np.diff fails when called on a Series

[wesm]

In [13]: s = Series(np.arange(10))

In [14]: np.diff(s)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/home/wesm/code/pandas/<ipython-input-14-b5f9fe77ab7c> in <module>()
----> 1 np.diff(s)

/usr/lib/epd-7.1/lib/python2.7/site-packages/numpy/lib/function_base.pyc in diff(a, n, axis)
    975         return diff(a[slice1]-a[slice2], n-1, axis=axis)
    976     else:
--> 977         return a[slice1]-a[slice2]
    978 
    979 def interp(x, xp, fp, left=None, right=None):

/home/wesm/code/pandas/pandas/core/series.pyc in __getitem__(self, key)
    392             key = np.asarray(key, dtype=bool)
    393 
--> 394         return self._get_with(key)
    395 
    396     def _get_with(self, key):

/home/wesm/code/pandas/pandas/core/series.pyc in _get_with(self, key)
    406         else:
    407             if isinstance(key, tuple):
--> 408                 return self._get_values_tuple(key)
    409 
    410             if not isinstance(key, (list, np.ndarray)):  # pragma: no cover

/home/wesm/code/pandas/pandas/core/series.pyc in _get_values_tuple(self, key)
    437 
    438         if not isinstance(self.index, MultiIndex):
--> 439             raise ValueError('Can only tuple-index with a MultiIndex')
    440 
    441         # If key is contained, would have returned by now

ValueError: Can only tuple-index with a MultiIndex

[changhiskhan]

On Dec 3, 2012, at 1:31 PM, Dan Allan notifications@github.com wrote:

There is still trouble here.

In [163]: s = Series(np.arange(10))

In [164]: np.diff(s)
Out[164]:
0 NaN
1 0
2 0
3 0
4 0
5 0
6 0
7 0
8 0
9 NaN
—
Reply to this email directly or view it on GitHub.

Hi Dan, you should be using s.diff() instead.
np.diff treats the input argument as index unaware so it does not do the right thing for a Series.
If you want to duplicate np.diff results, you can do np.diff(s.values)

[wesm]

I'm reopening and moving this to at minimum 0.11. The solution here is to make Series not an ndarray, once and for all. The issue is the numpy.diff uses asanyarray instead of asarray, and Series has different array behavior than diff

[jseabold]

I just ask that downstream packages have plenty of time to look at this before it becomes released / merged. I don't know the implications yet, but I'm almost positive that I abuse the Series is an array sub-class more than I think.

[michaelaye]

Do I understand it right that you want to make Series not a subclass of ndarray? Does that mean that you have to re-implement all the statistical helper functions inside? Or would this imply a lot of API changes for Series?

[jreback]

current behavior (0.12):

In [6]: np.diff(Series(np.arange(10)))
Out[6]: 
0   NaN
1     0
2     0
3     0
4     0
5     0
6     0
7     0
8     0
9   NaN
dtype: float64

after #3482

In [1]: s = Series(np.arange(10))

In [2]: np.diff(s)
Out[2]: array([1, 1, 1, 1, 1, 1, 1, 1, 1])

In [3]: np.diff(s.values)
Out[3]: array([1, 1, 1, 1, 1, 1, 1, 1, 1])
In [4]: s.diff()

Out[4]: 
0   NaN
1     1
2     1
3     1
4     1
5     1
6     1
7     1
8     1
9     1
dtype: float64

[jreback]

closed by #3482