如何将二维 numpy 数组直接拆分为 numpy 对象数组

Question

For a special application dealing with numpy arrays of different lengths, I need my preferably numpy array, not just a list , to have the form np.ndarray[np.ndarray[ ], np.ndarray[ ], ..., dtype=object] . If I have given sequence, list, etc. of numpy arrays, I want them always to have this form. However, for a list of numpy arrays of the same length, eg,

np.array(*np.array([np.arange(4), np.arange(4)], dtype=object)

gives me np.ndarray[np.ndarray[[]], dtype=object] so I came up with the workaround below.

Is there any other magic option, which could be passed to np.array() or another method which gives the desired result more directly?

Workaround:

inp_arr_a = np.asarray([np.arange(4), np.arange(3)], dtype=object)
inp_arr_b = np.array([np.arange(4), np.arange(4)])

def split_to_obj_arr(arr):
    return np.delete(np.array([*arr, 'dummy'], dtype=object), -1, 0)

gives for split_to_obj_arr(inp_arr_a)

array([array([0, 1, 2, 3]), array([0, 1, 2])], dtype=object)

and for split_to_obj_arr(inp_arr_b)

array([array([0, 1, 2, 3]), array([0, 1, 2, 3])], dtype=object)

Answer 1

np.array(...) by design tries to return as high a dimensional numeric array as possible. If the inputs are ragged it will raise a future-warning (unless you specify object dtype) and return the object array containing arrays. Or with some combinations of shapes it will raise an error.

Forcing an object dtype with the None element and then deleting that is one way around this. I prefer creating the None filled array first, and assigning the elements:

In [80]: def foo(alist):
    ...:     res = np.empty(len(alist), object)
    ...:     res[:] = alist
    ...:     return res
    ...: 
In [81]: foo([[],[]])
Out[81]: array([list([]), list([])], dtype=object)
In [82]: foo([np.array([]),np.array([])])
Out[82]: array([array([], dtype=float64), array([], dtype=float64)], dtype=object)
In [83]: foo([np.ones((2,3)),np.zeros((2,3))])
Out[83]: 
array([array([[1., 1., 1.],
       [1., 1., 1.]]),
       array([[0., 0., 0.],
       [0., 0., 0.]])], dtype=object)
In [84]: foo([np.array([2,3]),np.array([1,2])])
Out[84]: array([array([2, 3]), array([1, 2])], dtype=object)

Creating a 2d object array like this is also possible, but trickier. It may be simpler to reshape a 1d as needed after.