Setting each diagonal block with different plotting types

[xichenamoy]

Currently for all the diagonal blocks of a triangle plot, we could set types={'diagonal': 'hist'} or {'diagonal': 'kde'} . But sometimes one may want to have different type settings for each diagonal blocks. Could we add a feature that enable users to set the type of each diagonal block separately?

[williamjameshandley]

I think this should be achievable. The main thing to discuss is what the interface should be (@lukashergt may want to weigh in here).

Given that types is already provided as a dictionary, we could simply allow extra keys in this dictionary that would override the 'diagonal' 'upper' and 'lower' choices, for example

types={'diagonal':'hist', 
       'lower':'kde', 
       ('x0','x1'):'fastkde', 
       'x0':'kde'}

would set the x0 on the diagonal to be kde, and the rest of the diagonals to be hist, whilst the lower triangle would be filled with kdes, except the ('x0','x1') block, which would be a fastkde

[lukashergt]

That sort of addition to the dictionary would be possible, I guess.

Another option that possibly already works (haven't tried it though) could be to set up the axes with make_2d_axes (say e.g. with lower and diagonal set to true), then call plot_2d first on all axes with types={'lower': 'kde'} (this will leave the diagonal axes empty for now) and then call plot_2d with types={'diagonal': 'hist'} on the desired subset of axes (e.g. axes.iloc[:1, :1]) and analogously for types={'diagonal': 'kde'}.

[lukashergt]

Here an example of how to make it work with the current set-up. The colours will need manual tweaking. This might seem complicated, but extending the dictionary as proposed above would mean also the handling of kwargs would have to be extended and I fear that might complicate things a lot more.

What do you think, @williamjameshandley, @xichenamoy?

ns1 = NestedSamples(root="./tests/example_data/pc")
ns2 = NestedSamples(root="./tests/example_data/pc")
fig = plt.figure(figsize=(10, 10))
fig, axes = make_2d_axes(['x0', 'x1', 'x2', 'x3'], fig=fig, upper=False)
ns1.plot_2d(axes=axes,              types={'lower': 'kde'},     label='2D kde')
ns1.plot_2d(axes=axes.iloc[1:, 1:], types={'diagonal': 'kde'},  label='kde')
ns1.plot_2d(axes=axes.iloc[:1, :1], types={'diagonal': 'hist'}, label='hist', 
                                    diagonal_kwargs={'histtype': 'step', 'ls': '--'})
axes['x0']['x0'].legend()
axes['x0']['x1'].legend()
axes['x1']['x1'].legend(bbox_to_anchor=(1, 1), loc='lower left')
axes['x1']['x2'].legend(bbox_to_anchor=(1, 1), loc='lower left')

[xichenamoy]

That looks good to me, very nice fix! I think this could fully solve the issue in my plot. Thank you!

[williamjameshandley]

That is a good solution for @xichenamoy for now, although I think this kind of fine-grained control should be incorporated into a full interface.

This might seem complicated, but extending the dictionary as proposed above would mean also the handling of kwargs would have to be extended and I fear that might complicate things a lot more.

A dict-of-dict approach for the additional kwargs, whilst ghastly would certainly accommodate this extension. diagonal_kwargs etc were only introduced last release, I would be amenable to changing it if it allows us to implement the extension

[williamjameshandley]

Alternatively, if we'd rather stick with a similar interface, it would be entirely possible to add auto-generated kwargs like x0_x1_kwargs, where x0 and x1 are the user-specified parameter names, since the kwargs.pop functionality would allow this.

[williamjameshandley]

Resolved by lukas