On Di, 2015-05-05 at 11:13 -0400, Allan Haldane wrote: > Hello all, > > A question: > > Many ndarray methods (eg sum, mean, any, min) have a "keepdims" keyword > argument, but ndarray subclass methods sometimes don't. The 'matrix' > subclass doesn't, and numpy functions like 'np.sum' intentionally > drop/ignore the keepdims argument when called with an ndarray subclass > as first argument. > > This means you can't always use ndarray subclasses as 'drop in' > replacement for ndarrays if the code uses keepdims (even indirectly), > and it means code that deals with keepdims (eg np.sum and more) has to > detect ndarray subclasses and drop keepdims even if the subclass > supports it (since there is no good way to detect support). It seems to > me that if we are going to use inheritance, subclass methods should keep > the signature of the parent class methods. What does the list think? > > ---- Details: ---- > > This problem comes up in a PR I'm working on (#5706) to add the keepdims > arg to masked array methods. In order to support masked matrices (which > a lot of unit tests check), I would have to detect and drop the keepdims > arg to avoid an exception. This would be solved if the matrix class > supported keepdims (plus an update to np.sum). Similarly, > `np.sum(mymaskedarray, keepdims=True)` does not respect keepdims, but it > could work if all subclasses supported keepdims. > > I do not foresee immediate problems with adding keepdims to the matrix > methods, except that it would be an unused argument. Modifying `np.sum` > to always pass on the keepdims arg is trickier, since it would break any > code that tried to np.sum a subclass that doesn't support keepdims, eg > pandas.DataFrame. **kwargs tricks might work. But if it's permissible I > think it would be better to require subclasses to support all the > keyword args ndarray supports.
What is the advantage over having an error raised due to the invalid **kwargs trick when the subclass does not support it? First sight it seems like a far shot have a hard requirement. The transition period alone seems hard, unless we have add magic to test the subclass upon creation, and I am not sure that is easy to do (something like ABC conformance test). - Sebastian > > Allan > _______________________________________________ > NumPy-Discussion mailing list > NumPy-Discussion@scipy.org > http://mail.scipy.org/mailman/listinfo/numpy-discussion >
signature.asc
Description: This is a digitally signed message part
_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
