Mailman 3 September 2007 - NumPy-Discussion

2-d in-place operation performance vs 1-d non in-place
by George Sakkis Sept. 6, 2007

Sept. 6, 2007

I was surprised to see that an in-place modification of a 2-d array turns out to be slower from the respective non-mutating operation on 1- d arrays, although the latter creates new array objects. Here is the benchmarking code: import timeit for n in 10,100,1000,10000: setup = 'from numpy.random import random;' \ 'm=random((%d,2));' \ 'u1=random(%d);' \ 'u2=u1.reshape((u1.size,1))' % (n,n) timers = [timeit.Timer(stmt,setup) for stmt in # 1-d operations; create new arrays 'a0 = m[:,0]-u1; a1 = m[:,1]-u1', # 2-d in place operation 'm -= u2' ] print n, [min(timer.repeat(3,1000)) for timer in timers] And some results (Python 2.5, WinXP): 10 [0.010832382327921563, 0.0045706926438974782] 100 [0.010882668048592767, 0.021704993232380093] 1000 [0.018272154701226007, 0.19477587235249172] 10000 [0.073787590322233698, 1.9234369172618306] So the 2-d in-place modification time grows linearly with the array size but the 1-d operations are much more efficient, despite allocating new arrays while doing so. What gives ? George

3 3

Re: [Numpy-discussion] Use my own data type with NumPy
by James A. Bednar Sept. 5, 2007

Sept. 5, 2007

| Date: Wed, 05 Sep 2007 21:19:58 +0200 | From: G?nter Dannoritzer <dannoritzer(a)web.de> | Subject: Re: [Numpy-discussion] Use my own data type with NumPy | | The purpose of my (Python) class is to model a fixed point data | type. So I can specify how many bits are used for integer and how | many bits are used for fractional representation. Then it should be | possible to assign a value and do basic arithmetic with an instance | of that class. The idea is that based on fixed point arithmetic | rules, each operation tracks changes of bit width. You may already be aware of this, but there is a package available that does exactly what you describe: http://fixedpoint.sourceforge.net It is not currently actively aintained, and it's quite slow, but it does work reliably, and we use it daily at my site. It might be best for all concerned if you simply took over that project, making it faster and well supported, rather than creating a competing one... Jim

1 0

Use my own data type with NumPy
by Günter Dannoritzer Sept. 5, 2007

Sept. 5, 2007

Hi, I am trying to use my own data type with NumPy, but get some funny result when creating a NumPy array with it. My data type is indexable and sliceable and what happens now is when I create an array, NumPy is adding the instance as a list of the indexed values. How can I force NumPy to handle my data type as an 'Object' and use the value of __repr__ to display in the array? Maybe I am handling that too simple, as I had another data type before, -- not indexable though --, that just works fine with NumPy. Do I have to worry about a dtype object with my new data type or how can I use my new data type with NumPy? Thanks for your help. Cheers, Guenter

3 6

How-to: Uniform vector scale
by Robert Dailey Sept. 5, 2007

Sept. 5, 2007

Hi, I have a scalar value S. I want to perform the following math on vectors A and B (both of type array): A + B * S By order of operations, B * S should be done first. This is a vector multiplied by a real number and should be valid. However, the interpreter outputs: "ValueError: shape mismatch: objects cannot be broadcast to a single shape" I am not sure how I am supposed to multiply a vector with a scalar value. For example: array([5,2]) * 2 = [10,4] The above should happen. However, I get the error message instead. Any ideas? Thanks.

3 2

numpy.rot90 bug with >2D arrays
by Zachary Pincus Sept. 5, 2007

Sept. 5, 2007

Hello, numpy.rot90 (from twodim_base.by) says it works only on the first two axes of an array, but due to its use of the transpose method (which reverses the shape tuple), can affect other axes. For example: a = numpy.ones((50,40,3)) b = numpy.rot90(a) assert(b.shape == (3,40,50)) # desired result is b.shape == (40,50,3) I believe that the fix is to replace the two calls to 'transpose()' in the function with 'swapaxes(0,1)'. This would allow rotating an array along just the first two axes. Does this fix look right? Should I file a bug or can someone just check that in? Zach

1 0

Vector magnitude?
by Robert Dailey Sept. 5, 2007

Sept. 5, 2007

Hi, I have two questions: 1) Is there any way in numpy to represent vectors? Currently I'm using 'array' for vectors. 2) Is there a way to calculate the magnitude (length) of a vector in numpy? Thanks.

7 9

Improving bug triage in trac ?
by David Cournapeau Sept. 5, 2007

Sept. 5, 2007

Hi there, I am personnally a bit annoyed by the way trac handle bug reports, and would like to know if there is space for improvement. I do not know much about bug tracking systems, so maybe I just don't know how to use it, though. The main thing I dislike is the status of tickets and reports. In particular: - I don't know how other feel, but I am rather confused by the meta data of a ticket, and do not find them really useful from a developer point of view. It would be quite helpful to have information whether the bug is confirmed or not, another one which says wether a patch is available or not. This would make bug triage much easier, IMHO. This should be possible, since I have seen some trac installation with such features (wordpad trac, for example). - This one maybe a bit more difficult to implement I guess (I don't know anything about trac internals): I find the general view of bugs for a given repository really helpful in launchpad, in perticular, you can easily view the percentage of bugs wrt their status (eg 30 % bugs unconfirmed, etc...); see for example https://bugs.launchpad.net/+about. This gives a pretty good idea of what needs to be done for a particular release. How do other people feel about those suggestions ? cheers, David

2 1

Re: [Numpy-discussion] numpy build fails on powerpc ydl
by Vincent Broman Sept. 5, 2007

Sept. 5, 2007

Harris asked about long doubles. On my YDL, SIZEOF_LONG_DOUBLE and sizeof( long double) were both 8. Vincent Broman

2 1

Change array memory ownership status
by Marc POINOT Sept. 5, 2007

Sept. 5, 2007

I want to change the "status" of a numpy array. I mean this array was created by a server application using PyArray_FromDimsAndData that sets the NPY_OWNDATA flag to False. The server application believes the client would free the memory. But there are more than one client application and none knows who is in charge of freeing this memory. Then I want to set the flag NPY_OWNDATA to True to tell the server to do the job when it finishes the script. How can I do that, I mean at the Python interface level, not the C API. >>> print a.flags.owndata True >>> >>> a.flags.owndata=False Traceback (most recent call last): File "<stdin>", line 1, in ? TypeError: attribute 'owndata' of 'numpy.flagsobj' objects is not writable >>> Or should I have to write my own function set/get for OWNDATA ? The get is there but I can't get the set... -MP- ----------------------------------------------------------------------- Marc POINOT [ONERA/DSNA] Tel:+33.1.46.73.42.84 Fax:+33.1.46.73.41.66 Avertissement/disclaimer http://www.onera.fr/onera-en/emails-terms

1 0

Re: [Numpy-discussion] numpy build fails on powerpc ydl
by Vincent Broman Sept. 5, 2007

Sept. 5, 2007

Oliphant: > Would you be willing to help get the config.h file set up correctly? Yes. I thought it was automatic, tho. What to do? me: > Its kernel is 2.4.19-Asmp tailored by the vendor. Harris: > Which vendor? Curtiss-Wright Controls, back when they were called Synergy. Harris: > Ancient.... Curtiss-Wright now supports Linux and kernel 2.6.16 on > some of their newer hardware The military ends up supporting ancient systems for a long time. I don't think I have the option of upgrading the OS on these beasts, especially when it involves a new binary format. That would be nice, tho, especially if it made it possible to run subversion. Harris: > Any more detail on these? What causes the conflict. I've got to wonder about > the the libc/libm versions also. Does the include file math.h say anything > about the prototypes for these functions? I expect cosl et.al. to be > potential problems on the PPC anyway due to the way long doubles were > implemented. "grep -r sinl /usr/include" finds nothing. math.h defines sin and sinf with various underscores attached, using token pasting, but no mention of sinl. Similarly for fabs and cos which I checked. Harris: > What is the PPC model number? The CWC VSS4 contains four powerpc G4's, the 7400. Vincent Broman

2 1