numpy-discussion

[Numpy-discussion] slice assignment: strange behaviour

[koara <ko...@at...>]

Hello,

a piece of my code started giving strange results with certain data; i
managed to track down the cause to a slice array assignment. In the
following code snip; 'mat' is a numpy.array with shape=(22973, 1009),
'vec' is a numpy.array with shape=(22973,), both of type int:

for i in xrange(1009):
    ...
    fr = vec[10001]
    mat[:, i] = vec # assign whole column
    if mat[10001, i] != fr:
        print "how come?"
     ...

for elements beyond index 10000, nothing is assigned (ie,
numpy.sum(mat[row, :]) is zero for any row > 10000). As soon as i
replace the assignment with a cycle that assigns each element
explicitly (for j in xrange(22973): result[j, i] = vec[j]),
everything's OK. With some matrices, the above seems to work fine
though, as well as for smaller dimensions, so i am unable to provide a
full simple example.

Any ideas? I am using enthought python (python 2.4.3) which uses numpy
version 0.9.9.

Re: [Numpy-discussion] slice assignment: strange behaviour

[koara <ko...@at...>]

koara wrote:
> Hello,
>
> a piece of my code started giving strange results with certain data; i
> managed to track down the cause to a slice array assignment. In the
> ....

Also if i first build a sequence of columns and then use
numpy.transpose(numpy.vstack(sequence)) the result is ok. But the
arrays can be quite big so in this case i fear space overhead -- or am
i wrong?

Cheers.

Re: [Numpy-discussion] slice assignment: strange behaviour

[Robert K. <rob...@gm...>]

koara wrote:
> koara wrote:
>> Hello,
>>
>> a piece of my code started giving strange results with certain data; i
>> managed to track down the cause to a slice array assignment. In the
>> ....
> 
> Also if i first build a sequence of columns and then use
> numpy.transpose(numpy.vstack(sequence)) the result is ok. But the
> arrays can be quite big so in this case i fear space overhead -- or am
> i wrong?

You might want to use column_stack() instead, for clarity if nothing else.

-- 
Robert Kern

"I have come to believe that the whole world is an enigma, a harmless enigma
 that is made terrible by our own mad attempt to interpret it as though it had
 an underlying truth."
  -- Umberto Eco

Re: [Numpy-discussion] slice assignment: strange behaviour

[Alan G I. <ai...@am...>]

On Wed, 08 Nov 2006, koara apparently wrote: 
> 'mat' is a numpy.array with shape=(22973, 1009), 
> 'vec' is a numpy.array with shape=(22973,), both of type int: 
> for i in xrange(1009): 
>     ... 
>     fr = vec[10001] 
>     mat[:, i] = vec # assign whole column 
>     if mat[10001, i] != fr: 
>         print "how come?" 
>      ... 
> for elements beyond index 10000, nothing is assigned (ie, 
> numpy.sum(mat[row, :]) is zero for any row > 10000). 

Using numpy 1.0 and trying the code below,
I do not see the problem.

import numpy
from numpy import random
m = random.randint(0,1000,(22973,1009))
v =  random.randint(0,1000,(22973,))
fr = v[10001]
for i in xrange(1009):
    m[:,i] = v
    if m[10001, i] != fr:
        print m[10001,i], fr

Upgrading to numpy 1.0 should be easy enough and may fix 
your problem.

Cheers,
Alan Isaac

Re: [Numpy-discussion] slice assignment: strange behaviour

[koara <ko...@at...>]

Ok, so i installed the newest numpy and scipy over enthought defaults.
The original problem indeed disappeared! Given the highly suspicous
magic threshold number of 10000, i suspect it was a bug in old numpy.
Makes me wonder if it's really been solved or just stopped manifesting
itself in my test data :>

Btw numpy.sum now behaves differently when axis is left unspecified: an
example listed in help(numpy.sum) doesn't give the result the help
claims it does, as numpy.sum([[0, 1], [0, 5]]) now gives 6. Someone
ought to either change the description or the code 8)

Cheers