[MRG] BUG Fixes partial dependence bug by setting axes to be invisible

[thomasjpfan]

Reference Issues/PRs

Fixes bug with the title not displaying in examples/plot_partial_dependence_visualization_api.py

Instead of making the axes not visible, this checks if the axes is off to see if the axes was used twice.

What does this implement/fix? Explain your changes.

To see why this check was needed, please look at test_plot_partial_dependence_with_same_axes.

Any other comments?

CC @NicolasHug @amueller @glemaitre

[NicolasHug]

I'm fine with the changes but I don't understand what test_plot_partial_dependence_with_same_axes is supposed to explain

Why aren't we allowed to use the same ax object?

[thomasjpfan]

The expected behavior of using the same axes, is to draw on axes created by the first call. The following example will fail:

tree_disp = plot_partial_dependence(mlp, X, ["LSTAT", "RM"], ax=ax)

# will error
mlp_disp = plot_partial_dependence(mlp, X, ["LSTAT", "RM"], ax=ax)

This is because tree_disp will create two new axes to place on top of the ax passed in. This will result in 3 total axes in the figure. If we want the second call to plot on the axes created by the first call, then it needs to look at the axes passed in, find the axes that may be taking the same space as the original axes, go into the figure object to grab the two axes, and the plot on them.

Currently our api allows:

tree_disp = plot_partial_dependence(mlp, X, ["LSTAT", "RM"], ax=ax)

# will not raise
mlp_disp = plot_partial_dependence(mlp, X, ["LSTAT", "RM"], ax=tree_disp.axes_)

Because the second call gets the axes directly and uses them to plot the new curves.

[NicolasHug]

Thanks. Would be great if test_plot_partial_dependence_with_same_axes explained this (more succinctly)

[thomasjpfan]

Added more details in explaining test_plot_partial_dependence_with_same_axes to this PR, since it mostly related.

[glemaitre]

Added more details in explaining test_plot_partial_dependence_with_same_axes to this PR, since it mostly related.

Are you sure that you added something? I don't see any push since your comment?

[thomasjpfan]

Are you sure that you added something? I don't see any push since your comment?

Oops forgot to push

[glemaitre]

LGTM now :) Thanks @thomasjpfan

[amueller]

I just talked with @tacaswell and he suggested removing the axes from the figure instead of hiding it.

[thomasjpfan]

Initially, I resized the original axes to the size of one of the plots but it was really janky with constraints. I am happy with removing the original axes.

[amueller]

@thomasjpfan the other solution which was suggested was reshaping the axis to be the first axis of the stolen space. But maybe doing the removal for now would be ok?