I want to plot several "filled" kernel density estimates (KDE) in matplotlib, like the upper halfs of vertical violinplots or a non overlapping version of the cover art of Joy Division's Unknown Pleasures.
Ideally, I want matplotlib to create the density estimates itself, so that I don't have to use scipy's gaussian kde myself.
This answer shows how to modify Matplotlib's violinplots. Those violinplots can also be adapted to only show the upper half of a violin plot.
pos = np.arange(1, 6) / 2.0
data = [np.random.normal(0, std, size=1000) for std in pos]
violins = plt.violinplot(data, positions=pos, showextrema=False, vert=False)
for body in violins['bodies']:
paths = body.get_paths()[0]
mean = np.mean(paths.vertices[:, 1])
paths.vertices[:, 1][paths.vertices[:, 1] <= mean] = mean
A nice looking overlapping variant can be easily created by setting the bodies' transparency to 0, adding an edgecolor and making sure to plot underlying KDEs first:
pos = np.arange(1, 6) / 2
data = [np.random.normal(0, std, size=1000) for std in pos]
violins = plt.violinplot(
data[::-1],
positions=pos[::-1]/5,
showextrema=False,
vert=False,
)
for body in violins['bodies']:
paths = body.get_paths()[0]
mean = np.mean(paths.vertices[:, 1])
paths.vertices[:, 1][paths.vertices[:, 1] <= mean] = mean
body.set_edgecolor('black')
body.set_alpha(1)
Note that there is an existing package called joypy, building on top of matplotlib to easily produce such "Joyplots" from dataframes.
Apart, there is little reason not to use scipy.stats.gaussian_kde because it is directly providing the KDE. violinplot internally also uses it.
So the plot in question would look something like
from scipy.stats import gaussian_kde
import matplotlib.pyplot as plt
import numpy as np
pos = np.arange(1, 6) / 2.0
data = [np.random.normal(0, std, size=1000) for std in pos]
def plot_kde(data, y0, height, ax=None, color="C0"):
if not ax: ax = plt.gca()
x = np.linspace(data.min(), data.max())
y = gaussian_kde(data)(x)
ax.plot(x,y0+y/y.max()*height, color=color)
ax.fill_between(x, y0+y/y.max()*height,y0, color=color, alpha=0.5)
for i, d in enumerate(data):
plot_kde(d, i, 0.8, ax=None)
plt.show()
