Interaction between xtick label size and height changes FacetGrid ticks

[billbrod]

I've found what seems to be a weird bug where some interaction between the height argument of FacetGrid and the rcParam xtick.labelsize can result in xtick labels being visible on the upper facets of a multi-row FacetGrid, like so:

The upper facets should not have any xlabels on them, yet they do.

Code to reproduce this example:

1. Generate data

N = 6
x = np.linspace(0, 10)
x = np.vstack(N*[x])
A = np.random.rand(N) * np.eye(N)
b = np.random.rand(N, 1)
y = A@x+b

i = np.array([i*np.ones_like(x[i]) for i in range(N)]).flatten()
data = {'x': x.flatten(), 'y': y.flatten(), 'i': i}
tmp = pd.DataFrame(data)

2. Create plot above.

with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 10}):
    g = sns.FacetGrid(tmp, col='i', col_wrap=3, height=2.2)
    g.map(plt.plot, 'x', 'y')

Changing either height or the xtick.labelsize can change the behavior:

3. Create proper plot:

with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 10}):
    g = sns.FacetGrid(tmp, col='i', col_wrap=3, height=2.5)
    g.map(plt.plot, 'x', 'y')

4. Create another proper plot:

with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 12}):
    g = sns.FacetGrid(tmp, col='i', col_wrap=3, height=2.2)
    g.map(plt.plot, 'x', 'y')

5. Mess up a plot using textsize 12:

with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 12}):
    g = sns.FacetGrid(tmp, col='i', col_wrap=3, height=2.6)
    g.map(plt.plot, 'x', 'y')

Weirdly that last one looks different when I save it as a png or svg vs view it in a notebook (where 4 and 6 are visible on the upper facets, but not 10).

This is all on Ubuntu 18.04, with seaborn version 0.11.0 and matplotlib version 3.0.3 (svg backend)

[billbrod]

Actually, going back through my example, it appears the above is not sufficient to reproduce it. You'd also need to run the following: plt.style.use({'figure.subplot.right': .96, 'figure.subplot.left': .075})

Which makes me think this issue might be too niche to be of real concern. But it is weird -- how does seaborn determine whether to show the xticks / xticklabels on the upper facets?

[mwaskom]

I think that if this is a bug (and it seems to be?) it's probably a bug at the matplotlib layer.

As far as I recall (the FacetGrid code is mostly quite old) it's not doing anything that related to this. The ticklabels are hidden as part of matplotlib's sharex/sharey logic, and seaborn doesn't do anything explicitly beyond setting that when calling plt.subplots.

With that said, I'm having trouble reproducing using straight matplotlib, e.g.

plt.style.use({'figure.subplot.right': .96, 'figure.subplot.left': .075})
with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 12}):
    f, axs = plt.subplots(2, 3, sharex=True, sharey=True, figsize=(3 * 2.6, 2 * 2.6))
    f.tight_layout()
    for i, ax in enumerate(axs.flat):
        data = tmp[tmp["i"] == i]
        ax.plot(data["x"], data["y"])
        ax.set_title(f"i = {i}")
    plt.setp(axs[:, 0], ylabel="y")
    plt.setp(axs[-1, :], xlabel="x")
    f.tight_layout()
    sns.despine()

but the problem seems very narrow so it make take some futzing to get the order of operations exact...

[mwaskom]

FWIW mapping a plotting function doesn't seem required:

plt.style.use({'figure.subplot.right': .96, 'figure.subplot.left': .075})
with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 12}):
    g = sns.FacetGrid(tmp, col='i', col_wrap=3, height=2.6)

reproduces it (slightly differently)

[mwaskom]

Actually ... just noticed that your example uses col_wrap which actually does do a little bit extra: https://github.com/mwaskom/seaborn/blob/master/seaborn/axisgrid.py#L464

[billbrod]

It does seem to be something with col_wrap, since the problem doesn't happen if you remove it:

N = 6
x = np.linspace(0, 10)
x = np.vstack(N*[x])
A = np.random.rand(N) * np.eye(N)
b = np.random.rand(N, 1)
y = A@x+b

i = np.array([i*np.ones_like(x[i]) for i in range(N)]).flatten()
data = {'x': x.flatten(), 'y': y.flatten(), 'i': i, 'j': i//2, 'k': np.round((i/2)+.1)-i//2}
tmp = pd.DataFrame(data)

plt.style.use({'figure.subplot.right': .96, 'figure.subplot.left': .075})
with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 12}):
    g = sns.FacetGrid(tmp, col='j', row='k', height=2.6)

It looks like the main difference is that when col_wrap is not None, sharex and sharey are set to the first axis, whereas when it's None, they're set to True.

That said, I am unable reproduce the problem in a similar way to how col_wrap operates using pure matplotlib

plt.style.use({'figure.subplot.right': .96, 'figure.subplot.left': .075})
with sns.axes_style('white'), sns.plotting_context('notebook', rc={'xtick.labelsize': 12}):
    f = plt.figure(figsize=(3*2.6, 2*2.6))
    axs = np.empty(6, object)
    axs[0] = f.add_subplot(2, 3, 1)
    for i in range(1, 6):
        axs[i] = f.add_subplot(2, 3, i+1, sharex=axs[0], sharey=axs[0])
    axs = axs.reshape((2, 3))
    f.tight_layout()
    for i, ax in enumerate(axs.flat):
        data = tmp[tmp["i"] == i]
        ax.plot(data["x"], data["y"])
        ax.set_title(f"i = {i}")
    plt.setp(axs[:, 0], ylabel="y")
    plt.setp(axs[-1, :], xlabel="x")
    f.tight_layout()
    for ax in axs[0, :]:
        for label in ax.get_xticklabels():
            label.set_visible(False)
        ax.xaxis.offsetText.set_visible(False)
    for ax in axs[:, 1]:
        for label in ax.get_yticklabels():
            label.set_visible(False)
        ax.yaxis.offsetText.set_visible(False)
    for ax in axs[:, 2]:
        for label in ax.get_yticklabels():
            label.set_visible(False)
        ax.yaxis.offsetText.set_visible(False)
    sns.despine()

Finally: I dropped some print statements in that if sharex block you linked, and it appears like, at that point, not all xtick labels get returned by ax.get_xticklabels(), and so the label.set_visible(False) ignores the ones that remains visible (10 in the examples above). It's unclear to me why that would be the case though....

[mwaskom]

I'm finding that I can't reproduce this on current versions of things, so I am going to close as a weird upstream problem. Ping if you run into it again.

[donok1]

Hi @mwaskom,

I recently updated seaborn from v.010 and found this problem again, but strangely only when using sns context "poster".

Here the example bothering me:

import seaborn as sns
import pandas as pd

sns.set(style="whitegrid", palette="Pastel2", context="poster")

# initialise some data 
data = {'X':[2019, 2020, 2021,2019, 2020, 2021,2019, 2020, 2021,2019, 2020, 2021,2019, 2020, 2021,2019, 2020, 2021,2019, 2020, 2021,2019, 2020, 2021],
        'Y':['AAA', 'AAA', 'AAA', 'AAA','AAA', 'AAA', 'BBB', 'BBB','BBB','BBB','BBB','BBB','CCC','CCC','CCC','CCC','CCC','CCC','DDD','DDD','DDD','DDD','DDD','DDD'],
        'H':['B2','B1','B2','B1','B2','B1','B2','B1','B2','B1','B2','B1','B2','B1','B2','B1','B2','B1','B2','B1','B2','B1','B2','B1'],
        'Z':[20, 21, 19, 18,20, 21, 19, 18,20, 21, 19, 18, 18,20, 21, 19, 18, 18,20, 21, 19, 18,20,21]}
df = pd.DataFrame(data)
 
g = sns.FacetGrid(df, col='Y', 
                  col_wrap=2, 
                 )
g.map(sns.pointplot, 
      'X', 'Z', 'H',
      order=[2019, 2020, 2021],
      hue_order=['B1','B2'],
      palette='Pastel2',
      ci=None);

g.set(xlabel=None,ylabel=None);

The last xtick label, here "2021" is still visible on the upper graphs.

But for other context it disappears. Here with context = talk:

This probably comes from some strange issue with the size formating, but I don't know matplotlib enought to work it out.
Thanks for work on seaborn !

Versions used:

Python 3.9.5
sns.__version__ = 0.11.1
matplotlib.__version__ = 3.4.2

[mwaskom]

Thanks @donok1 that's reproducible in my dev environment too. I really have no clue what might be causing this, but here is an experiment that may shed some insight:

• Add the following line to the end of the script to show the full extent that the ticklabel objects are occupying:

plt.setp(g.axes[-1].get_xticklabels(), backgroundcolor=(1, 0, 0, .5))

Play around with font_scale in sns.set to get a gradual change in font size. The threshold for the problem is about font_scale=1.906. And that is just about where the extent boxes start to overlap the actual text of their adjacent lablels. It's not perfect, so this may be a red herring. But it's better than being totally clueless?

[donok1]

Thanks @mwaskom for the great tip.
It also works when I tweak the size of the graphs with height directly in FacetGrid. That will do well enough for now.
And for sure, any think you know is better than being clueless ! :)