python - How can I create a Matlab struct array from scipy.io?

Question

Consider the following Matlab code:

pmod(1).name{1}  = 'regressor1';
pmod(1).param{1} = [1 2 4 5 6];
pmod(1).poly{1}  = 1; 
pmod(2).name{1}  = 'regressor2-1';
pmod(2).param{1} = [1 3 5 7]; 
pmod(2).poly{1}  = 1;

This creates a struct array. Each struct in the array contains three fields of type cell. As such, we have the following hierarchy in pmod:

pmod  // struct array
|
*- struct
|  |
|  *- cell  // contains 1 or more strings
|  *- cell  // contains 1 or more arrays
|  *- cell  // contains 1 or more arrays
|
*- struct [...]

I'm trying to use scipy.io to produce the above data structures in Python, such that they can be loaded into Matlab (this hierarchy is required by SPM).

Creating a struct is straightforward, as scipy.io.savemat saves any dict whose keys are all of type str as a Matlab struct:

from scipy.io import savemat

struct = {
    'field1': 1,
    'field2': 2,
}

savemat('/tmp/p.mat', {'a_struct': struct})

However, when trying to generalize this to a struct array, I hit the following roadblock:

struct_array = [struct, struct]
savemat('/tmp/p.mat', {'s_array': struct_array})

This does not behave as expected; when loading p.mat into Matlab, I get a 1x2 cell array, not a struct array.

How can I create a struct array using scipy.io?

Notes:

1. I have tried savemat('/tmp/p.mat', np.array(struct_array)) and savemat('/tmp/p.mat', np.array(struct_array, dtype=object)), to no avail.

Answer 1

You can use np.core.records.fromarrays to construct a record array, which is roughly equivalent to a MATLAB struct, and will be converted to a MATLAB struct by scip.io.savemat.

from numpy.core.records import fromarrays
from scipy.io import savemat

myrec = fromarrays([[1, 10], [2, 20]], names=['field1', 'field2'])
savemat('p.mat', {'myrec': myrec})

When opened in MATLAB, this gives:

>> load('p.mat')
>> myrec

myrec = 

1x2 struct array with fields:

    field1
    field2