How to visit custom numpy datatype array's member as an array?

Question

I was trying to implement my custom quaternion datatype which has 4 members: w, x, y, z. And I found the official example code: https://github.com/numpy/numpy-dtypes/tree/master/npytypes/quaternion

I tested this implementation by following:

import numpy as np
import npytypes.quaternion

a = np.zeros((2, 2), dtype=np.float).astype(np.quaternion)
print(a)
print(a[0][0].w) # correct, get 0.0
print(a.w) # wrong, AttributeError: 'numpy.ndarray' object has no attribute 'w'

And I got:

[[quaternion(0, 0, 0, 0) quaternion(0, 0, 0, 0)]
 [quaternion(0, 0, 0, 0) quaternion(0, 0, 0, 0)]]
0.0
Traceback (most recent call last):
  File "e:/..../test.py", line 7, in <module>
    print(a.w)
AttributeError: 'numpy.ndarray' object has no attribute 'w'

What I expect was like:

>>> a.w
array([[0.0, 0.0], [0.0, 0.0]], dtype=np.float)

And my question is that how can I modify that code to achive this goal?

np.complex did it well:

>>> import numpy as np
>>> a = np.random.rand(2, 3).astype(np.complex)
>>> a
array([[0.94226049+0.j, 0.71994713+0.j, 0.718848  +0.j],
       [0.57285105+0.j, 0.35576711+0.j, 0.51016149+0.j]])
>>> a.real
array([[0.94226049, 0.71994713, 0.718848  ],
       [0.57285105, 0.35576711, 0.51016149]])
>>> a.real.dtype
dtype('float64')

Answer 1

You might think arrays of complex dtype have extra attributes, but that's probably because you haven't tried to access arr.real or arr.imag on an array of non-complex dtype. It works. Those attributes aren't something specific to complex dtypes - they're baseline NumPy array functionality. (Also, np.complex is just a backwards compatibility alias for the regular Python complex type - when you specify complex as a dtype, NumPy will automatically interpret that as requesting NumPy's complex128 dtype.)

np.ndarray does not have any support for what you're attempting. You could subclass np.ndarray if you really wanted, but that gets messy and wouldn't help with regular arrays.