Setting values on slice of multi-index gives NaNs

[jim22k]

Best shown with an example.

import numpy as np, pandas as pd
timestamps = map(pd.Timestamp, ['2014-01-01', '2014-02-01'])
categories = ['A', 'B', 'C', 'D']
df = pd.DataFrame(index=pd.MultiIndex.from_product([timestamps, categories], names=['ts', 'cat']),
                  columns=['Col1', 'Col2'])

>>> df
                Col1  Col2
ts         cat            
2014-01-01 A     NaN   NaN
           B     NaN   NaN
           C     NaN   NaN
           D     NaN   NaN
2014-02-01 A     NaN   NaN
           B     NaN   NaN
           C     NaN   NaN
           D     NaN   NaN

I want to set the values for all categories in a single month. These examples work just fine.

df.loc['2014-01-01', 'Col1'] = 5
df.loc['2014-01-01', 'Col2'] = [1,2,3,4]

>>> df
               Col1 Col2
ts         cat          
2014-01-01 A      5    1
           B      5    2
           C      5    3
           D      5    4
2014-02-01 A    NaN  NaN
           B    NaN  NaN
           C    NaN  NaN
           D    NaN  NaN

These examples don't work.

df.loc['2014-01-01', 'Col1'] += 1
df.loc['2014-02-01', 'Col2'] = df.loc['2014-01-01', 'Col2']

>>> df
               Col1 Col2
ts         cat          
2014-01-01 A    NaN    1
           B    NaN    2
           C    NaN    3
           D    NaN    4
2014-02-01 A    NaN  NaN
           B    NaN  NaN
           C    NaN  NaN
           D    NaN  NaN

It doesn't seem to be a "setting a value on a copy" issue. Instead, Pandas is writing the NaNs.

My current workaround is to unstack each column into a DataFrame with simple indexes. This works, but I have lots of columns to work with. One dataframe is much easier to work with than a pile of dataframes.

The computations for each month depend on the values computed in the previous month, hence why it can't be done fully vectorized on an entire column.

[jreback]

see docs here

In [17]: idx = pd.IndexSlice

# this is multi-index slicing
# this will *always* work
In [18]: df.loc[idx['2014-01-01',:], 'Col1'] += 1

In [19]: df
Out[19]: 
               Col1 Col2
ts         cat          
2014-01-01 A      6    1
           B      6    2
           C      6    3
           D      6    4
2014-02-01 A    NaN  NaN
           B    NaN  NaN
           C    NaN  NaN
           D    NaN  NaN

# pandas automatically aligns the labels for you, so you need to tell it to just shove these values in (as the ``.values`` makes it a numpy array w/o the labels)
In [22]: df.loc[idx['2014-02-01',:], 'Col2'] = df.loc[idx['2014-01-01',:], 'Col2'].values

In [23]: df
Out[23]: 
               Col1  Col2
ts         cat           
2014-01-01 A      6     1
           B      6     2
           C      6     3
           D      6     4
2014-02-01 A    NaN     1
           B    NaN     2
           C    NaN     3
           D    NaN     4

[jreback]

I'll mark this as a api-issue. I think that the 'shortcuts' about prob could work, they just may not be hooked up exactly right.

[jreback]

a possibility here would be to allow this:

df.loc(align=False)[idx['2014-02-01',:], 'Col2'] = df.loc[idx['2014-01-01',:], 'Col2']

IOW a keyword that would essentially do .values on the rhs and forgo alignment

[jim22k]

df.loc(align=False)[] looks really strange to me. Using .values on the rhs works fine.

If I want to avoid dropping into implicit alignment via numpy arrays, I found a way.

rhs = df.loc[idx['2014-01-01', :], 'Col2']
rhs.index.set_levels([pd.Timestamp('2014-02-01')], level=0, inplace=True)
df.loc[idx['2014-02-01', :], 'Col2'] = rhs

It's a little verbose, but functional. Between that and .values, I have two working solutions.

Thanks for the tip about idx = pd.IndexSlice. That makes working with multi-indexes much nicer.

[adamdivak]

I just ran into the same issue and was about to report it when I found it has already been reported. I think this is a bug and very inconsistent, I see no reason why some of the assignments below should result in NaNs.

df = pd.DataFrame([('a', 0, 1), ('a', 1, 2),
                   ('b', 0, 1), ('b', 1, 2)], 
                  columns=['c1', 'c2', 'c3'])
df = df.set_index(['c1', 'c2'])
df = df.sortlevel(0)
print(df)

           c3
    c1 c2    
    a  0    1
       1    2
    b  0    1
       1    2

df.loc['a'] = 1
print(df.loc['a'])

        c3
    c2    
    0    1
    1    1

df.loc['b'] -= 1
print(df.loc['b'])

        c3
    c2    
    0  NaN
    1  NaN

df.loc['a'] = df.loc['b'].values
print(df.loc['a'])

        c3
    c2    
    0    1
    1    2

df.loc['a'] = df.loc['b']
print(df.loc['a'])

        c3
    c2    
    0  NaN
    1  NaN

[jreback]

@yosuah you realize this shouldn't work, right. the point of pandas is to ALIGN data. This is correct as these are different columns, they don't align.

the only issue here is the inplace arithmetic.

df.loc['a'] = df.loc['b']
print(df.loc['a'])

        c3
    c2    
    0  NaN
    1  NaN