Best practices with plot_kde

[rpgoldman]

Short Description

I was trying to redo some of the plot_kde plots for the "rat tumor" example of Chapter 5 BDA3, building on the notebooks in the PyMC3 resources repo. I had a couple of problems, that may be more documentation issues than anything else:

1. It seems that one must label the x and y axes in a separate call -- there doesn't seem to be any way to specify the x and y axis labels in the call to plot_kde. This seems normal enough to want to do that it should be supported. Perhaps it's supported by one of the keywords arguments, but I couldn't figure out how and the examples in the docs weren't sufficient to figure it out.

2. I tried to use the contour_kwds argument to set the number of contours, but that didn't seem to work. Again, an example would have helped.

I had two less specific issues, that may be related:

1. In general I couldn't easily figure out how to use the extra keyword arguments, and I couldn't find examples.

2. The ArviZ plots I did ended up looking significantly uglier than the Seaborn plots that were in the old version of the notebook. I could use some help figuring out how to style the ArviZ plots.

Code Example or link

From the diagram, here is the seaborn contour plot:

and here is the best ArviZ equivalent I could come up with. It's not very impressive:

The seaborn code is as follows:

fig, ax = plt.subplots(figsize=(8, 7))
sns.kdeplot(np.log(alpha_trunc /  beta_trunc), np.log(alpha_trunc + beta_trunc),
            cmap=plt.cm.viridis, n_levels=11, ax=ax)
# ax.set_xlim(1.5, 2.1)
ax.set_ylim(1.7, 3.75)
ax.set_xlabel('log(alpha / beta)')
ax.set_ylabel('log(alpha + beta)');

and here's what I came up with as an ArviZ alternative:

fig, ax = plt.subplots(figsize=(8, 7))

az.plot_kde(values2=idata_10.posterior['log(alpha + beta)'], values=idata_10.posterior['log(alpha / beta)'], ax=ax,
                contour=True, 
                contour_kwargs={'levels': 10,}) # Setting different levels here does not have any effect.
    
# ax.set_xlim(1.5, 2.1)
ax.set_ylim(1.7, 3.75)
ax.set_xlabel('log(α / β)')
_ = ax.set_ylabel('log(α + β)')

I'm not sure why it looks so bad. Possibly seaborn's style sheet is messing up the ArviZ plot but... I don't know how styling works with ArviZ. There's an example in the gallery that uses az.style, but az.style does not appear in the online API docs page.

I cargo-culted in az.style.use("arviz-darkgrid") from the ArviZ gallery, but it didn't change my grayscale contour plot.

For full details of what I was trying to do, see my fork of BDA3.

ArviZ version: master at 8f1677cfd985b2b553dc52571943a830adc29d1c

[aloctavodia]

Hi @rpgoldman , thanks for the report.

The grayscale is weird because Matplotlib's default cmap is viridis, I guess you change it somewhere and thus you are seeing a gray scale.

Below there are a few examples of customizing the plots. Probably we should add a few to the gallery and/or docstring.

az.style.use("arviz-darkgrid")
data = {"a":np.random.normal(0, 1, size=1000), "b":np.random.normal(0, 2, size=1000)}

az.plot_kde(values2=data["a"], values=data["b"], 
                    contour=True, fill_last=False) 

az.plot_kde(values2=data["a"], values=data["b"],
                contour=True, fill_last=False, contour_kwargs={"alpha":0})

az.plot_kde(values2=data["a"], values=data["b"],
            contour=True, fill_last=False, 
            contour_kwargs={"colors":None, "cmap":plt.cm.viridis},
            contourf_kwargs={"alpha":0});

az.plot_kde(values2=data["a"], values=data["b"],
            contour=True, fill_last=False, 
            contour_kwargs={"colors":None,"cmap":plt.cm.viridis,"levels":2},
            contourf_kwargs={"alpha":0.5, "levels":2});

[canyon289]

@rpgoldman is this closed?