熊猫：回填DataFrameGroupBy对象

Question

I have a DataFrame object that looks like this:

Out[9]:         pts   res  sensor
        0         0     Y   accel
        1         0   NaN   accel
        2         0     N    beta
        3         0   NaN    beta
        4         5   NaN    beta
        5         8   NaN   accel

I would like to write some code that first uses the .groupby() function to group by sensor . And then backfills the pts column of each group and forward fills the res column of each group. My code attempt looks like this:

df_g = df.groupby('sensor')
next_pts = pd.Series(df_g.pts.bfill())
next_res = pd.Series(df_g.res.ffill())

df['next_pts'] = next_pts
df['next_res'] = next_res
df

The output was this:

Out[11]:         pts  res  sensor next_pts next_res
        0         0     Y   accel       0         Y
        1         0   NaN   accel       0         Y
        2         0     N    beta       0         N
        3         0   NaN    beta       0         N
        4         5   NaN    beta       5         N
        5         8   NaN   accel       8         Y

So it seems like the ffill() on the res column worked, but the bfill() on the pts column didn't. How do I make it look like this?:

Out[12]:         pts  res  sensor next_pts next_res
        0         0     Y   accel       8         Y
        1         0   NaN   accel       8         Y
        2         0     N    beta       5         N
        3         0   NaN    beta       5         N
        4         5   NaN    beta       5         N
        5         8   NaN   accel       8         Y

I found this Stack Overflow link that asks a similar question but on a DataFrame object, not a DataFrameGroupBy object: How to copy pandas DataFrame values down to fill 0's?

But when I tried to do something like this on my DataFrameGroupBy object, it threw an error: Cannot access callable attribute 'astype' of 'SeriesGroupBy' objects, try using the 'apply' method

Any help would be so much appreciated!!

Answer 1

It doesn't look like it has a lot to do with groupby , but rather with the ?fill functions, which don't fill int series with 0s.

Perhaps there's a more elegant way to do it, but this worked:

>> df.pts = np.where(df.pts == 0, np.NaN, df.pts)
>> df.pts.groupby(df.sensor).apply(lambda g: g.bfill())
0    8
1    8
2    5
3    5
4    5
5    8
dtype: float64

Note that you can convert the float series back to ints easily using .astype .

Edit The first line can be written as

>> df.pts.replace(0, np.NaN)