Re: [Numpy-discussion] nonzero() behaviour has changed

[Ed S. <sch...@ft...>]

Alan G Isaac wrote:
> On Tue, 16 May 2006, Alan G Isaac apparently wrote: 
>   
>> 2. How are people using this?  I trust that the numarray 
>> behavior was well considered, but I would have expected 
>> coordinates to be grouped rather than spread across the 
>> arrays in the tuple. 
>>     
>
> OK, just to satisfy my curiosity:
> does the silence mean that nobody is using
> 'nonzero' in a fashion that leads them to
> prefer the current behavior to the "obvious"
> alternative of grouping the coordinates?
> Is the current behavior just an inherited
> convention, or is it useful in specific applications?
>   

I also think a function that groups coordinates would be useful.  Here's
a prototype:

def argnonzero(a):
    return transpose(a.nonzero())

This has the same effect as:

def argnonzero(a):
    nz = a.nonzero()
    return array([[nz[i][j] for i in xrange(a.ndim)] for j in
xrange(len(nz[0]))])

The output is always a 2d array, so

>>> a
array([[ 0,  1,  2,  3],
       [ 4,  0,  6,  7],
       [ 8,  9,  0, 11]])

>>> argnonzero(a)
array([[0, 1],
       [0, 2],
       [0, 3],
       [1, 0],
       [1, 2],
       [1, 3],
       [2, 0],
       [2, 1],
       [2, 3]])

>>> b = a[0]
>>> argnonzero(b)
array([[1],
       [2],
       [3]])

>>> c = array([a, a-1, a-2])
>>> argnonzero(c)
array([[0, 0, 1],
       [0, 0, 2],
       ...

It looks a little clumsy for 1d arrays, but I'd argue that, if NumPy
were to offer a function like this, it should always return a 2d array
whose rows are the coordinates for consistency, rather than returning
some squeezed version for indices into 1d arrays.

I'd support the addition of such a function to NumPy.  Although it's
tiny, it's not obvious, and it might be useful.

-- Ed