Obtaining the exact data coordinates of seaborn boxplot boxes

Question

I have a seaborn boxplot (sns.boxplot) on which I would like to add some points. For example, say I have this pandas DataFrame:

[In]  import pandas as pd
      import numpy as np
      import matplotlib.pyplot as plt
      import seaborn as sns

      df = pd.DataFrame({'Property 1':['a']*100+['b']*100,
                         'Property 2': ['w', 'x', 'y', 'z']*50,
                         'Value': np.random.normal(size=200)})
      df.head(3)

[Out]         Property 1    Property 2  Value
           0    a               w       1.421380
           1    a               x      -1.034465
           2    a               y       0.212911
[In]  df.shape
[Out] (200, 3)

I can easily generate a boxplot with seaborn:

[In]  sns.boxplot(x='Property 2', hue='Property 1', y='Value', data=df)
[Out]

Now say I want to add markers for a specific case in my sample. I can get close with this:

[In]   specific_case = pd.DataFrame([['a', 'w', '0.5'],
                                     ['a', 'x', '0.2'],
                                     ['a', 'y', '0.1'],
                                     ['a', 'z', '0.3'],
                                     ['b', 'w', '-0.5'],
                                     ['b', 'x', '-0.2'],
                                     ['b', 'y', '0.3'],
                                     ['b', 'z', '0.5']
                                    ], 
                                    columns = df.columns
                                   )
[In]  sns.boxplot(x='Property 2', hue='Property 1', y='Value', data=df)
      plt.plot(np.arange(-0.25, 3.75, 0.5),
               specific_case['Value'].values, 'ro')
[Out] 

That is unsatisfactory, of course.

I then used this answer that talks about getting the bBox and this tutorial about converting diplay coordinates into data coordinates to write this function:

[In] def get_x_coordinates_of_seaborn_boxplot(ax, x_or_y):
         display_coordinates = []
         inv = ax.transData.inverted()    
         for c in ax.get_children():
             if type(c) == mpl.patches.PathPatch:
                 if x_or_y == 'x':         
                     display_coordinates.append(
                         (c.get_extents().xmin+c.get_extents().xmax)/2)
                 if x_or_y == 'y':
                     display_coordinates.append(
                         (c.get_extents().ymin+c.get_extents().ymax)/2)
         return inv.transform(tuple(display_coordinates)) 

That works great for my first hue, but not at all for my second:

[In]  ax = sns.boxplot(x='Property 2', hue='Property 1', y='Value', data=df)
      coords = get_x_coordinates_of_seaborn_boxplot(ax, 'x')
      plt.plot(coords, specific_case['Value'].values, 'ro')
[Out] 

How can I get the data coordinates of all my boxes?

Answer 1

I'm unsure about the purpose of those transformations. But it seems the real problem is just to plot the points from the specific_case at the correct positions. The xcoordinate of every boxplot is shifted by 0.2 from the whole number. (That is because bars are 0.8 wide by default, you have 2 boxes, which makes each 0.4 wide, half of that is 0.2.) You then need to arrange the x values to fit to those of the specific_case dataframe.

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

df = pd.DataFrame({'Property 1':['a']*100+['b']*100,
                 'Property 2': ['w', 'x', 'y', 'z']*50,
                 'Value': np.random.normal(size=200)})

specific_case = pd.DataFrame([['a', 'w', '0.5'],
                             ['a', 'x', '0.2'],
                             ['a', 'y', '0.1'],
                             ['a', 'z', '0.3'],
                             ['b', 'w', '-0.5'],
                             ['b', 'x', '-0.2'],
                             ['b', 'y', '0.3'],
                             ['b', 'z', '0.5']
                            ],  columns = df.columns )
ax = sns.boxplot(x='Property 2', hue='Property 1', y='Value', data=df)

X = np.repeat(np.atleast_2d(np.arange(4)),2, axis=0)+ np.array([[-.2],[.2]]) 
ax.plot(X.flatten(), specific_case['Value'].values, 'ro', zorder=4)

plt.show()

Answer 2

I got it figured out:

In your code do this to extract the x-coordinate based on hue. I did not do it for the y, but the logic should be the same:

Create two lists holding your x coordinate:

display_coordinates_1=[]
display_coordinates_2=[]

Inside your for loop that starts with:

for c in ax.get_children():

Use the following:

display_coordinates_1.append(c.get_extents().x0) 

You need x0 for the x-coordinate of boxplots under first hue.

The following gives you the x-coordinates for the subplots in the second hue. Note the use of x1 here:

display_coordinates_2.append(c.get_extents().x1) 

Lastly, after you inv.transform() the two lists, make sure you select every other value, since for x-coordinates each list has 6 outputs and you want the ones at indices 0,2,4 or [::2].

Hope this helps.