Issues faceting violinplots / boxplots

[sjobeek]

I ran into a KeyError while attempting to facet a single violinplot into each plot of a FacetGrid.

Here is some short code that replicated the problem:

df = pd.DataFrame(dict(one=[1,2,4,2,3], two=[2,2,4,4,4], three=[1,1,1,2,2]))
fg = sns.FacetGrid(df, col='three')
fg.map(sns.boxplot, 'one')

Supplying a second parameter ('groupby') results in the proper plots, even if there is only a single group.

I'd expect the single plot to simply work. Perhaps violinplot/boxplot is expecting a continuous index if given a dataframe?

[sjobeek]

There also seems to be some other problems with the locations of the labels, and the actual grouping of the data when faceting.

While a single plot works as expected:

df = pd.DataFrame(dict(one=[1,2,4,2,1,3], two=[1,1,2,2,4,4], three=[1,1,1,1,2,2]))
sns.violinplot(df.one, df.two)

a FacetGrid with the same data will fail to group it or apply the labels where I expect them (one / two are switched):

fg = sns.FacetGrid(df, col='three')
fg.map(sns.violinplot, 'one', 'two')

And finally, inner='sticks' doesn't appear to actually draw any sticks when I've tried to use it, though 'points' is working properly.

[mwaskom]

Yeah, basically the dilemma here is that I either have to break the API of voilinplot/boxplot or they don't get to work with FacetGrid. I punted on this decision for 0.3 (although there is a kind="box" in factorplot that somewhat gets around the issue (example here).

[sjobeek]

I've also noticed that violinplots run into an LinAlgError: singular matrix exception if any single plot has several instances of a single value.

This means I can't use it to compare a large collection of relatively sparse datasets that may or may not have a single "facet" with repeat values. Obviously it's not possible to actually calculate a KDE if there is 0 variation, but it would be nice if we could at least continue to plot the rest of the data without completely failing.

I find violinplots with inner='points' to be quite useful for visually comparing large sets of data with just 3-5 datapoints each. Unfortunately this bug makes it unfeasible for many datasets which have a small chance of seeing zero variation for a single facet.

[harper357]

I am also running into the LinAlgError: singular matrix exception. Is there any hope of a fix/work around/exception for it? It would be nice if, like sjobeek said, it could continue to plot the rest of the data.

[mwaskom]

What version of seaborn do you have? I'm pretty sure I've fixed this.

[harper357]

I have 0.5.0.

[mwaskom]

It's possible you are seeing something else, then, this works for me:

[harper357]

I tried repeating your last post and I got the same thing you did. The problem I am having, and maybe it is actually something different, is if you have only have one value repeated. So
data = [np.random.randn(100), [.83,.83,.83]]
will give the error.

[mwaskom]

Can you make a separate issue with a self-contained example so it's easier to keep track of?

[mwaskom]

Closing with #410