[Numpy-discussion] Should arr.diagonal() return a copy or a view? (1.7 compatibility issue)

Nathaniel Smith njs at pobox.com
Fri May 11 19:54:11 EDT 2012


On Fri, May 11, 2012 at 9:26 PM, T J <tjhnson at gmail.com> wrote:
> On Fri, May 11, 2012 at 1:12 PM, Mark Wiebe <mwwiebe at gmail.com> wrote:
>>
>> On Fri, May 11, 2012 at 2:18 PM, Pauli Virtanen <pav at iki.fi> wrote:
>>>
>>> 11.05.2012 17:54, Frédéric Bastien kirjoitti:
>>> > In Theano we use a view, but that is not relevant as it is the
>>> > compiler that tell what is inplace. So this is invisible to the user.
>>> >
>>> > What about a parameter to diagonal() that default to return a view not
>>> > writable as you said. The user can then choose what it want and this
>>> > don't break the inferface.
>>> [clip]
>>>
>>> Agreed, it seems this is the appropriate way to go on here
>>> `diagonal(copy=True)`. A more obscure alternative would be to add a
>>> separate method that returns a view.
>>
>>
>> This looks like the best way to deal with it, yes.
>>
>> Cheers,
>> Mark
>>
>>>
>>>
>>> I don't think changing the default behavior in a later release is a good
>>> idea. It's a sort of an API wart, but IMHO better that than subtle code
>>> breakage.
>>>
>>>
>
> copy=True seems fine, but is this the final plan?   What about long term,
> should diag() eventually be brought in line with transpose() and reshape()
> so that it is a view by default?  Changing default behavior is certainly not
> something that should be done all the time, but it *can* be done if
> deprecated appropriately.  A more consistent API is better than one with
> warts (if this particular issue is actually seen as a wart).

This is my question as well. Adding copy=True default argument is
certainly a fine solution for 1.7, but IMHO in the long run it would
be better for diagonal() to return a view by default. (Aside from it
seeming generally more "numpythonic", I see from auditing the code I
have lying around my homedir that it would generally be a free speed
win, and having to remember to type a.diagonal(copy=False) all the
time in order to get full performance seems a bit annoying.)

I mean, I'm all for conservatism in these things, which is why I
raised the issue in the first place :-). But it also seems like there
should be *some* mechanism for getting there from here (assuming
others agree we want to). There's been grumblings about trying to do
more evolving of numpy in-place instead of putting everything off to
the legendary 2.0, right?

Unfortunately just putting a deprecation warning on everyone who calls
diagonal() without an explicit copy= argument seems like it would be
*really* obnoxious, though. If necessary we could get more creative...
add a special-purpose ndarray flag like WARN_ON_WRITE so that code
which writes to the returned array continues to work like now, but
also triggers a deprecation warning? I dunno.

-- Nathaniel



More information about the NumPy-Discussion mailing list