[Numpy-discussion] Use-case for np.choose
josef.pktd@gmai...
josef.pktd@gmai...
Mon Nov 9 19:16:58 CST 2009
On Mon, Nov 9, 2009 at 7:59 PM, <josef.pktd@gmail.com> wrote:
> On Mon, Nov 9, 2009 at 7:54 PM, David Goldsmith <d.l.goldsmith@gmail.com> wrote:
>> May I infer from the sudden silence that I finally have it?
>
> I think so,
> I assume that the result of broadcasting is unique, I haven't seen an
> example yet where broadcasting would depend on the sequence in which
> it is done.
on a related note:
I looked at the doc editor discussion for numpy.where and it seems
broadcasting has been fixed also for where (in 1.3.0). I don't know
whether the c implementation is related to choose. I only tried Paulis
example.
Josef
>
> Josef
>
>
>>
>> DG
>>
>> On Sun, Nov 8, 2009 at 8:50 PM, David Goldsmith <d.l.goldsmith@gmail.com>
>> wrote:
>>>
>>> OK, let me see if I'm interpreting this example correctly:
>>>
>>> >>> c1=np.arange(2).reshape(2,1,1); c1
>>> array([[[0]],
>>>
>>> [[1]]])
>>> >>> c2=2+np.arange(2).reshape(1,1,2); c2
>>> array([[[2, 3]]])
>>> >>> a=np.eye(2,dtype=int)
>>> >>> np.choose(a, [c1, c2])
>>> array([[[2, 0],
>>> [0, 3]],
>>>
>>> [[2, 1],
>>> [1, 3]]])
>>>
>>> First, everything is being broadcast to (2,2,2); a is broadcast to
>>> [[[1,0], [0,1]], [[1,0], [0,1]]], c1 is broadcast to [[[0,0], [0,0]],
>>> [[1,1], [1,1]]] and c2 is broadcast to [[[2,3], [2,3]], [[2,3], [2,3]]].
>>> Now result is created by "stepping through" broadcast a and using,
>>> respectively, the positionally corresponding element from broadcast c1
>>> (resp. c2) if the value in a at the position is 0 (resp. 1). At least, this
>>> gives the result above (but I have not examined other possible broadcasts of
>>> the arguments to see if they would also give the result - I conjectured what
>>> appeared to me to be the most "natural" broadcasts and checked to see if it
>>> worked and it does; is there something I should know - e.g., uniqueness of
>>> the result, or a rule governing how choose broadcasts - to *know* that the
>>> broadcasts above are indeed the broadcasts choose is using?)
>>>
>>> Thanks again,
>>>
>>> DG
>>> On Sun, Nov 8, 2009 at 8:19 PM, Anne Archibald <peridot.faceted@gmail.com>
>>> wrote:
>>>>
>>>> 2009/11/8 David Goldsmith <d.l.goldsmith@gmail.com>:
>>>> > On Sun, Nov 8, 2009 at 7:40 PM, Anne Archibald
>>>> > <peridot.faceted@gmail.com>
>>>> > wrote:
>>>> >>
>>>> >> As Josef said, this is not correct. I think the key point of confusion
>>>> >> is
>>>> >> this:
>>>> >>
>>>> >> Do not pass choose two arrays.
>>>> >>
>>>> >> Pass it one array and a *list* of arrays. The fact that choices can be
>>>> >> an array is a quirk we can't change, but you should think of the
>>>> >> second argument as a list of arrays,
>>>> >
>>>> > Fine, but as you say, one *can* pass choose an array as the second
>>>> > argument
>>>> > and it doesn't raise an exception, so if someone is stupid/careless
>>>> > enough
>>>> > to pass an array for `choices`, how is choose interpreting it as a
>>>> > list? Is
>>>> > the first dimension "list converted" (so that, e.g., my (2,1,2) example
>>>> > is
>>>> > interpreted as a two element list, each of whose elements is a (1,2)
>>>> > array)?
>>>>
>>>> It seems to me that this is the only reasonable interpretation, yes.
>>>> After all, arrays behave like sequences along the first axis, whose
>>>> elements are arrays of one less dimension. Thus if you pass an array,
>>>> any broadcasting happens ignoring the first axis, which is a rather
>>>> abnormal pattern for numpy broadcasting, but necessary here.
>>>>
>>>> As a bonus, I think this is what is implemented in current versions of
>>>> numpy. (In 1.2.1 it raises an exception if broadcasting is necessary.)
>>>>
>>>> Anne
>>>> _______________________________________________
>>>> NumPy-Discussion mailing list
>>>> NumPy-Discussion@scipy.org
>>>> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>>>
>>
>>
>> _______________________________________________
>> NumPy-Discussion mailing list
>> NumPy-Discussion@scipy.org
>> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>>
>>
>
More information about the NumPy-Discussion
mailing list