Python: Leave Numpy NaN values from matplotlib heatmap and its legend

Question

I have a numpy array that I need to plot as a heatmap. The numpy array would also contain NaN values which I need to exclude from plotting. I was told in other posts that numpy automatically masks the NaN values in the plot, but its somehow not working for me. Here is a sample code

column_labels = list('ABCDEFGH')
row_labels = list('WXYZ')
fig, ax = plt.subplots()
data = np.array([[ 0.96753494,  0.52349944,  0.0254628 ,  0.5104103 ],
         [ 0.07320069,  0.91278731,  0.97094436,  0.70533351],
         [ 0.30162006,  0.49068337,  0.41837729,  0.71139215],
         [ 0.19786101,  0.15882713,  0.59028841,  0.06242765],
         [ 0.51505872,  0.07798389,  0.58790067,  0.44782683],
         [ 0.68975694,  0.53535385,  0.15696023,  0.35641951],
         [ 0.66481995,  0.03576846,  0.9623601 ,  0.96006395],
         [ 0.45865404,  0.50433582,  0.18182575,  0.35126449],])

data[3,:] = np.nan
heatmap = ax.pcolor(data, cmap=plt.cm.seismic)

fig.colorbar(heatmap)
# put the major ticks at the middle of each cell
ax.set_xticks(np.arange(data.shape[1])+0.5, minor=False)
ax.set_yticks(np.arange(data.shape[0])+0.5, minor=False)

# want a more natural, table-like display
ax.invert_yaxis()
ax.xaxis.tick_top()

ax.set_xticklabels(row_labels, minor=False)
ax.set_yticklabels(column_labels, minor=False)
plt.show()

The plots looks like

Clearly this is very different from the plot without the Nan which looks like

I want to avoid the NaN values completely from the legend and preferably mark it with some symbol such as X. How can I achieve the same?

score 18 · Accepted Answer · edited May 23 '17 at 11:52

The nans interfere with pcolor determining the range of values contained in data since

In [72]: data.min(), data.max()
Out[72]: (nan, nan)

You can work around the problem by declaring the range of values yourself using np.nanmin and np.nanmax to find the minimum and maximum non-NaN values in data:

heatmap = ax.pcolor(data, cmap=plt.cm.seismic, 
                    vmin=np.nanmin(data), vmax=np.nanmax(data))

since

In [73]: np.nanmin(data), np.nanmax(data)
Out[73]: (0.025462800000000001, 0.97094435999999995)

import numpy as np
import matplotlib.pyplot as plt

column_labels = list('ABCDEFGH')
row_labels = list('WXYZ')
fig, ax = plt.subplots()
data = np.array([[ 0.96753494,  0.52349944,  0.0254628 ,  0.5104103 ],
         [ 0.07320069,  0.91278731,  0.97094436,  0.70533351],
         [ 0.30162006,  0.49068337,  0.41837729,  0.71139215],
         [ 0.19786101,  0.15882713,  0.59028841,  0.06242765],
         [ 0.51505872,  0.07798389,  0.58790067,  0.44782683],
         [ 0.68975694,  0.53535385,  0.15696023,  0.35641951],
         [ 0.66481995,  0.03576846,  0.9623601 ,  0.96006395],
         [ 0.45865404,  0.50433582,  0.18182575,  0.35126449],])

data[3,:] = np.nan
heatmap = ax.pcolor(data, cmap=plt.cm.seismic, 
                    vmin=np.nanmin(data), vmax=np.nanmax(data))
heatmap.cmap.set_under('black')

bar = fig.colorbar(heatmap, extend='both')

# put the major ticks at the middle of each cell
ax.set_xticks(np.arange(data.shape[1])+0.5, minor=False)
ax.set_yticks(np.arange(data.shape[0])+0.5, minor=False)

# want a more natural, table-like display
ax.invert_yaxis()
ax.xaxis.tick_top()

ax.set_xticklabels(row_labels, minor=False)
ax.set_yticklabels(column_labels, minor=False)
plt.show()

Another option (based on Joe Kington's solution) would be to draw rectangular patches with hatch marks wherever data is NaN.

The above example shows that pcolor colors in cells with NaN values as though the NaNs were very negative numbers. In contrast, if you pass pcolor a masked array, pcolor leaves the masked areas transparent. Thus, you can draw hatches on the axes background patch, ax.patch, to show hatch marks on the masked areas.

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches

column_labels = list('ABCDEFGH')
row_labels = list('WXYZ')
fig, ax = plt.subplots()
data = np.array([[ 0.96753494,  0.52349944,  0.0254628 ,  0.5104103 ],
         [ 0.07320069,  0.91278731,  0.97094436,  0.70533351],
         [ 0.30162006,  0.49068337,  0.41837729,  0.71139215],
         [ 0.19786101,  0.15882713,  0.59028841,  0.06242765],
         [ 0.51505872,  0.07798389,  0.58790067,  0.44782683],
         [ 0.68975694,  0.53535385,  0.15696023,  0.35641951],
         [ 0.66481995,  0.03576846,  0.9623601 ,  0.96006395],
         [ 0.45865404,  0.50433582,  0.18182575,  0.35126449],])

data[3,:] = np.nan
data = np.ma.masked_invalid(data)

heatmap = ax.pcolor(data, cmap=plt.cm.seismic, 
                    vmin=np.nanmin(data), vmax=np.nanmax(data))
# https://stackoverflow.com/a/16125413/190597 (Joe Kington)
ax.patch.set(hatch='x', edgecolor='black')
fig.colorbar(heatmap)

# put the major ticks at the middle of each cell
ax.set_xticks(np.arange(data.shape[1])+0.5, minor=False)
ax.set_yticks(np.arange(data.shape[0])+0.5, minor=False)

# want a more natural, table-like display
ax.invert_yaxis()
ax.xaxis.tick_top()

ax.set_xticklabels(row_labels, minor=False)
ax.set_yticklabels(column_labels, minor=False)
plt.show()

If you have wish to use more than one type of hatch mark, say one for NaNs and another for negative values, then you could use a loop to add the hatched Rectangles:

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches

column_labels = list('ABCDEFGH')
row_labels = list('WXYZ')
fig, ax = plt.subplots()
data = np.array([[ 0.96753494,  0.52349944,  0.0254628 ,  0.5104103 ],
         [ 0.07320069,  0.91278731,  0.97094436,  0.70533351],
         [ 0.30162006,  0.49068337,  0.41837729,  0.71139215],
         [ 0.19786101,  0.15882713,  0.59028841,  0.06242765],
         [ 0.51505872,  0.07798389,  0.58790067,  0.44782683],
         [ 0.68975694,  0.53535385,  0.15696023,  0.35641951],
         [ 0.66481995,  0.03576846,  0.9623601 ,  0.96006395],
         [ 0.45865404,  0.50433582,  0.18182575,  0.35126449],])
data -= 0.5
data[3,:] = np.nan
data = np.ma.masked_invalid(data)
heatmap = ax.pcolor(data, cmap=plt.cm.seismic, 
                    vmin=np.nanmin(data), vmax=np.nanmax(data))

# https://stackoverflow.com/a/16125413/190597 (Joe Kington)
ax.patch.set(hatch='x', edgecolor='black')

# draw a hatched rectangle wherever the data is negative
# http://matthiaseisen.com/pp/patterns/p0203/
mask = data < 0
for j, i in np.column_stack(np.where(mask)):
      ax.add_patch(
          mpatches.Rectangle(
              (i, j),     # (x,y)
              1,          # width
              1,          # height
              fill=False, 
              edgecolor='blue',
              snap=False,
              hatch='x' # the more slashes, the denser the hash lines 
          ))

fig.colorbar(heatmap)

# put the major ticks at the middle of each cell
ax.set_xticks(np.arange(data.shape[1])+0.5, minor=False)
ax.set_yticks(np.arange(data.shape[0])+0.5, minor=False)

# want a more natural, table-like display
ax.invert_yaxis()
ax.xaxis.tick_top()

ax.set_xticklabels(row_labels, minor=False)
ax.set_yticklabels(column_labels, minor=False)
plt.show()

Is there a way to add a symbol like x or something to the nan values — rambalachandran, Mar 10 '16 at 00:43
I added `heatmap.cmap.set_under('black')` to color the NaNs black -- a color not in the seismic colormap. Would that do? — unutbu, Mar 10 '16 at 00:49
Is it possible to add multiple patches, say one for NaN and one for negative values, or is it best to achieve it by applying the masked array twice — rambalachandran, Mar 10 '16 at 04:22
I think the overlay trick can only be used with one hatch. If you need two kinds of hatches, you could add the second using a loop. I've added an example above to show what I mean. — unutbu, Mar 10 '16 at 14:29

Python: Leave Numpy NaN values from matplotlib heatmap and its legend

1 Answers1

Linked