Make np.ma.take works on scalars - #7586

Nice catch!

I'm not sure I like how this always returns a 0-d array though. Eg, with this PR np.ma.array([1,2,3]).take(0) will return a 0-d array, unlike np.array([1,2,3]).take(0) which returns a scalar.

I think it might be better to detect the scalar case and return np.ma.masked if needed. Eg, leave most of take alone but add something like

if np.isscalar(out) and _mask[indices] is True:
    return masked
return out

(As desired, this will not affect the case the user supplied out as a kwd, as in that case out will never be a scalar)

Oh, sorry I misunderstood what is going on. I see that you don't return a 0-d array in the end.

Still, maybe having an explicit if-statement is clearer than converting to 0-d with [...] and later converting back to scalar with [()].

I'd argue that the version where we don't special case scalars is better - your fix introduces a bug when indexing with np.ma.masked.

Incidentally, np.isscalar(np.ma.masked) is False (#7588). Is this deliberate?

That's true, you'd need to account for the indices is masked case.

One reason I prefer to avoid using [()] is it is buggy: see #7267. So take will not work for string and unicode types if you use it.

By the way, the reason np.isscalar(np.ma.masked) is False is related to the problem you noticed above, that scalars can't be viewed as ndarray subclasses: masked is an instance of MaskedArray which is a subclass of ndarray, but scalars are not subclasses of ndarrays.

These are all long-standing difficult problems people have thought about changing . For instance, in my opinion the very existence of np.ma.masked is a design mistake, and if/when we rewrite MaskedArray we should get rid of it completely. I know other numpy contributors also think we should get rid of scalars completely in favor of 0-d arrays, to avoid problems like this one.

(actually I think #7267 should be fixable without too much fuss, as I outlined in this comment, I just never got around to it)

Sorry, I should have looked a little more before posting. Actually #7267 does not affect this PR since out is guaranteed to be an ndarray.

So actually I think your code is fine. It even does the right think if the user supplies a 0-d array as out.

I would be happier with a comment though. Would you mind adding something to the effect of "account for scalar case by converting to 0-d with [...] and back with [()]"?

Comments added

Looks good! I'll merge i a little bit.

Merged. Thanks @eric-wieser