Add plot option to transpose layouts

[philippjfr]

As requested in #909 this adds a plot option to transpose layouts. It was trivial to implement and provides useful functionality.

[jbednar]

Thanks so much!

[jlstevens]

Would be good to say the normal behavior is scanline order (left to right, top to bottom) and that transposing makes it work top to bottom and left to right.

[philippjfr]

Sounds good, I'll make sure to add some tests too.

[jlstevens]

Other than a suggestion for improving the docstring, this looks good and I'm happy to merge.

[jbednar]

I can't tell from the code -- will this make the subfigure labels transpose as well? Presumably we'd want them to, because if the ordering in the layout has any meaning, the layout's order will now be top to bottom instead of left to right.

[philippjfr]

I can't tell from the code -- will this make the subfigure labels transpose as well? Presumably we'd want them to, because if the ordering in the layout has any meaning, the layout's order will now be top to bottom instead of left to right.

True, currently not the case, it simply labels as normal now.

[jbednar]

I can't tell from the code -- will this make the subfigure labels transpose as well? Presumably we'd want them to, because if the ordering in the layout has any meaning, the layout's order will now be top to bottom instead of left to right.

True, currently not the case, it simply labels as normal now.

It's a start, then, but changing the label order seems dangerous to put off until later, because if we ever do implement that, a previously labeled transposed figure will no longer match its caption.

[philippjfr]

Now handles the subfigure labels too. Ready to merge when tests pass.

[jlstevens]

All tests are green. Merging.

[jbednar]

Sorry I didn't get a chance to test this earlier, but I think this is a problem: the cols() method now seems to determine the number of rows, which is very unintuitive:

Shouldn't cols() still determine the number of columns? Being able to determine the number of rows is nice, but would seem to require having a method other than cols().

The code in that example is from https://github.com/bokeh/colorcet/blob/master/docs/index.ipynb , where the definition of colormaps is replaced with the above.

[philippjfr]

Shouldn't cols() still determine the number of columns? Being able to determine the number of rows is nice, but would seem to require having a method other than cols().

It could do, but it's not really a transpose in that case, and the implementation would be a lot uglier.

[jbednar]

Drat. If you look at that example notebook as updated with the new colormaps function above, you'll see that the transpose option doesn't really work, then -- the idea was to have a fixed number of columns, but to start to fill them top to bottom, not left to right. As it is now, sometimes there are two columns, sometimes 50, which doesn't work from a layout perspective because the width of a web page is fixed:

I.e., a fixed number of columns is about the only thing that makes sense on a webpage that scrolls vertically; cramming in all the plots left to right doesn't seem like it would normally be useful in practice.

[jbednar]

It can be faked by having setting the cols() to a fraction of the length of the layout, but that's very fragile.

[philippjfr]

It can be faked by having setting the cols() to a fraction of the length of the layout, but that's very fragile.

True, it depends on the layout not changing length. I'll investigate what the alternative looks like.

[jlstevens]

I'm wondering if the (Generic)LayoutPlot could use a bit of modulo arithmetic on the number of elements to convert the cols value appropriately (if transposed).

[jbednar]

This definition of cols seems to work fine:

def colormaps(named_cms,cols=3,array=xs):
    images = [hv.Image(array, group=name)(style=dict(cmap=cm)) 
              for name,cm in named_cms]
    return hv.Layout(images)(plot=dict(transpose=True)).display('all').cols(int(np.ceil(len(images)*1.0/cols)))

So is it possible to stick that when the columns value is used, so that it uses that version instead of the cols number directly, if it's transposed?

[jlstevens]

So is it possible to stick that when the columns value is used, so that it uses that version instead of the cols number directly, if it's transposed?

Right. That is what I was suggesting...and now you've shown a concrete implementation, I can't see why the plotting class can't use similar logic.

[philippjfr]

I can't see why the plotting class can't use similar logic.

It probably can easily enough, I'll look into it.

[philippjfr]

One annoyance is that we have the current Layout logic encoded in Layout.shape and NdLayout.shape. I'd vote for either deprecating those or actually allow transposing at the object level, e.g. by extending the cols method.

[jlstevens]

I don't think Layout.shape is public API so I am happy to see it changed. Changing cols is a different matter though but I am open to any backwards-compatible approach.