Why does assert np.nan == np.nan cause an error?

NaN has the property that it doesn't equal itself, you should use np.isnan to test NaN values, here np.isnan(np.nan) will yield True:

In[5]:
np.nan == np.nan

Out[5]: False

In[6]:
np.nan != np.nan

Out[6]: True

In[7]:
np.isnan(np.nan)

Out[7]: True

Use np.isnan(value). NaN doesn't compare equal to itself because it indicates a failure, and might not have been produced the same way. I'm not sure why isnan is missing in the CPython documentation, but it's present in math for both CPython 3.4 and 2.7, and as a ufunc in numpy.