I normally create numpy dtypes like this:
C = np.dtype([('a',int),('b',float)])
However in my code I also use the fields a and b individually elsewhere:
A = np.dtype([('a',int)])
B = np.dtype([('b',float)])
For maintainability I'd like to derive C from types A and B somehow like this:
C = np.dtype([A,B]) # this gives a TypeError
Is there a way in numpy to create complex dtypes by combining other dtypes?
You can combine the fields using the .descr attribute of the dtypes. For example, here are your A and B. Note that the .descr attrbute is a list containing an entry for each field:
In [44]: A = np.dtype([('a',int)])
In [45]: A.descr
Out[45]: [('a', '<i8')]
In [46]: B = np.dtype([('b',float)])
In [47]: B.descr
Out[47]: [('b', '<f8')]
Because the values of the .descr attributes are lists, they can be added to create a new dtype:
In [48]: C = np.dtype(A.descr + B.descr)
In [49]: C
Out[49]: dtype([('a', '<i8'), ('b', '<f8')])
According to dtype documentation, dtypes have an attribute descr which provides an "Array-interface compliant full description of the data-type". Therefore:
A = np.dtype([('a',int)]) # A.descr -> [('a', '<i4')]
B = np.dtype([('b',float)]) # B.descr -> [('b', '<f8')]
# then
C = np.dtype([A.descr[0], B.descr[0]])
There is a backwater module in numpy that has a bunch of structured/rec array utilities.
zip_descr does this sort of descr concatenation, but it starts with arrays rather than the dtypes:
In [77]: import numpy.lib.recfunctions as rf
In [78]: rf.zip_descr([np.zeros((0,),dtype=A),np.zeros((0,),dtype=B)])
Out[78]: [('a', '<i4'), ('b', '<f8')]
In [81]: rf.zip_descr([np.array((0,),dtype=A),np.array((0,),dtype=B)])
Out[81]: [('a', '<i4'), ('b', '<f8')]
