numpy.array with elements of different shapes

Question

I want to have a numpy array of two another arrays (each of them has different shape). As I know, for this reason one must use: dtype = object in the definition of the main array.

For example, let us define (in Python 2.7) our arrays as

     a0 = np.arange(2*2).reshape(2,2)
     a1 = np.arange(3*3*2).reshape(3,3,2)
     b = np.array([a0,a1], dtype = object)

This works perfect: b[1] is the same as a1. But if I change the dimension in a0 from (2,2) to (3,3) something strange happens:

     a0 = np.arange(3*3).reshape(3,3)
     a1 = np.arange(3*3*2).reshape(3,3,2)
     b = np.array([a0,a1], dtype = object)

This time b[1] and a1 are not equal, they even have different shapes. What is the reason of this strange behavior?

Perhaps there is a completely different solution for me. But I don't want to use lists or tuples because I want to allow addition such as b + b. It is clear that I can write my own class for this purpose but is there any simpler way?

Answer 1

If you explicitly want an objects array, you can create an empty array with type object first and assign to it:

x = empty(5, dtype=object)
x[0] = zeros((3,3))
x[1] = zeros((3,2)) #does not merge axes.
x[2] = eye(4)
x[3] = ones((2,2))*2
x[4] = arange(10).reshape((5,2))

>>> x+x
array([array([[ 0.,  0.,  0.],
   [ 0.,  0.,  0.],
   [ 0.,  0.,  0.]]),
   array([[ 0.,  0.],
   [ 0.,  0.],
   [ 0.,  0.]]),
   array([[ 2.,  0.,  0.,  0.],
   [ 0.,  2.,  0.,  0.],
   [ 0.,  0.,  2.,  0.],
   [ 0.,  0.,  0.,  2.]]),
   array([[ 4.,  4.],
   [ 4.,  4.]]),
   array([[ 0,  2],
   [ 4,  6],
   [ 8, 10],
   [12, 14],
   [16, 18]])], dtype=object)

You will have to fill all elements before you can perform arithmetic, or grow the element from size zero using np.append.