Slicing a Series returns an Array and fails

[jseabold]

From pystatsmodels ML "[pandas] slicing a series returns an array?"

Got it. It only happens if the series is longer than 200

x = pandas.Series(np.random.random(201), name='x')
x.reshape((-1,1))
<snip>

This is part of what I was talking about that Series aren't quite
array-like. Something like this

np.dot(x, np.random.random((1,10))

doesn't work, and I can't coerce x to be 2d. However, if X is less
than length 201, I sure can coerce it to be 2d, which is why I
couldn't reproduce

x = pandas.Series(np.random.random(199), name='x')
x = x.reshape((-1,1))

But I still can't dot

np.dot(x,np.random.random((1,10)))
<snip>
ValueError: The truth value of an array with more than one element is
ambiguous. Use a.any() or a.all()

I realize that trying to 'dot' might not make sense with a Series or
there's no intuitive interpretation to the output in terms of the
Series that I can think of offhand. Just an example of the 1d vs 2d
issue. Mainly, I guess I'd expect a series to be able to pass through
an atleast_2d check.

[wesm]

This works:

np.dot(x[:, None], randn(1, 10))

But this does not:

np.dot(x.reshape((-1, 1)), randn(1, 10))

This is definitely a bug; guess I need to override reshape and return an array in the case of > 1D output.

Frankly, I'm partially considering making Series not a subclass of ndarray for reasons like these. I'm fearful of all the code people have written like isinstance(obj1d, np.ndarray) that this would break, but otherwise as long as Series obeys the array interface ufuncs and all that will continue to work fine. This would allow me to push Series completely into Cython if I wanted.

[jseabold]

I'm tentatively +1 on having isinstance(series, np.ndarray) return False. I'd be interested to hear what other people think though. IMO, it's not array-like enough. I've pretty much abandoned trying to use Series in any calculations that involve other data for this reason.

OT: I've been playing with a factor class that subclasses Series and it also doesn't make sense as an array.

[wesm]

Yeah, you sacrifice too much by preserving the array interface in array computations. I guess making Series minimally array-compatible (so ufuncs work, like DataFrame) would be the way to go. Need to assess how much havoc it would cause for users.

[wesm]

the bug above is all set, will return to the Series-as-ndarray-subclass issue at a later time