[Numpy-discussion] Incrementing with advanced indexing: why don't repeated indexes repeatedly increment?

[Robert Cimrman]

On 06/06/2012 05:06 PM, Nathaniel Smith wrote:
> On Wed, Jun 6, 2012 at 9:48 AM, John Salvatier
> <jsalvati at u.washington.edu>  wrote:
>> Hello,
>>
>> I've noticed that If you try to increment elements of an array with advanced
>> indexing, repeated indexes don't get repeatedly incremented. For example:
>>
>> In [30]: x = zeros(5)
>>
>> In [31]: idx = array([1,1,1,3,4])
>>
>> In [32]: x[idx] += [2,4,8,10,30]
>>
>> In [33]: x
>> Out[33]: array([  0.,   8.,   0.,  10.,  30.])
>>
>> I would intuitively expect the output to be array([0,14, 0,10,30]) since
>> index 1 is incremented by 2+4+8=14, but instead it seems to only increment
>> by 8. What is numpy actually doing here?
>>
>> The authors of Theano noticed this behavior a while ago so they python loop
>> through the values in idx (this kind of calculation is necessary for
>> calculating gradients), but this is a bit slow for my purposes, so I'd like
>> to figure out how to get the behavior I expected, but faster.
>>
>> I'm also not sure how to navigate the numpy codebase, where would I look for
>> the code responsible for this behavior?
>
> Strictly speaking, it isn't actually in the numpy codebase at all --
> what's happening is that the Python interpreter sees this code:
>
>    x[idx] += vals
>
> and then it translates it into this code before running it:
>
>    tmp = x.__getitem__(idx)
>    tmp = tmp.__iadd__(vals)
>    x.__setitem__(idx, tmp)
>
> So you can find the implementations of the ndarray methods
> __getitem__, __iadd__, __setitem__ (they're called
> array_subscript_nice, array_inplace_add, and array_ass_sub in the C
> code), but there's no way to fix them so that this works the way you
> want it to, because there's no way for __iadd__ to know that the
> temporary values that it's working with are really duplicate copies of
> "the same" value in the original array.
>
> It would be nice if numpy had some sort of standard API for doing what
> you want, but not sure what a good API would look like, and someone
> would have to implement it.

This operation is also heavily used for the finite element assembling, and a 
similar question has been raised already several times (e.g. 
http://old.nabble.com/How-to-assemble-large-sparse-matrices-effectively-td33833855.html). 
So why not adding a function np.assemble()?

r.