I have a complex-valued NumPy array that I'd like to convert into a contiguous NumPy array with real and imaginary parts separate.
This
import numpy
u = numpy.array([
1.0 + 2.0j,
2.0 + 4.0j,
3.0 + 6.0j,
4.0 + 8.0j
])
u2 = numpy.ascontiguousarray(numpy.vstack((u.real, u.imag)).T)
does the trick, but transposing, vstacking, and converting to a contiguous array is probably a step or two too much.
Is there a native NumPy function that does this for me?
None of the alternatives are native or save on reshape, transposes etc.
For example internally column_stack converts its inputs to 2d 'column' arrays. Effectively it is doing
In [1171]: np.concatenate((np.array(u.real,ndmin=2).T,np.array(u.imag,ndmin=2).T),axis=1)
Out[1171]:
array([[ 1., 2.],
[ 2., 4.],
[ 3., 6.],
[ 4., 8.]])
vstack passes its inputs through atleast_2d(m), making sure each is a 1 row 2d array. np.dstack uses atleast_3d(m).
A new function is np.stack
In [1174]: np.stack((u.real,u.imag),-1)
Out[1174]:
array([[ 1., 2.],
[ 2., 4.],
[ 3., 6.],
[ 4., 8.]])
It uses None indexing to correct dimensions for concatenation; effectively:
np.concatenate((u.real[:,None],u.imag[:,None]),axis=1)
All end up using np.concatenate; it and np.array are the only compiled joining functions.
Another trick is to use view
In [1179]: u.view('(2,)float')
Out[1179]:
array([[ 1., 2.],
[ 2., 4.],
[ 3., 6.],
[ 4., 8.]])
The complex values are saved as 2 adjacent floats. So the same databuffer can be viewed as pure floats, or with this view as a 2d array of floats. In contrast to the concatenate functions, there's no copying here.
Another test on the alternatives is to ask what happens when u is 2d or higher?