Feature request: df that only return dfs when indexing.

[twiecki]

With more complex dataframes, I often stumble over this:

In [1]: x = pandas.DataFrame({'a': [1,2,3], 'b': [4,5,6]})
In [2]: x.ix[0].a
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-8-8358f19a1e70> in <module>()
----> 1 x.ix[0].a

AttributeError: 'Series' object has no attribute 'a'

In [3]: x.a.ix[0]
Out[3]: 1

This is just an example, at other times I end up converting the Series back to a DataFrame because that's what the rest of the code expects.

I know that there have been attempts to solve this issue by adding attribute lookup to Series (e.g. #1904) but they seem to come with a performance penalty.

Often, however, I care more about expressiveness than performance. I thus propose the addition of an option like:

x = DataFrame(data, slicing_returns_df=True)

Which will cause x.ix[0] or x.a to return again a DataFrame rather than a Series and make the above work. The default would be False so that there are no backward compat issues. Alternatively there could be a new DataFrame class that inherits from DataFrame and has the desired behavior.

I'm happy to gives this a crack, however, I wanted to first make sure that it's not only me who thinks that'd be a good idea or that this can't work for obvious reason X.

[jreback]

are you looking for something like this? (this is 0.11-dev)
equiv in 0.10.1 is x.get_value(0,'a')

In [6]: x.at[0,'a']
Out[6]: 1

[jreback]

And if you ALWAYS want to force a df to return (the above is ALWAYS a scalar)

In [12]: x.loc[[0],['a']]
Out[12]: 
   a
0  1

[twiecki]

.get_value() does not seem to support multi-indices x.get_value([0,1], 'a').

However, x.loc seems to do exactly what I need. Thanks!

[jreback]

@twiecki great!

yes, by definition at (and get_value) return are for fast scalar access, see: http://pandas.pydata.org/pandas-docs/dev/indexing.html#fast-scalar-value-getting-and-setting

while loc (and ix) are more flexibile label based (and have a small performance penatly in order to figure out what you are after)

[twiecki]

I tried this again just now but df.loc[0,:] returns a series again. Was this changed by any chance? This is with current master.

[jreback]

You asked for a Series, this will always return a series, enclose the list of rows with a [] to always get a frame

and FYI, the 0 is a label here (and not a location!),

In [1]: df = DataFrame(randn(5,2),columns=list('AB'))

In [2]: df.loc[0,:]
Out[2]: 
A   -1.080064
B   -0.499382
Name: 0, dtype: float64

In [3]: df.loc[[0],:]
Out[3]: 
          A         B
0 -1.080064 -0.499382

[twiecki]

Perfect. Thanks!