fix pointplot(native_scale=True, ...)

[v4hn]

Just ran into this plotting something.

Sorry if this should be done cleaner.
I removed the backward compatible code because it failed with working native_scale.

[mwaskom]

Please supply an example that reproduces the problem you ran into. This works fine for me:

import seaborn as sns
tips = sns.load_dataset("tips")
sns.pointplot(tips, x="size", y="total_bill", native_scale=True)

[v4hn]

I hit this in the middle of designing a plot combining lineplot and barplot and eventually moved to manual x axis assignment for individual plots, so I cannot reproduce this in my original context anymore.

Note that your example cannot illustrate the effect native_scale has if I understood it correctly because the values are evenly spaced.
Playing around with your example a bit I can at least produce the issue below, which might reflect my previous situation. I'm not sure whether you consider this a user error or a bug on the framework side:

import seaborn as sns
import matplotlib.pyplot as plt
tips = sns.load_dataset("tips")
tips['ssize'] = tips['size'].apply(lambda x: x**4).astype('category')
tips['ssize'] = tips['ssize'].cat.reorder_categories(sorted(tips['ssize'].cat.categories), ordered=True)
sns.pointplot(data=tips, x='ssize', y='total_bill', native_scale=True, legend=False)

In the example I would have expected this result, which I can also get by explicitly converting the category to float:

I believe my original example was slightly more convolved with grouped categories though if that can make a difference in interpretation.

I also believe my current patch is insufficient to address the issue as I see errors when I do not specify legend.

[mwaskom]

I think your issue is perhaps just that the "native scale" for a series with a category dtype is a categorical representation, not a numeric one. I believe that is consistent and shouldn't change.

[v4hn]

That does make sense. Thank you for the clarification!

As it does not seem to be a reasonable/useful function call, I would have appreciated a warning when native_scale=True is specified with a categorical variable. Something like

native_scale=True has no effect on categorical variables.
Did you intend a different dtype?

[mwaskom]

I could see how that would be helpful but I'm generally opposed to issuing warnings where the user isn't doing something unambiguously "wrong", otherwise they can get very noisy and annoying.

[v4hn]

I do not quite agree as someone who tripped over it, but the argument is valid.

Last suggestion before close then:

The documentation for the parameter (in at least pointplot and boxplot) states

When True, **numeric or datetime values on the categorical axis** will maintain
their original scaling rather than being converted to fixed indices.

I would suggest to rephrase this to indicate that the dtype of the categorical axis must not be category for this. Something like:

When True and dtype of the categorical axis is numeric or datetime,
the categorical axis will maintain the original scaling rather than convert to even spacing.

[mwaskom]

If you're trying to emphasize "dtype" then I'm not sure that's accurate: seaborn will treat series that have an object dtype but numeric values as numeric, it's only the categorical dtype that is interpreted as an explicit request for categorical treatment.

[v4hn]

Yes, I tried. Do you have a more accurate suggestion? :)

With this insight my personal (user) intuition is that dtype "category" should be treated the same as dtype "object" with regards to the native_scale parameter, but you already stated you disagree.

Another suggestion:

When True, numeric or datetime values on the categorical axis will maintain
their original scaling rather than being converted to fixed indices.
The native scale of a Series with the explicit category dtype always uses fixed indices.