[Numpy-discussion] Style for pad implementation in 'pad' namespace or functions under np.lib
Fri Mar 30 10:02:43 CDT 2012
On Fri, Mar 30, 2012 at 2:20 PM, Tim Cera <email@example.com> wrote:
>> My suggestion is:
>> Step 1: Change the current PR so that it has only one user-exposed
>> function, something like pad(..., mode="foo"), and commit that.
>> Everyone seems to pretty much like that interface, implementing it
>> would take <1 hour of work, and then the basic functionality would be
>> landed and done.
> This is all done in my working directory.
> Currently I have 'mode' as the first argument and not a keyword. Could you
> explain the utility of having it be a keyword, if that is indeed what you
> were advocating earlier?
Eh, just that it feels vaguely more consistent with all the other np
functions, which generally take arrays as their first argument and
sometimes have a mode="string" argument. Examples of the latter:
np.linalg.qr, np.correlate, np.convolve, np.take, np.put, np.choose
And unlike Charles, my first guess at the default would have been
zero-padding (isn't that the most common sort of padding?).
But both of these are minor quibbles -- bikeshedding, really. My
opinion isn't strong.
>> Step 2: Add the option to pass a user-defined function as the mode=
>> argument, since there's still uncertainty about the best way to do it
>> and working through uncertainty adds time and risk that shouldn't hold
>> up the parts that we do agree on.
> This is done also. I don't do any checks. If it isn't a string, then I take
> it to be a function.
Yes, that's probably the Right Way in python.
> The function signature is:
> myfunc(vector, pad_tuple, iaxis, kwds)
> and it has to return a rank 1 array the same length as the input `vector`
>From your other mail, this API does make more sense to me now.
However, looking at the code, I think the returned vector is ignored,
and ideally should be  so as to prevent apply_along_axis from
generating a large temporary matrix that we will in any case discard?
>> Even if we do want to keep around the pad_with_mean, pad_with_median
>> etc. functions as additional user-exposed entry-points, I think the
>> current names in the PR met with objections? (The current names are
>> like "np.lib.pad.pad_mean".)
I wasn't sure if Charles was talking about merging the current pull
request as-is, so I was re-raising the original question you started
the thread with: whether we like having the functions referred to like
"np.lib.pad.pad_mean", or something else would be better (e.g.
"np.lib.pad_with_mean"). If you're changing the PR anyway then you can
just ignore that paragraph :-).
More information about the NumPy-Discussion