Seaborn: displot(kind='ecdf',..., log_scale=True) not working

[JanHomann]

The following line of code gives an error:

sns.displot(kind='ecdf', data=df, x='col_1', log_scale=True)

UserWarning: Data has no positive values, and therefore cannot be log-scaled.

My data is all positive and kind='hist' or 'kde' works just fine.

[mwaskom]

Please share:

• A sample of your data, or a simulated dataset with similar characteristics that produces the error
• The full traceback with the error message
• The versions of matplotlib and seaborn

As requested here.

Thanks!

[JanHomann]

This here works:

sns.displot(kind='ecdf', x=np.random.lognormal(mean=-1, sigma=1, size=1000), log_scale=True);

While this here doesn't (note that the values are smaller (roughly between 10^-6 and 10^-3)):

sns.displot(kind='ecdf', x=np.random.lognormal(mean=-10, sigma=1, size=1000), log_scale=True);

Using hist works:

sns.displot(kind='hist', x=np.random.lognormal(mean=-10, sigma=1, size=1000), log_scale=True);

There is no traceback because it's just a warning and not an error. the plot appears empty (screenshot below).

sns.__version__
'0.11.1'

matplotlib.__version__
'3.3.1'

[mwaskom]

Thanks; I can reproduce.

You can turn a warning into a full stack trace with

import warnings
warnings.simplefilter("error", UserWarning)

when I do that, the problem is in the guts of matplotlib and will take some figuring out.

In the meantime, because the ecdf is just based on quantiles, it doesn't matter if the log scaling happens before or after you calculate it. So you could do

g = sns.displot(..., kind="ecdf")  # No log_scale
g.set(xscale="log")

for the same effect.

[JanHomann]

Actually, I think there is a bigger problem. Even with the numbers that are bigger it doesn't give a reasonable cumulative distribution.

Have a look at this:

sns.displot(kind='ecdf',x=np.random.lognormal(mean=-1,sigma=1,size=1000), log_scale=True);

sns.displot(kind='hist',x=np.random.lognormal(mean=-1,sigma=1,size=1000),
            log_scale=True, cumulative=True,
            element='step', bins=1000);

I think those plots should end up roughly the same. The upper plot is definitely wrong.

[JanHomann]

Sorry, I don't know why my plots are so big. I use a 4K monitor.

[JanHomann]

Your suggested solution (thank you!) gives the correct result. So I think there is generally a problem with displot 'ecdf' and log_scale=True, even when it works. The numbers seem to be transformed in a way that is not correct.

g = sns.displot(kind='ecdf',x=np.random.lognormal(mean=-1,sigma=1,size=1000));
g.set(xscale='log')

[mwaskom]

Yes I think ecdfplot is missing the step that transforms the data back from log space for plotting after calculating the distribution — this is currently a rough edge.

[JanHomann]

Okay, good to know.

So essentially NOONE should be using displot(kind='ecdf', log_scale=True) right now. because even if it works, it gives (or can give) the wrong result. Correct?

[mwaskom]

Indeed, though fortunately the result ends up being fairly obviously wrong visually, as you discovered.

[JanHomann]

Yes. Me and you have enough "common sense" to see that. But there are a lot of people walking around that don't have that. :sweat_smile:.