Inconsistent return types for scores

[simon-hirsch]

Hi @frazane, @sallen12

I've noticed some inconsistent return types, e.g.

obs = np.ones(5)
fct = np.random.normal(size=(100, 5))

es = sr.es_ensemble(obs, fct)
dss = sr.dssmv_ensemble(obs, fct)

print(type(es), type(dss))

returns: <class 'numpy.float64'> <class 'numpy.ndarray'>.

The issue is a bit tricky, e.g. when doing stuff like:

df.loc[t, "dss"] = sr.dssmv_ensemble(obs, fct)

then you've got a series of arrays which needs post processing for storage and aggregation.

If I find some time in the next week I might start to do some digging, I suspect it is something in the numba backend/handling of array shapes there

[frazane]

Hi @simon-hirsch, thanks for raising this. I agree consistency here would be more than desirable. Happy to hear you're willing to have a look. Generally speaking, what do you think should be the return type when the result is a scalar?

[simon-hirsch]

So currently sr.dssmv_ensemble(obs, fct).shape is () this sounds pretty odd for an array that is containing a value.

For scalar values I'd lean towards base types (floats) in favor of arrays of shape () - if one want to output arrays, they should at least have shape (1, ) - but that might be personal preference.

[frazane]

For scalar values I'd lean towards base types (floats) in favor of arrays of shape ()

Same preference here!

[simon-hirsch]

Okay so the inconsistency comes from the fact that the DSS is using the lazy_gufunc_wrapper but the ES is not.

Proposal: Explicitly cast in the gufunc wrapper

def lazy_gufunc_wrapper_mv(func):
    """
    Wrapper for lazy/dynamic generalized universal functions so
    they return a value instead of modifying an output array in-place.
    This is the multivariate version, which assumes that the shape of the
    return value is reduced to the shape of the first input argument
    (the observations) along the last axis.
    """

    def wrapper(*args):
        out = np.empty_like(args[0][..., 0])
        func(*args, out)
        if out.shape == ():
            return out.item()
        else:
            return out

    return wrapper

and then also wrap the ES (and potentially some other scores).

Same logic should also apply for the univariate scores.

[simon-hirsch]

Made a PR here: #109