Mailman 3 March 2007 - NumPy-Discussion

Re: [Numpy-discussion] scalar recordarrays
by Matthew Koichi Grimes 22 Mar '07

22 Mar '07

If there are no objections, I'll file this ticket in the trac site: <snip> Title: Return type inconsistency in recarray Description: The sub-arrays of rank-0 recarrays are returned as scalars rather than rank-0 ndarrays. Example: >>> import numpy as N >>> dt = N.dtype([('x','f8'),('y','f8')]) >>> rarr = N.zeros((), dtype = dt).view(N.recarray) >>> rarr recarray((0.0, 0.0), dtype=[('x', '<f8'), ('y', '<f8')]) >>> # oddly, rarr.x is not an array >>> rarr.x 0.0 >>> # This makes it impossible to fill values in the usual manner >>> rarr.x[...] = 2.0 TypeError: object does not support item assignment >>> # A workaround is to reshape to non-zero rank >>> rarr.shape = [1] >>> # Now rarr.x returns a ndarray of the same rank as rarr >>> rarr.x array([ 0.]) >>> # Value-setting now works as expected. >>> rarr.x[...] = 2.0 >>> rarr recarray([(2.0, 0.0)], dtype=[('x', '<f8'), ('y', '<f8')]) This is a problem for any function that takes record arrays of arbitrary rank, as it necessitates special-case checks for ndim == 0. </snip> -- Matt Francesc Altet wrote: > Yeah, I agree. In fact, one of the main reasons to keep rank-0 arrays > around is to provide generality enough to address to these kind of > problems in an elegant way. > > Cheers, >

1 0

distutils for a Pyrex module
by Pierre GM 22 Mar '07

22 Mar '07

All, I'm trying to write a numpy.distutils setup.py for a pyrex module that involves both external C and fortran sources, and where the fortran sources need to be linked w/ blas and lapack. Here's what I have so far: ################################## def configuration(parent_package='',top_path=None): from numpy.distutils.misc_util import Configuration, get_mathlibs config = Configuration('cloess',parent_package,top_path) from numpy.distutils.system_info import get_info, dict_append libs = [] generate_libraries = [] f_sources = ['loessf.f', 'linpack_lite.f'] c_sources = ['loess.c', 'loessc.c', 'misc.c', 'predict.c',] # blas_info = get_info('blas_opt') lapack_info = get_info('lapack_opt') build_info = {} dict_append(build_info, **blas_info) dict_append(build_info, **lapack_info) # config.add_extension('cloess', sources=[join('src', x) for x in (f_sources+c_sources)], depends = [join('src','*.h'), join('src','*.pyx'), join('src','*.pxd') ], **build_info ) config.add_data_dir(join('.','examples')) config.add_data_files('pyloess.py') return config ############################ But that doesn't do what I want: during the setup, the .f files are processed by f2py. I don't need that, I just want to have the Fortran files compiled along the C sources. What's the M.O., then ? FYI, here are the commands I'm using to build the library w/o distutils: pyrexc cloess.pyx || exit gcc -g -c -fPIC loess.c loessc.c misc.c predict.c loessf.f linpack_lite.f gcc -g -c -fPIC cloess.c -I/usr/include/python2.4 \ -I/usr/lib64/python2.4/site-packages/numpy/core/include gcc -g -shared loessf.o linpack_lite.o loess.o loessc.o misc.o predict.o cloess.o -llapack -lblas -o cloess.so Any help will be welcome. Thanks a lot in advance (PS: initially posted on the pyrex mailing list and moved here on Robert Kern's request).

3 7

memory leak w/ numpy & Numeric together
by Bryan Cole 22 Mar '07

22 Mar '07

I'm not sure where best to post this, but I get a memory leak when using code with both numpy and FFT(from Numeric) together: e.g. >>> import numpy >>> import FFT >>> def test(): ... while 1: ... data=numpy.random.random(2048) ... newdata = FFT.real_fft(data) >>> test() and memory consumption goes through the roof. I'm using numpy-1.0.1 and Numeric-24.2 (on FC6). This has now forced me to go 100% numpy... BC

2 1

Little module to get numpy examples
by Bill Baxter 22 Mar '07

22 Mar '07

I wrote a little python module to go fetch the Numpy examples from the scipy wiki page, parse them, and print out entries. Is there a good place on the wiki for this? It didn't really seem right in the cookbook, and it doesn't quite fit with documentation. Here's an example of use: ----------------------------------------- In [1]: import numpy In [2]: from NumpyExamples import examples In [3]: examples(numpy.hstack) --- hstack --- >>> from numpy import * >>> a =array([[1],[2]]) # 2x1 array >>> b = array([[3,4],[5,6]]) # 2x2 array >>> hstack((a,b,a)) # only the 2nd dimension of the arrays is allowed to be different array([[1, 3, 4, 1], [2, 5, 6, 2]]) See also: column_stack, concatenate, dstack, vstack, hsplit ----------------------------------------- --bb

3 4

Re: [Numpy-discussion] Translation of Matlab function find()
by Miquel Poch 22 Mar '07

22 Mar '07

Hi, I'm trying to translate some Matlab functions. I don't know exactly how make and equivalent for function find(). This give us the index of an array where a condition it's true. I've found some options like this one: (a>0).nonzero(), where a is the array and (a>0) the condition. But the problem appear with this function return. It's something like this: >>> from numpy import * >>> a=array([0,1,0,1,2,0,1,2,0,0]) >>> print a [0 1 0 1 2 0 1 2 0 0] >>> b=(a>0).nonzero() >>> print b (array([1, 3, 4, 6, 7]),) >>> print (a>0).nonzero() (array([1, 3, 4, 6, 7]),) I can't acces to b data, doing something like b[x]. Why b is not an array? I think another possible is use where() function, also avilable in numpy library. It's better or worse? Thanks in advance, Miquel

3 2

problems building exported version of numpy from svn
by Fred Romelfanger 22 Mar '07

22 Mar '07

When I use co to check the code out of svn, running setup.py and building the code works fine, but when I export the code or create a source distribution that includes numpy the .svn directories get stripped. svn export http://svn.scipy.org/svn/numpy/trunk numpy I then get the following error when I run "python setup.py build" AssertionError: hmm, why I am not inside SVN tree??? How do I export a copy of numpy that I can build outside of an svn tree? We were using numpy 1.0.1, but it doesn't build cleanly on RHEL v3 with the old version of ATLAS that comes with the system software. I was hoping a newer version of numpy from svn might correct that problem. Thanks -- fred.

2 1

basearray & dimarray : you have dead URL; + some questions
by dmitrey 22 Mar '07

22 Mar '07

Hallo! Excuse my bad English. In the web page http://www.scipy.org/BaseArray/Application in the section Project details and tentative schedule you refer to the dead link with name http://numeric.scipy.org/array_interface.html which leads to http://numpy.scipy.org//array_interface.html which is dead (with singular '/' before array_interface.html - too) I'm using Mozilla browser if it's important. Now I'm writing some code and want to know, what will dotwise and matrix operations look like? Is there any discussion where these and related questions are concerned? So, will they be MATLAB/Octave/omatrix/SciLab/etc look-like: dot is present => dotwise operation dot is absend => matrix operation or you'll continue numpy array operations, which seems to be very unclear for newbies? As for me I'm sure that in long-term period 1st approach is much more better, because users will not have to dig in numpy documentation for to chose appropriate operator, and translating code from those languages will be much more easier. Of course, many people can say that 2nd way simplifies translating code from already-written py-files, but it will be done only once, while migrating MATLAB, Octave, etc users will continue for years or dozens of years or even more. Maybe you'll organize a voting or separate discussion thread or something else to gather users opinions? And 2nd question is - is it so important to have own class? Maybe something from c/c++ boost (www.boost.org), or collaboration with Octave developers will be more fast & require less workers and money? And it will be possible to fork any time when something will be going wrong. As far as I understood from my discussions in octave mailing lists with David Bateman, they still have many problems with for example sparse matrices. Even Mathworks still have some troubles, and sparse witn dimension > 2 isn't implemented at all. So collaborating could be much more productive. And, if you still want to create 100% your own code, what are tasks for students participating in GSoC? I want to estimate the complexity of jobs (am I able to finish them successfully till the deadlines or no for any reasons). Thank you in advance for your answer, Dmitrey.

2 1

Array of Callables
by Shane Holloway 22 Mar '07

22 Mar '07

To the vector-processing masters of numpy! I'm wanting to optimize calling a list (or array) of callable objects. Consider the following: vCallables = numpy.array([<python callables: fns, methods, lambdas, classes, builtin functions>]) vParam1 = numpy.array([<parameters for vCallables>]) vParam2 = numpy.array([<parameters for vCallables>]) vParam3 = numpy.array([<parameters for vCallables>]) vResults = numpy.array([None for e in vCallables]) Is it possible to have numpy call them in a way something like:: for i in xrange(vCallables.shape[0]): vResults[i] = vCallables[i] ( vParam1[i], vParam2[i], vParam3[i] ) With the core of that loop implemented in C? Thanks for reading and pondering! -Shane Holloway

5 8

Matlab -> NumPy translation and indexing
by David Koch 21 Mar '07

21 Mar '07

Hello, I am trying to translate some Matlab code to NumPy. I started reading the NumPy book and, yeah it's a very long read :-/ One thing I am completely confused about are the concpets of "basic" vs. "advanced" indexing. Are there some good examples out there where for the same piece of code - Matlab and NumPy are compared? I feel looking at those maybe more helpful to get me started than an in-depth study of the book. I already read "Numpy for Matlab users" and at a glance it doesn't seem to contain equivalents to all the common indexing operations you'd do in Matlab. Some concrete problems/questions I have right now: - Am I correct in assuming that all arrays have to be initialized to their final number of elements in NumPy (using empty/zero for instance)? - Given a matrix R, is there an equvialent to the Matlab operation R(:,j) = [] (which removes column j and "shrinks" the matrix? - For (basic ?) indexing of an ndarray - is there any reason to prefer A[i,j] over A[i][j]? - Is it "pythonic" to initialize vectors to 2 dimensions so that vec.shape== (len(vec), 1) instead of vec.shape == (len(vec),)? Alright, Thanks for now, David

12 22

PyChecker
by Charles R Harris 21 Mar '07

21 Mar '07

All, I've run pychecker on some of my programs that import numpy and it issues many warnings about numpy. Many (most?) of the warnings look bogus, but some also look valid. I will try to work through some of these, but I want to suggest that we run pychecker, and maybe pylint, over the python code to look for errors. Chuck

3 3