Open mrocklin opened 9 years ago
You could adopt the h5py custom 'O' dtype for this. np.dtype('O', metadata={'vlen' : str}) This is what dynd converts strings to when going to numpy.
Dealing with numpy's dtype metadata is a bit funny though, it doesn't participate in equality, so you have to check that manually.
I'm more concerned about the case where someone gives me a numpy array and I'm asked to discover its datashape. My concern here is somewhat abated by adding a heuristic to the discover(np.ndarray)
function in #121 .
Object type isn't seeing any love and, at least in Blaze, I find it useful to interpret numpy dtype
'O'
asstring
. There are trade-offs here. Thought I'd ask for discussion.@mwiebe thoughts?