Funny thing about the not a number (nan) implementation in numpy.

>>> import numpy as np
>>> np.nan is np.NaN is np.NAN
True
>>> np.nan == np.NaN == np.NAN
False

Which makes sense, because nan is never equal to anything, but any nan *is* any other implementation of it. Moral of the story: don't assign stuff with arr[arr==np.nan]=x. It won't work. Use np.isnan(x).