BUG/ENH: groupby head returns empty result for negative n

[TomAugspurger]

More drive-by reports. I'll try to close some of these issues sometime...

In [8]: s = pd.Series(np.random.randn(10))

In [9]: s.groupby(s > 0).apply(lambda x: x.head(-1))  # works nicely
Out[9]:
False  0   -0.277400
       1   -0.875231
       3   -0.315087
       4   -0.966649
       6   -0.474389
       7   -0.121225
True   2    0.237861
       5    1.311601
dtype: float64

In [10]: s.groupby(s > 0).head(-1)  # doesn't.
Out[10]: Series([], dtype: float64)

[cpcloud]

this seems like a strange way to express what you're showing here. why wouldn't one use tail(1) instead?

[TomAugspurger]

I wanted all but the last one.

[jorisvandenbossche]

I found it a bit strange that it works like this. The docstring of head also does not really mentions this ("Returns first n rows").
But apparantly that is the way it works for DataFrame.head(), so it should also for groupby.head()?

In [32]: df = pd.DataFrame(np.random.randn(5,2))

In [33]: df
Out[33]:
          0         1
0  0.902403  0.270362
1 -0.139103 -0.680811
2  1.284748 -0.404867
3  0.857957 -0.341599
4  0.741684 -1.402058

In [34]: df.head(-1)
Out[34]:
          0         1
0  0.902403  0.270362
1 -0.139103 -0.680811
2  1.284748 -0.404867
3  0.857957 -0.341599

[jreback]

In [103]: df = DataFrame(np.arange(10).reshape(5,2))

In [104]: df
Out[104]: 
   0  1
0  0  1
1  2  3
2  4  5
3  6  7
4  8  9

# .head(2)
In [105]: df.iloc[0:2]
Out[105]: 
   0  1
0  0  1
1  2  3

# head(-2)
In [106]: df.iloc[0:-2]
Out[106]: 
   0  1
0  0  1
1  2  3
2  4  5

so this is how its implemented. it does look odd, but follows slicing semantics.

[TomAugspurger]

I realized last night that I could, and should, just use iloc. I'm ok with closing this as not a priority.

On Jan 9, 2015, at 06:04, jreback notifications@github.com wrote:

In [103]: df = DataFrame(np.arange(10).reshape(5,2))

In [104]: df
Out[104]:
0 1
0 0 1
1 2 3
2 4 5
3 6 7
4 8 9

.head(2)

In [105]: df.iloc[0:2]
Out[105]:
0 1
0 0 1
1 2 3

head(-2)

In [106]: df.iloc[0:-2]
Out[106]:
0 1
0 0 1
1 2 3
2 4 5
so this is how its implemented. it does look odd, but follows slicing semantics.

—
Reply to this email directly or view it on GitHub.

[jorisvandenbossche]

I think we should aim for consistency with DataFrame.head(), so I wouldn't close it.

[hayd]

Commented on the SO question:

def negative_head(g, n):
    return g._selected_obj[g.cumcount(ascending=False) >= n]

def negative_tail(g, n):
    return g._selected_obj[g.cumcount() >= n]

I didn't realize this was a thing when implementing this, sorry.

[smcinerney]

Unaware that this one from 2015 existed, I independently just filed #30192. Its title and body make clear that it's suggesting the undocumented behavior is useful and should be documented.

[TomAugspurger]

Closing in favor of #30192