Series constructor doesn't handle object implementing numpy array interface

[dalejung]

In [1]: import pandas as pd
In [2]: import rpy2
In [7]: import numpy as np
In [4]: arr = rpy2.robjects.IntVector(range(10))
In [5]: arr
Out[5]:
<IntVector - Python:0x102ed3908 / R:0x105a7c178>
[       0,        1,        2, ...,        7,        8,        9]

In [6]: pd.Series(arr) # wrong behavior
Out[6]:
0    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
dtype: object

In [9]: np.array(arr)
Out[9]: array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=int32)

In [10]: pd.Series(np.array(arr)) # correct behavior
Out[10]:
0    0
1    1
2    2
3    3
4    4
5    5
6    6
7    7
8    8
9    9
dtype: int32

I guess this is a regression due to Series no longer being a subclass of np.ndarray. Not sure if numpy exposes a catchall isarray that does the checks.

[jreback]

IIRC this is an issue with rpy2, you need a fairly recent version of it.

[dalejung]

hm. I'm on 2.4.3 which is the latest. I'm not seeing where in the constructor we'd be handling it, unless numpy itself is overloading the isinstance check. I gotta step out for a bit, but I'll dig into it when I get back.

I know this used to work before. I don't use rpy2 regulary...I'm just fixing some tests that have been stuck at red for a long time :)

[jreback]

this doesnt' have anything really to do with numpy. Its the rpy2.robjects.IntVector(range(10))
why does Series not being a sub-class have to do with anythin here?

[dalejung]

Ah right. I was meaning that being a ndarray subclass would help the chances of data making it pa.array untouched. Which is what happens on 12.0.

The problem in master is:

# series.py 198
if index is None:
    if not is_list_like(data):
        data = [data]       
    index = _default_index(len(data))

In 12.0 there is no data = [data] and the index check gets called after _sanitize_array anyways.

So I'm thinking that we should have something like:

if data has attr '__array_interface__' or '__array_struct__':
  data = pa.array(data)

I'm not entirely sure what the rpy2 vector could do, it doesn't pull the data from R until you treat it like a numpy array. But it is properly implementing the interface. Something like

arr = rpy2.robjects.IntVector(range(10))
np.sum(arr) # 45

works fine. If we're cool with the array interface check, I'll throw up a PR tonight.

[jreback]

you can check for 'array' is a more compat way

you can try not sure if will work

[dalejung]

yar, it was meant to be pseudo-ish, gotta look around to see if there's an established way to check. Will report back.

[jreback]

closing, but pls reopen if the issue persists.