一张 seaborn 热图中的两个范围/色阶

Question

I was doing quite some research, but couldn't find a satisfying solution yet.

I'm trying to build a heatmap using seaborn. As my dataset is a bit volatile in a lower range (0-20) but reaches up to 7000 using only one color-scale for all of the data doesn't allow a good graphical interpretation. That's why I thought about using two scales, two different color-spectrums.

I would like to merge those two heatmaps into one:

What works so far is that I get both axes (scales) displayed in my plot, but when it comes to plotting the data only the last active axis is taken into account. The upper range is not considered.

2

I have also tried to split the data set according to the spectrum, but this was not working. Here is my code:

df = pd.DataFrame(merged, classes)


vmax = np.amax(merged)

ax1 = sns.heatmap(df, vmin = 25, vmax = vmax, cmap = "crest", )
ax2 = sns.heatmap(df, vmin = 0, vmax = 25, cmap = "flare", )

plt.xlabel("Time")
plt.ylabel("Method")

plt.show()

Answer 1

If you want to plot both parts together, you can update the colormaps such that the "over" and the "under" color are set to invisible ("none"). The "over" color is used for values that are larger than vmax . The "under" color is used for values below vmin .

Optionally, you can set negative padding for the first colorbar to have both colorbars closer together.

import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
from scipy.ndimage.filters import gaussian_filter

# first create some test data
data = gaussian_filter(np.random.rand(50, 50), sigma=5)
data -= data.min()
data *= 2 / data.max()
data = np.where(data < 1, data*25, (data ** 3) * 25)

# adapt the colormaps such that the "under" or "over" color is "none"
cmap1 = plt.get_cmap('crest').copy()
cmap1.set_under('none')
cmap2 = plt.get_cmap('flare').copy()
cmap2.set_over('none')

ax1 = sns.heatmap(data, vmin=25, cmap=cmap1, cbar_kws={'pad': -0.02})
sns.heatmap(data, vmin=0, vmax=25, cmap=cmap2, ax=ax1)

plt.show()

Answer 2

You need to create the two Axes to plot beforehand, and then use the ax argument when calling sns.heatmap in order to tell seaborn which Axes should have which colormap.

Example (using mock data):

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

values = np.random.uniform(0, 7000, size=(50, 50))
vmax = np.amax(values)

fig, (ax1, ax2) = plt.subplots(ncols=2)

sns.heatmap(values, ax=ax1, vmin=25, vmax=vmax, cmap="crest")
sns.heatmap(values, ax=ax2, vmin=0, vmax=25, cmap="flare")

for ax in (ax1, ax2):
    ax.set_xlabel("Time")
    ax.set_ylabel("Method")

plt.show()

Resulting figure:

Answer 3

You can create a Colormap class that uses both other Colormaps

import seaborn as sns
import matplotlib as mpl
import matplotlib.pyplot as plt

class SplitCMap(mpl.colors.Colormap):
  def __init__(self, name, vmin, vsplit, vmax, N=256):
    super().__init__(name, N)
    self.lowcmap = mpl.cm.get_cmap('flare')
    self.highcmap = mpl.cm.get_cmap('crest')
    self.split_level = (vsplit-vmin) / (vmax-vmin)
    self.scale_low = 1.0 / self.split_level
    self.scale_high = 1.0 / (1.0 - self.split_level)
  def mapcolor(self, v, **kwds):
    if v < self.split_level:
      return self.lowcmap(v * self.scale_low, **kwds)
    return self.highcmap((v-self.split_level)*self.scale_high, **kwds)
  def __call__(self, *args, **kwds):
    if isinstance(args[0], (int, float)):
      self.mapcolor(args[0], **kwds)
    return [self.mapcolor(v, **kwds) for v in args[0] ]

df = pd.DataFrame(merged, classes)

vmax = np.amax(merged)

cmap = SplitCMap('split', 0, 25, vmax)
ax = sns.heatmap(df, cmap=cmap)

plt.xlabel("Time")
plt.ylabel("Method")

plt.show()