I think it's certainly functionality that we want to have. Details of the
design will have to go through the list. One limitation of your initial
straw man proposal is that we'll want to handle gufuncs with multiple
arguments and specify the axes for each. (E.g. think of matrix multiply.)
On 17 Oct 2014 06:40, "Stephan Hoyer" notifications@github.com wrote:

I would like to write generalized ufuncs, using numba
numba/numba#825, to create fast functions
such as nanmean (signature '(n)->()') or rolling_mean (signature
'(n),()->(n)') that take the axis along which to aggregate as a keyword
argument, e.g., nanmean(x, axis=0) or rolling_mean(x, window=5, axis=0).

Of course, I have the option of writing my own wrapper function to do
something like x.swapaxes(axis, -1) before passing the array off to the
gufunc. This is not so terrible.

However, it occurs to me that this sort of dimension reordering might
be handled more cleanly (and efficiently) within the gufunc machinery
itself, somewhat analogously to numpy.ufunc.reduce, except the axis
argument (if supplied) would need to be an integer or tuple with length
equal to the number of core dimensions on the first gufunc argument.

Thoughts?

—
Reply to this email directly or view it on GitHub
#5197.

I think it would be a great idea, if we can come up with a syntax that's both general enough and simple enough. The example of matrix_multiply, and its (m,n),(n,p)->(m,p) illustrates this perfectly. You could then specify something like axis=[(-3, -1), (0, 1)], or even axis=[(-3, -1), (0, 1), (-1, -2)] and specify also the axes of the output. As an example already in the API, np.cross has axis selection for multiple operands, although its interface is probably not the best for gufuncs.

There are also plans to add optional dimensions to gufunc signatures, e.g. make matrix_mutiply's signature (m?,n),(n,p?)->(m?,p?) so that it can also handle products with vectors. We also need to figure out how these two would interact.

But yes, it would be great if you could bring this up on the list.

Possibly we should take advantage of the fact that gufunc axes have names
to allow something more symbolic like axis=[{"m": 0, "n": 2}, {"n": 0, "p":
1}].

On Fri, Oct 17, 2014 at 3:16 PM, Jaime notifications@github.com wrote:

I think it would be a great idea, if we can come up with a syntax that's
both general enough and simple enough. The example of matrix_multiply,
and its (m,n),(n,p)->(m,p) illustrates this perfectly. You could then
specify something like axis=[(-3, -1), (0, 1)], or even axis=[(-3, -1),
(0, 1), (-1, -2)] and specify also the axes of the output. As an example
already in the API, np.cross
http://docs.scipy.org/doc/numpy/reference/generated/numpy.cross.html
has axis selection for multiple operands, although its interface is
probably not the best for gufuncs.

There are also plans to add optional dimensions to gufunc signatures, e.g.
make matrix_mutiply's signature (m?,n),(n,p?)->(m?,p?) so that it can
also handle products with vectors. We also need to figure out how these two
would interact.

But yes, it would be great if you could bring this up on the list.

—
Reply to this email directly or view it on GitHub
#5197 (comment).

Nathaniel J. Smith
Postdoctoral researcher - Informatics - University of Edinburgh
http://vorpus.org

That would be very nice indeed.

Posted to the list: http://mail.scipy.org/pipermail/numpy-discussion/2014-October/071455.html

@ovillellas Thoughts?

I just raised issues with the same points as done here, but in addition a bit of a roadmap on how to implement it: see #8810 (and (#8811)

Implemented.