NumPy: array assignment issue when using custom dtype

I've found the following puzzling behavior with NumPy and a custom dtype for an ndarray:

import numpy as np

# Make a custom dtype with a single triple开发者_JS百科t of floats (my actual dtype has other
# components, but this suffices to demonstrate the problem.
dt = np.dtype([('a', np.float64, 3)])

# Make a zero array with this dtype:
points = np.zeros((4, 4), dtype=dt)

# Try to edit an entry:
points[0][0]['a'] = np.array([1, 1, 1])

print points[0][0]['a']

Now, this comes back as containing not [1. 1. 1.] as I would expect, but instead [1. 0. 0.], only performing the assignment on the first coordinate. I can get around this by performing the assignment coordinate-wise, but that seems unnecessary given that the full assignment should certainly be the default behavior in this case.

Any thoughts on what's going on here?

If you change the ordering of the indices, like this: points['a'][0][0] = np.array([1, 1, 1]), it works ok for me (python 2.6.5, numpy 1.3.0 on Ubuntu 10.04). I wish I knew why.

There are many method to assign points, if you want your method works:

points[0][0]['a'][:] = np.array([1, 1, 1])

or:

points[0,0]['a'][:] = np.array([1, 1, 1])

because points[0,0]['a'] is an array, if you want to change the content of the array, you shoud use index.