns need
to

be built with the same version of the compiler, but it isn't installe=
d.

=20

D:\Razafindrakoto\PythonEnthought\Lib\site-packages\PyMC-0.2.6>      =
=20

=20

=20

______________________________________________________________

AXA CESSIONS

=20

Jean-Claude RAZAFINDRAKOTO

=20

Re-ARMS

Reinsurance - Actuarial and Risk Management Services

=20

109 rue La Bo=E9tie

75008 Paris

=20

Tel. : +33 1 56 43 78 54

Fax. : +33 1 56 43 78 70

E-mail : jea...@ax...
<mailto:mar...@ax...>=20

_______________________________________________________________

=20