Allow * to work on an NdLayout (and more)

[nleclercq]

The story began with the following use case: how to overlay a selection box on images attached a to hv.NdLayout?

Implementing such a feature with a hv.Holomap is quite trivial and would only require 5 lines of code:

images = {en:hv.Image(ed) for en, ed in xrarray_dataset.data_vars.items()}
box = hv.streams.BoundsXY(source=hm, bounds=(0,0,0,0))
bounds = hv.DynamicMap(selection_callback, streams=[box])
hv.HoloMap(images, kdims=['img']) * bounds

Accordingly, one would legitimately expect the same approach to work with a hv.NdLayout:

images = {en:hv.Image(ed) for en, ed in xrarray_dataset.data_vars.items()}
box = hv.streams.BoundsXY(source=hm, bounds=(0,0,0,0))
bounds = hv.DynamicMap(selection_callback, streams=[box])
hv.NdLayout(image).cols(2) * bounds

Unfortunately, with holoviews <= 1.9, the overlay operator * doesn’t work with hv.NdLayout.
The main feature request would consequently be to allow the overlay operator to work with hv.NdLayout.

@philippjfr and @jlstevens already had a short discussion about this feature request on gitter. @jlstevens suggested to extend support to the hv.Layout. Here is the main part of that discussion:

@philippjfr:
a feature request to allow * to work on an NdLayout seems sensible.

@jlstevens:
I'm not sure I recommend that
the code for mul is spread out and pretty impenetrable imho!

@philippjfr:
Yeah, I don't expect it to be too hard, the main issue arises from the fact that NdLayout lives in core/layout which can't import from core/overlay, may have to move it to core/spaces before doing that.

@jlstevens:
one thing though…if * works for hv.NdLayout, why not hv.Layout?
isn't it the same semantics?
which I haven't yet figured out btw
is it a layout * element where the element is broadcast to all the parts of the layout?
or layout * layout where you overlay pairwise....though that if they have different lengths?

@philippjfr:
I don't have any major objection to that, the difference is that hv.NdLayout is a homogeneous collection so wanting to overlay something on all the components is more common.
I wouldn't go so far as support hv.layout * hv.layout

@jlstevens:
I'm just saying that I don't see why to make hv.NdLayout a special case if the semantics also apply equally to hv.Layout
I have mixed feelings then... if we want * on layout I would like it to be as general as * everywhere else

@philippjfr:
I have no objection to hv.layout * hv.element, matching up potentially heterogeneous collections of things as in layout * layout is not a rabbithole I necessarily want to go down.

@jlstevens:
I think hv.layout * hv.layout sort of makes sense... you overlay the zip portion and append what is left from the longer one

@philippjfr:
I suppose, can't see myself wanting that very much but those semantics are easy to implement at least.

@jlstevens:
hv.Layout([(el1 * el2) for el1, el2 in zip(layout1, layout2) ]) + hv.Layout(remaining_items)
it isn't something I would necessarily encourage or use much myself
but I think that is conceptually simple enough

[jbednar]

Seems like a reasonable request with clear semantics to me, more or less like NumPy broadcasting for elementwise operations. I wouldn't want to have to implement it, though!

[philippjfr]

Looking back at this I have to say I'm not a huge fan of implementing this for Layout, since they are heterogeneous collections, while NdLayout is a clear cut homogeneous collection where this should be supported for consistency with GridSpace, which already works.

[github-actions]

This issue has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.