How can I create a numpy array whose elements are other numpy array objects?

Question

I find myself needing to create a numpy array of dtype="object" whose elements are themselves numpy arrays. I can manage to do this if the arrays are different lengths:

arr_of_arrs = np.empty((2,2), dtype=np.object)
arr_list = [np.arange(i) for i in range(4)]
arr_of_arrs.flat[:] = arr_list
print(arr_of_arrs)

array([[array([], dtype=int32), array([0])],
   [array([0, 1]), array([0, 1, 2])]], dtype=object)

But if they happen to be the same length, it doesn't work and I am not entirely sure how it is generating the values it gives me:

arr_list = [np.arange(2) for i in range(4)]
arr_of_arrs.flat[:] = arr_list
print(arr_of_arrs)

[[0 1]
[0 1]]

Is this even doable? numpy seems to try and coerce the data into "making sense" despite my best efforts to prevent it from doing so...

Answer 1

If the array is 1d, the assignment works fine:

In [767]: arr = np.empty(4,object)                                                             
In [768]: arr[:] = [np.arange(6) for _ in range(4)]                                            
In [769]: arr                                                                                  
Out[769]: 
array([array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5]),
       array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])], dtype=object)
In [770]: arr.reshape(2,2)                                                                     
Out[770]: 
array([[array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])],
       [array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])]],
      dtype=object)

We can also start with (2,2), but assign to ravel() (a view ):

In [771]: arr = np.empty((2,2),object)                                                         
In [772]: arr.ravel()[:] = [np.arange(6) for _ in range(4)]                                    
In [773]: arr                                                                                  
Out[773]: 
array([[array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])],
       [array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])]],
      dtype=object)

flat apparently serializes the RHS:

In [774]: arr.flat = [np.arange(6) for _ in range(4)]                                          
In [775]: arr                                                                                  
Out[775]: 
array([[0, 1],
       [2, 3]], dtype=object)

If the RHS list is nested right we can assign directly to the 2d array:

In [779]: alist = Out[770].tolist()                                                            
In [780]: alist                       # list of lists of arrays                                                         
Out[780]: 
[[array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])],
 [array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])]]
In [781]: arr = np.empty((2,2),object)                                                         
In [782]: arr[:] = alist                                                                       
In [783]: arr                                                                                  
Out[783]: 
array([[array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])],
       [array([0, 1, 2, 3, 4, 5]), array([0, 1, 2, 3, 4, 5])]],
      dtype=object)