I see this on my Mac as well.

Interestingly, you get the same segfault for masked values as well, which removes the excuse of "don't use NaN".

It does work if you trim the bad values from x, y, and z, so it seems that tricontour just doesn't like dealing w/ bad data points...

mport numpy as np
import matplotlib.pyplot as plt

x,y=np.meshgrid(np.arange(10),np.arange(10))
z=x**2.+y**2.
z[x<y]=np.nan
x = x.ravel()
y = y.ravel()
z = z.ravel()
ind = x>=y
fig,ax=plt.subplots()
ax.tricontourf(x[ind], y[ind], z[ind])
#ax.tripcolor(x.ravel(),y.ravel(),z.ravel(),vmax=100,vmin=0)
plt.show()

attn @ianthomas23

tricontour and tricontourfexpect the z argument to be an array of finite values, so no np.nan or np.infvalues, and the array cannot be a numpy masked array. Evidently we need to add explicit checks for these and raise an exception instead of allowing a segmentation fault to occur.

There are two ways around this limitation. The first, as suggested by @jklymak, is to only triangulate the points with finite z values. An example code snippet for this, assuming x, y and z are 1D arrays is

point_mask = np.isfinite(z)  # Points to keep.
ax.tricontourf(x[point_mask], y[point_mask], z[point_mask])

The second approach is to use the set_mask method of the Triangulation object, which exists for just this sort of problem where you have a triangulation but want to exclude specific triangles from being rendered. For example

import matplotlib.tri as mtri
triang = mtri.Triangulation(x, y)  # Delaunay triangulation of all points.
point_mask = ~np.isfinite(z)   # Points to mask out.
tri_mask = np.any(point_mask[triang.triangles], axis=1)  # Triangles to mask out.
triang.set_mask(tri_mask)
ax.tricontourf(triang, z, levels=np.arange(0, 151, 25))

Unfortunately, having looked into this I've discovered that you need to specify the contour levels as there is a second bug that non-finite z values are not filtered out before the z limits are determined, leading to incorrect contour levels. This needs a second PR, but is very simple to fix.

Just keeping the finite values is not enough. Sometimes having nans is important. In my case it's an indication that I did not measure at some point. I cannot put some arbitrary number, such as 0, since I did not measure that arbitrary value. If I just get rid of the nans, the tricontourf will just interpolate in that region. See the script below:

import numpy as np
import matplotlib.pyplot as plt

x,y=np.meshgrid(np.arange(10),np.arange(10))
z=x**2.+y**2.
z[np.logical_and(x<y,x+y>10)]=np.nan
fig,ax=plt.subplots()
ax.contourf(x,y,z)
ax.set_title('contourf')
fig.savefig('contourf.png')
ok=np.isfinite(z)
fig,ax=plt.subplots()
ax.tricontourf(x[ok],y[ok],z[ok])
ax.set_title('tricontourf')
fig.savefig('tricontourf.png')

I think the second approach @ianthomas23 has provided seems good. It seems _process_args could be easily extended to do this book keeping?

Much better with separate triangulation:

@ianthomas23 did you have a chance to look into this more?

@jklymak No, I've been ignoring it. But you are right to ping me about it. I have a change in work situation at the end of March that should free up some time so that I can look at such things. I'll write a PR then.

re-milestoned for the first 3.1.x bug-fix release.

Closed by #14040.