Add ndarray.keys() and ndarray.values() methods

It would be useful for duck typing if ndarray had keys() and values() methods. Here’s how I would implement them in Python:

def keys(self):
    return self.dtype.names

def values(self):
    return self.flat

The way I’ve shown, keys() would return a tuple and values() would return an iterator (as dict does in Python 3).

This would make it easier to write generic code which handles structured arrays, dicts, and other mapping types transparently, for example:

np.array([(1, 3), (4, 5)], dtype=[('a', int), ('b', int)])

{ 'a': [1, 4], 'b': [3, 5] }

Without this, I currently have ugly code such as:

def hasColumn(data, name):
    columns = data.dtype.names if isinstance(data, np.ndarray) else data.keys()
    return name in columns

As far as I know, these methods would actually be implemented in C, specifically by adding to array_methods in multiarray/methods.c (and using flat from multiarray/getset.c).

[jzwinck]

The proposed keys() would return a set of valid arguments to __getitem__, though of course ndarray supports many other arguments too. I think most people would consider the names of columns in a structured array to be its “keys” even if indexing is also possible by row numbers, boolean masks, etc.

If your impression is that what I am suggesting is for keys() to return things not usable with __getitem__ then I will have to apologize for miscommunicating. It absolutely is my intention that indexing would be possible using the results of keys().

[jzwinck]

What’s proposed in #7552 is good enough to satisfy this.