suggestion: set current ax to g.ax_joint at end of jointplot

[jhncls]

Tested with seaborn 0.11.1

jointplot (or JointGrid in general) leaves ax_marg_y as current axis. Some people try to use plt.xlim() or similar and are changing the marginal subplot instead of the main one. It could be nice to automatically set plt.sca(g.ax_joint) when finishing the jointplot.

Example code:

import matplotlib.pyplot as plt
import seaborn as sns

tips = sns.load_dataset('tips')
g = sns.jointplot(data=tips, x='total_bill', y='tip')
# plt.sca(g.ax_joint)
plt.axhline(tips['tip'].mean(), ls='--', c='r')
plt.axhspan(tips['tip'].quantile(.25), tips['tip'].quantile(.75), color='y', alpha=.3)
plt.show()

[mwaskom]

I think this is a good idea!

[mwaskom]

@jhncls would you like to do a PR? Should be a simple code change, but it would be good to have a small test and a mention in the release notes.

[jhncls]

OK, I'll look into it. At first sight the only code change would be either setting plt.sca() at the end of JointGrid's __init__. Or simply change the order in which the three axes are created (and share in the opposite direction: ax_joint = f.add_subplot(gs[1:, :-1], sharex=ax_marg_x, sharey=ax_marg_y)). Just changing the order might be the simplest.

Should the function plot_marginals also save the current axis and set it back? By the way, a simple test of a non-seaborn function called with plot_marginals isn't trivial. Also, there don't seem to be a lot of practical use cases.

[mwaskom]

I think the easiest change would be to do it at the end of jointplot . People who are using JointGrid directly are probably less likely to get bitten by this. That said, calling plt.gca at the end of the JointGrid constructor is probably fine too; but I would worry about the unintended consequences of changing the direction of axis sharing so I'd be -1 on the second proposal.

By the way, a simple test of a non-seaborn function called with plot_marginals isn't trivial.

A simple case would be:

g = sns.JointGrid(data=penguins, x="bill_length_mm", y="bill_depth_mm")
g.plot_marginals(plt.hist)

But yes, without an established API pattern for "which axis to draw the distribution on", it's harder to be generic in that function.

[jhncls]

It probably would make sense to do the same for ax_heatmap of sns.clustermap. That way e.g. plt.xlabel('features') would affect the main subplot instead of the colorbar.