How do you create a multidimensional numpy array from an iterative of tuples?


I would like to create a numpy array from an iterable, which yields tuples of values, such as a database query.

Like so:

data = db.execute('SELECT col1, col2, col3, col4 FROM data')
A = np.array(list(data))

Is there a way faster way of doing so, without converting the iterable to a list first?

I am not an experienced user of numpy, but here is a possible solution for the general question:

>>> i = iter([(1, 11), (2, 22)])
>>> i
<listiterator at 0x5b2de30>                    # a sample iterable of tuples
>>> rec_array = np.fromiter(i, dtype='i4,i4')  # mind the dtype
>>> rec_array                                  # rec_array is a record array
array([(1, 11), (2, 22)],
    dtype=[('f0', '<i4'), ('f1', '<i4')])
>>> rec_array['f0'], rec_array[0]              # each field has a default name
(array([1, 2]), (1, 11))
>>> a = rec_array.view(np.int32).reshape(-1,2) # let's create a view
>>> a
array([[ 1, 11],
       [ 2, 22]])
>>> rec_array[0][1] = 23
>>> a                                          # a is a view, not a copy!
array([[ 1, 23],
       [ 2, 22]])

I assume that all columns are of the same type, otherwise rec_array is already what you want.

Concerning your particular case, I do not completely understand what is db in your example. If it is a cursor object, then you can just call its fetchall method and get a list of tuples. In most cases, the database library does not want to keep a partially read query result, waiting for your code processing each line, that is by the moment when the execute method returns, all data is already stored in a list, and there is hardly a problem of using fetchall instead of iterating cursor instance.