Arguably, push_back (and so also +=) is rather inefficient as it run in a Python loop. To be sure, initializing std::vector from a numpy array isn’t terribly better unless that array is 1D, but if you start out from an empty vector, it’s better to use swap(). You can copy the one you’re taking the contents from if need be. Examples:
import cppyy
import numpy as np
vec = cppyy.gbl.std.vector['unsigned int']
val = 2**32-1
a = np.array([val], dtype=np.uint32)
assert a[0] == val
v1 = vec(a)
assert v1[0] == val
v2 = vec()
v2.swap(vec(a))
assert v2[0] == val
v3 = vec()
v3.swap(vec(v1))
assert v1[0] == val
assert v3[0] == val