Pandas DataFrame: select row-wise max value in absolute terms

I have a Dataframe with only numeric data:

[ In1]: df = pd.DataFrame(np.random.randn(5, 3).round(2), columns=['A', 'B', 'C'])
        df

[Out1]:	        A	    B	    C
        0	-0.27	 1.22	 1.10
        1	-3.22	 0.48	-1.64
        2	 1.42	 0.24	-0.12
        3	-1.12	 0.44	 0.23
        4	 1.88	-0.38	 0.62

How do I select, for each row, the max value in absolute terms while preserving the sign?

In this case it would be:

0     1.22
1    -3.22
2     1.42
3    -1.12
4     1.88

I got as far as determining which column to use:

[ In2]: loc_max = df.abs().idxmax(axis=1)
        loc_max

[Out2]: 
        0    B
        1    A
        2    A
        3    A
        4    A

Performance is important because my actual dataframe is big.

SOLUTIONS COMPARISON:

All answers below will give the desired outcome.

Performance comparison on a slightly bigger dataframe:

df = pd.DataFrame(np.random.randn(1000, 100).round(2))

def numpy_argmax():
    idx_max = np.abs(df.values).argmax(axis=1)
    val = df.values[range(len(df)), idx_max]
    return pd.Series(val, index=df.index)

def check_sign():
    row_max = df.abs().max(axis=1)
    return row_max * (-1) ** df.ne(row_max, axis=0).all(axis=1)

def loop_rows():
    return df.apply(lambda s: s[s.abs().idxmax()], axis=1)

def pandas_loc():
    s = df.abs().idxmax(axis=1)
    val = [df.loc[x, y] for x, y in zip(s.index, s)]
    return pd.Series(val, index=df.index)

%timeit numpy_argmax()
%timeit check_sign()
%timeit loop_rows()
%timeit pandas_loc()

Results:

As usual going to the numpy level behind pandas curtain achieves the best performance. (Let me know if that's not always true.)

Let us use argmax on the absolute values to find the indices of the maximum absolute values. Then use these indices to get the corresponding values from each row.

v = df.values
v[range(len(v)), np.abs(v).argmax(axis=1)]

array([ 1.22, -3.22,  1.42, -1.12,  1.88])

A method that builds on your attempt, using numpy indices and numpy broadcasting rules:

import pandas as pd

df = pd.DataFrame(np.random.randn(5, 3).round(2), columns=['A', 'B', 'C'])
idx_max = np.argmax( df.abs(), axis=1)
df.values[range(len(df)), idx_max]

Another possible solution:

s = df.abs().idxmax(axis=1)
[df.loc[x, y] for x, y in zip(s.index, s)]