Mailman 3 January 2008 - NumPy-Discussion

Building a static libnumpy
by Jussi Enkovaara 14 Feb '08

14 Feb '08

Hi, I need to use numpy in an environment which does not support shared libraries. Previously, I have used the old Numeric where I created a Makefile to build a static Numeric library which was later on linked to the python interpreter. As far as I understood, numpy uses sort of extended distutils. I was wondering if this extended distutils enables building of static libraries or do I have to go the cumbersome Makefile-way again? Regards, Jussi Enkovaara

5 6

segfault problem with numpy and pickle
by David Bolme 08 Feb '08

08 Feb '08

A am having some trouble when pickling numpy arrays. Basically I use one python script to create the array and then pickle it. When I load the pickled array using a different python script it appears to load fine. When I try to perform a matrix multiply on the array with a vector (using dot) python crashes with a segmentation violation. The array contains 64bit floating point numbers and has a shape of (47, 16384) and the vector has a shape of (16384,1). Other shapes I have tested also have this problem. A workaround is to create a new array that is initialized with the data from the first. For example: a = array(pickle.loads(buffer)) After calling this command the script runs fine. Here is some information on my system setup: OS: Macos X Leopard 64bit Intel numpy.version.version = 1.0.5.dev4624 Below is a simple script that reproduces this error on my system. I would like to get this problem fixed so that I don't have to use this work around every time I load a pickled array. Is this a known bug? I could not find it in the bug tracker. Thanks in advance for the help. Dave import pickle from numpy import array,dot,ones,random # Create the arrays a = random.normal(size=(47,128*128)) b = ones((128*128,2)) c = ones((128*128,1)) # Test dot print "Computing dot product on original arrays..." print dot(a,b).shape # WORKS print dot(a,c).shape # WORKS print "Pickle and unpickel the array with workaround." buffer = pickle.dumps(a) a = array(pickle.loads(buffer)) # Test dot print "Computing dot product with workaround..." print dot(a,b).shape # WORKS print dot(a,c).shape # WORKS print "Pickle and unpickel the array without workaround." a = pickle.loads(buffer) # Test dot print "Computing dot product without workaround..." print dot(a,b).shape # WORKS print dot(a,c).shape # CRASH

3 4

Can not update a submatrix
by Nadav Horesh 01 Feb '08

01 Feb '08

In the following piece of code: >>> import numpy as N >>> R = N.arange(9).reshape(3,3) >>> ax = [1,2] >>> R array([[0, 1, 2], [3, 4, 5], [6, 7, 8]]) >>> R[ax,:][:,ax] = 100 >>> R array([[0, 1, 2], [3, 4, 5], [6, 7, 8]]) Why R is not updated? I was expecting: >>> R array([[0, 1, 2], [3, 100, 100], [6, 100, 100]]) I am using numpy 1.04 on gentoo-linux/amd64 Nadav

7 12

searchsorted bug
by James Philbin 01 Feb '08

01 Feb '08

Hi, The following gives the wrong answer: In [2]: A = array(['a','aa','b']) In [3]: B = array(['d','e']) In [4]: A.searchsorted(B) Out[4]: array([3, 0]) The answer should be [3,3]. I've come across this while trying to come up with an ismember function which works for strings (setmember1d doesn't seems to assume numerical arrays). Thanks, James

11 26

Accepted Numpy Core Datatype Formats
by Mack Smith 31 Jan '08

31 Jan '08

Where can I find a list of accepted tokens for the Numpy Core formats as used in numpy.core.rec.fromfile for instance. I'm working with a binary file format that includes single byte buffers and other formats that I need to correctly parse. Any help would be appreciated. Thanks in advance. Mack

1 0

matrix <-> ndarray bug
by dmitrey 31 Jan '08

31 Jan '08

I don't know, maybe it's already fixed in more recent versions? >>> from numpy import * >>> a=mat('1 2') >>> b = asfarray(a).flatten() >>> print b[0] [[ 1. 2.]] # ^^ I expected getting 1.0 here >>> numpy.version.version '1.0.3'

2 2

blcr checkpoint/restart
by Neal Becker 31 Jan '08

31 Jan '08

Hi numeric processing fans. I'm pleased to report that you can now have convenient checkpoint/restart, at least if you are running fedora linux. Berkeley Lab Checkpoint/Restart for Linux (BLCR) http://ftg.lbl.gov/CheckpointRestart/CheckpointDownloads.shtml What I've done is: 1) built 2 rpm packages for blcr. First installs everything except kernel module. 2nd is the kernel module, packaged for use with akmods, which will be (I think) the kmod format of choice in rpmfusion. 2) made a python ctypes module to use it. This allows your python program to checkpoint itself. You will need a couple of packages from livna development to use this: rpm -q --whatprovides kmodtool kmodtool-1-7.fc8.noarch rpm -q --whatprovides akmods akmods-0.2.1-1.fc8.noarch You can find this stuff here: http://livna-dl.reloumirrors.net/fedora/development/SRPMS/ I grabbed the development srpms from livna and built them for my F8 machine. My stuff is here: https://nbecker.dyndns.org/RPM/blcr_mod.py https://nbecker.dyndns.org/RPM/blcr-0.6.4-1.src.rpm https://nbecker.dyndns.org/RPM/blcr-kmod-0.6.4-2.fc8.src.rpm

1 0

Visual studio with mingw g77 built libraries
by David Cournapeau 31 Jan '08

31 Jan '08

Hi, I would like to ask some details about the build process of numpy when visual studio is used for C compilation, and g77 for blas/lapack. As I understand it, the current situation consists in using libraries built the "Unix way" (e.g. libblas.a, static library built with ar + ranlib), taking the gcc runtimes on the fly (libgcc and g2c), and put everything together to feed VS compiler. I wonder if we should not use another way, that is use libraries built the "windows way" (dll + lib). I think it is more natural (and would also make my life easier for numscons): - it would means that on windows, all compilers combination would follow a similar process. - dll and lib are more natural for windows developers (who are the ones using VS in the first place). The only difference for users would be in the way blas/lapack should be built. Is there a reason why this was not the chosen path ? Is it because it would have been too difficult to do it with distutils (using dlltool and co) ? Or other reasons I am not aware of ? cheers, David

1 0

Median again
by Matthew Brett 30 Jan '08

30 Jan '08

Hi, I was wondering whether there was any plan to change the anomalous interface of median, for example compared to 'mean', 'min', 'max' np.mean(a, axis=None, dtype=None, out=None) whereas: np.median(m) 'median(m) returns a median of m along the first dimension of m.' I think it would be a good to harmonize the median call signature and functionality with mean and the rest, but that will obviously break code. I was wondering whether the team would consider a migration plan on the lines of: median moved mediandim0 implementation of medianwithaxis or similar, with same call signature as mean. Deprecation warning for use of median, and return of mediandim0 for now. Eventual move of median to return medianwithaxis. Does that sound sensible? Matthew

8 12

Vectorizing a function
by LB 30 Jan '08

30 Jan '08

Hi, I've got some questions on the numpy.vectorize function. Currently, i'm doing this kind of work : [ code] def calc_0d(x, y): """ make complexe calculation using two scalars x and y """ [ ... ] return res1, res2, res 3 # vectorize the function calc = vectorize(calc_0d) res1, res_2, res_3 = calc(array_x, array_y) [/code] This works fine. Really impressive. Good work for this ! My problem is that the complexe calculations made in calc_0d use some parameters, which are currently defined at the head of my python file. This is not very nice and I can't define a module containing theses two functions and call them with different parameters. I would like to make this cleaner and pass theses parameter as keyword argument, but this don't seems to be possible with vectorize. Indeed, some of theses parameters are array parameters and only the x and y arguments should be interpreted with the broadcasting rules.... What is the "good way" for doing this ? Regards, -- LB

7 9