I'm trying to create a Matlab cell array in python and save it as a .mat file, but am running into problems when all the cells contain 2 values: 

import scipy.io as sio

twoValues = {'a': array([[array([[2, 2]]), array([[3, 3]])]])}
sio.savemat('test.mat',twoValues) 

In Matlab: 

load('test.mat')
>>> a

a(:,:,1,1) =

           2           3


a(:,:,1,2) =

           2           3


>>> class(a)

ans =

int32

Back in python: 

threeValues = {'a': array([[array([[2, 2, 2]]), array([[3, 3]])]])}
sio.savemat('test.mat',threeValues)

In Matlab: 

>>> a

a = 

    [3x1 int32]    [2x1 int32]


>>> class(a)

ans =

cell 

What's the reason for this? 

Answer:

When you do this: 

a = np.array([[np.array([[2, 2]]), np.array([[3, 3]])]]) 

the final call to np.array actually concatenates the inner two, so you get one array at the end: 

>>> a
array([[[[2, 2]],

        [[3, 3]]]])

>>> a.shape
(1, 2, 1, 2) 

But to mimic a cell array you want to basically have an array of arrays. You can acheive this by setting dtype=object, but you must create the array and set the elements separately to avoid the automatic merging.

three = array([[array([[2, 2, 2]]), array([[3, 3]])]])
two = np.empty(three.shape, dtype=object)
two[0,0,0] = np.array([[2,2]])
two[0,1,0] = np.array([[3,3]])

Then: 

sio.savemat('two.mat', {'two': two}) 

to see what they look like: 

>>> two
array([[[array([[2, 2]])],
        [array([[3, 3]])]]], dtype=object)

>>> two.shape
(1, 2, 1)

Note that I may have gotten confused about your desired shape, since you have so many nested brackets, so you might have to reshape some of this, but the idea should hold regardless.