Loading a cfg_dep from a batch.mat is not working

[balbasty]

I have saved a cfg_dep that is displayed like this in matlab:

dep = 
 
cfg_dep object: 1-by-1
           tname: 'Select GCM_*.mat'
    tgt_exbranch: [0×0 struct]
       tgt_input: [0×0 struct]
        tgt_spec: {[1×2 struct]}
          jtsubs: [0×0 struct]
           sname: 'Specify group: GCM mat File(s)'
    src_exbranch: [1×10 struct]
      src_output: [1×1 struct]

I then load it and display it in spm-python:

In [1]: from spm import Runtime

In [2]: dep = Runtime.call('load', 'dep.mat').dep
Initializing Matlab Runtime...
WARNING: package sun.awt.X11 not in java.desktop
WARNING: package sun.awt.X11 not in java.desktop

In [3]: dep
Out[3]: <spm.__matlabbatch.cfg_dep.cfg_dep at 0x1659e87c0>

In [4]: Runtime.call('disp', dep, nargout=0)
cfg_dep object: 1-by-1
           tname: 'Target'
    tgt_exbranch: [0x0 struct]
       tgt_input: [0x0 struct]
        tgt_spec: {}
          jtsubs: [0x0 struct]
           sname: [1x1 struct]
    src_exbranch: [0x0 struct]
      src_output: [0x0 struct]

[johmedr]

Oh, interesting, could you share the file you have?
We need to enable some debug logging in mpython_endpoint.

[johmedr]

Hi @balbasty, have you retried since the last updates?

[balbasty]

Still the same with spm-python==25.1.2.post2 & mpython-core==25.4rc1 (tried on R2024b and R2025b).

I've attached the test object.

dep.mat.gz

[balbasty]

I've played a bit with it:

1. I believe there is a bug in mpython_endpoint here. Iterating over a cell of strings yields cells of size 1x1, so it should be

s.(f{1}) = S.data__.(f{1});

2. Even with this fix, the problem with cfg_dep is that the first conversion option here does not fail, but returns a "default" cfg_dep. We need a more robust way to detect that the "field-by-field" conversion must be used, but I don't have any good idea. Maybe check that isequal(S.__data, struct(feval(S.class__, S.data__)))? But it could be expensive when the object contains arrays.
3. It might not be an issue with SPM objects, but it seems that the current field-by-field conversion may not be enough in the presence of private fields.

[johmedr]

Then is the solution to do our best to convert the object to a structure in mpython, and to suggest people to implement a struct class method on this object that explicitly handles the conversion, as suggested here?

[balbasty]

I've tried to use the mex hack, but it does not work on late recent of MATLAB (as reported by the author in 2023).

In the meantime, I propose the following fix to mpython_endpoint (L118-129):

                try
                    s = feval(S.class__);
                catch
                    s = feval(S.class__, S.data__);
                end
                for f = fieldnames(s)'
                    f = f{1}
                    try
                        s.(f) = S.data__.(f);
                    catch
                        % ignore
                    end
                end

The idea is to always copy fields, and favour the default constructor over the copy constructor. I haven't run tests on this.

There may still be cases where a class has neither default nor copy constructors, and it does not solve the problem of private fields in "new" (R2008+) class objects.

---

Edit: I am also wondering whether we currently support object arrays? It seems that struct(objectarray) returns a scalar struct rather than a struct array.

[balbasty]

As an alternative, there are two undocumented functions that can be use to serialize/deserialize objects: getArrayFromByteStream and getByteStreamFromArray.

See this.

Since, for objects, we don't do anything with the serialized data on the python side apart from holding it, we could simply pass the byte stream back and forth in mpython_endpoint.