[Numpy-discussion] Using nditer + external_loop to Always Iterate by Column

[Sebastian Berg]

On Tue, 2020-05-19 at 17:11 +0100, Eric Wieser wrote:
> Hi Will,
> 
> To force an iteration to run along certain axes, I believe you should
> be
> using `op_axes`. Your diagnosis is correct that `external_loop` is
> trying
> to help you be more optimal, since it's purpose is exactly that:
> optimization.
> 
> Unfortunately, if you use `op_axes` you'll run into
> https://github.com/numpy/numpy/issues/9808.


Yeah, I do not think what you want is possible with nditer. `op_axes`
allows you to ignore certain axis, but as Eric points out, that
actually ignores that axis entirely. So it acts similar to slicing the
whole axis away.

I previously thought it could be nice to have a new
`arr.iteraxis(axis=None)` command. As a base, it would act like a flat
iter, otherwise iterate the axis listed. Maybe with an additional flag
whether its allowed to optimize iteration order (although from the
python side I expect we do not want to do that by default).

The reason is also that I am not sure I like `arr.flat` and `for subarr
in arr` too much, because the list-of-list like iteration seems only
semi-natural for an N-D array.

- Sebastian


> 
> Eric
> 
> On Tue, 19 May 2020 at 00:42, William Ayd <william.ayd at icloud.com>
> wrote:
> 
> > I am trying to use the nditer to traverse each column of a 2D
> > array,
> > returning the column as a 1D array. Consulting the docs, I found
> > this
> > example which works perfectly fine:
> > 
> > In [*65*]: a = np.arange(6).reshape(2,3)
> > 
> > 
> > 
> > In [*66*]: *for* x *in* np.nditer(a, flags=['external_loop'],
> > order='F'):
> >     ...:     print(x, end=' ')
> >     ...:
> > 
> > 
> > [0 3] [1 4] [2 5]
> > 
> > When changing the shape of the input array to (1, 3) however, this
> > doesn’t
> > yield what I am hoping for any more (essentially [0], [1] [2]):
> > 
> > In [*68*]: *for* x *in* np.nditer(a, flags=['external_loop'],
> > order='F'):
> >     ...:     print(x, end=' ')
> >     ...:
> > 
> > 
> > [0 1 2]
> > 
> > I suspect this may have to do with the fact that the (1, 3) array
> > is both
> > C and F contiguous, and it is trying to return as large of a 1D
> > F-contiguous array as it can. However, I didn’t see any way to
> > really force
> > it to go by columns. My best guess was the *itershape* argument
> > though I
> > couldn’t figure out how to get that to work and didn’t see much in
> > the
> > documentation.
> > 
> > Thanks in advance for the help!
> > 
> > - Will
> > 
> > 
> > 
> > _______________________________________________
> > NumPy-Discussion mailing list
> > NumPy-Discussion at python.org
> > https://mail.python.org/mailman/listinfo/numpy-discussion
> > 
> 
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at python.org
> https://mail.python.org/mailman/listinfo/numpy-discussion