Pandas Dataframes: Apply Examples

Pandas Dataframes: Apply Examples

Last updated:
Table of Contents

WIP Alert This is a work in progress. Current information is correct but more content may be added in the future.

Pandas version 1.0+ used.

All code available online on this jupyter notebook

Apply example

To apply a function to a dataframe column, do df['my_col'].apply(function), where the function takes one element and return another value.

import pandas as pd

df = pd.DataFrame({
    'name': ['alice','bob','charlie','david'],
    'age': [25,26,27,22],
})[['name', 'age']]

# each element of the age column is a string
# so you can call .upper() on it
df['name_uppercase'] = df['name'].apply(lambda element: element.upper())

source-dataframe BEFORE: Original dataframe
         
after-applying-map AFTER: Created new column using Series.apply()

Apply example, custom function

To apply a custom function to a column, you just need to define a function that takes one element and returns a new value:

import pandas as pd

df = pd.DataFrame({
    'name': ['alice','bob','charlie','david'],
    'age': [25,26,27,22],
})

# function that takes one value, returns one value
def first_letter(input_str):
    return input_str[:1]

# pass just the function name to apply
df['first_letter'] = df['name'].apply(first_letter)

source-dataframe Source dataframe
  
alt text Crated a new column passing a
custom function to apply

Take multiple columns as parameters

Double square brackets return another dataframe instead of a series

To apply a single function using multiple columns, select columns using double square brackets ([[]]) and use axis=1:

import pandas as pd

df = pd.DataFrame({
    'name': ['alice','bob','charlie','david'],
    'age': [25,26,27,22],
})

# define a function that takes two values, returns 1 value
def concatenate(value_1, value_2):
    return str(value_1)+ "--" + str(value_2) 

# note the use of DOUBLE SQUARE BRACKETS!
df['concatenated'] = df[['name','age']].apply(
    lambda row: concatenate(row['name'], row['age']) , axis=1)

source-dataframe Original dataframe
  
dataframe-with-new-concatenated-column Created a new column by applying a function that
takes two columns and concatenates them as strings

Apply function to row

To apply a dunction to a full row instead of a column, use axis=1 and call apply on the dataframe itself:

Example: Sum all values in each row:

import pandas as pd

df = pd.DataFrame({
    'value1': [1,2,3,4,5],
    'value2': [5,4,3,2,1],
    'value3': [10,20,30,40,50],
    'value4': [99,99,99,99,np.nan],
})

def sum_all(row):
    return np.sum(row)

# note that apply was called on the dataframe itself, not on columns
df['sum_all'] = df.apply(lambda row: sum_all(row) , axis=1)

source-dataframe-with-observations Source dataframe where each row contains
observations for one sample
  
Dataframe-with-new-column-based-on-row-application Generated a new column by summing all
values in the row, with numpy.sum

Apply function to column

Just use apply. Example here

Return multiple columns

To apply a function to a column and return multiple values so that you can create multiple columns, return a pd.Series with the values instead:

Example: produce two values from a function and assign to two columns

import pandas as pd

df = pd.DataFrame({
    'name': ['alice','bob','charlie','david','edward'],
    'age': [25,26,27,22,np.nan],
})

def times_two_times_three(value):
    value_times_2 = value*2
    value_times_3 = value*3

    return pd.Series([value_times_2,value_times_3])

# note that apply was called on age column
df[['times_2','times_3']]= df['age'].apply(times_two_times_three)

source-dataframe-with-columns Source dataframe
  
dataframe-with-two-new-columns Modified dataframe with two new columns,
both returned by apply

Apply function in parallel

If you have costly operations you need to perform on a dataframe, (e.g. text preprocessing), you can split the operation into multiple cores to decrease the running time:

import multiprocessing

import numpy as np
import pandas as pd

# how many cores do you have?
NUM_CORES=8

# replace load_large_dataframe() with your dataframe
df = load_large_dataframe()

# split the dataframe into chunks, depending on hoe many cores you have
df_chunks = np.array_split(df ,NUM_CORES)

# this is a function that takes one dataframe chunk and returns
# the processed chunk (for example, adding processed columns)
def process_df(input_df):
    # copy the dataframe to prevent mutation in place
    output_df = input_df.copy()

    # apply a function to every row *in this chunk*
      output_df['new_column'] = output_df.apply(some_function, axis=1)    

    return output_df

with multiprocessing.Pool(NUM_CORES) as pool:
    # process each chunk in a separate core and merge the results
    full_output_df = pd.concat(pool.map(process_df, df_chunks), ignore_index=True)

Vectorization and Performance

TODO

map vs apply

map() apply()
Series functionSeries function
and Dataframe function
Returns new SeriesReturns new dataframe,
possibly with a single column
Can only be applied to a single
column (one element at
a time)
Can be applied to multiple
columns at the same time
Operates on array elements,
one at a time
Operates on whole columns or rows
Very slow, no better than a
Python for loop
Much faster when you can use
numpy vectorized functions