Can you use a view to get this, instead of poking at the internal data buffer?

Good point, the following appears to work for regular arrays, but not masked arrays

reconstructed = np.ma.array(data.view(dt))

So the issue with masked arrays is that when the dtype is chnaged from 'comlex128' to 'float64, float64' the shape of the mask changes, so we cannot do data.view(dt) on masked data.

To support both masked and not-masked input data, we can do:

reconstructed = np.ma.array(np.asarray(data).view(dt))

To me, this looks like an improvement over poking the internal data buffer

The mask is in either case addressed in the lines following this:

if np.ma.is_masked(data):
     for descriptor in dt.descr:
         reconstructed[descriptor[0]][data.mask] = np.ma.masked

Since we're already in a isinstance(data, np.ndarray), would it make more sense to do np.ma.getdata instead of np.asarray, which at first glance looks superfluous?