How to do this in pandas:
I have a function extract_text_features
on a single text column, returning multiple output columns. Specifically, the function returns 6 values.
The function works, however there doesn't seem to be any proper return type (pandas DataFrame/ numpy array/ Python list) such that the output can get correctly assigned df.ix[: ,10:16] = df.textcol.map(extract_text_features)
So I think I need to drop back to iterating with df.iterrows()
, as per this?
UPDATE: Iterating with df.iterrows()
is at least 20x slower, so I surrendered and split out the function into six distinct .map(lambda ...)
calls.
UPDATE 2: this question was asked back around v0.11.0, before the useability df.apply
was improved or df.assign()
was added in v0.16. Hence much of the question and answers are not too relevant.
df.ix[: ,10:16]
. I think you'll have to merge
your features into the dataset.
apply
I usually do this using zip
:
>>> df = pd.DataFrame([[i] for i in range(10)], columns=['num'])
>>> df
num
0 0
1 1
2 2
3 3
4 4
5 5
6 6
7 7
8 8
9 9
>>> def powers(x):
>>> return x, x**2, x**3, x**4, x**5, x**6
>>> df['p1'], df['p2'], df['p3'], df['p4'], df['p5'], df['p6'] = \
>>> zip(*df['num'].map(powers))
>>> df
num p1 p2 p3 p4 p5 p6
0 0 0 0 0 0 0 0
1 1 1 1 1 1 1 1
2 2 2 4 8 16 32 64
3 3 3 9 27 81 243 729
4 4 4 16 64 256 1024 4096
5 5 5 25 125 625 3125 15625
6 6 6 36 216 1296 7776 46656
7 7 7 49 343 2401 16807 117649
8 8 8 64 512 4096 32768 262144
9 9 9 81 729 6561 59049 531441
In 2020, I use apply()
with argument result_type='expand'
applied_df = df.apply(lambda row: fn(row.text), axis='columns', result_type='expand')
df = pd.concat([df, applied_df], axis='columns')
pd.Series
which is always nice regarding performance issues
df.apply
returns a dict
, the columns will come out named according to the keys.
appiled_df =
with df[["col1", "col2", ...]] =
. This will also give named columns.
result_type='expand'
. E.g. df[new_cols] = df.apply(extract_text_features, axis=1, result_type='expand')
just works. Although you'd need to know names of the new columns.
Building off of user1827356 's answer, you can do the assignment in one pass using df.merge
:
df.merge(df.textcol.apply(lambda s: pd.Series({'feature1':s+1, 'feature2':s-1})),
left_index=True, right_index=True)
textcol feature1 feature2
0 0.772692 1.772692 -0.227308
1 0.857210 1.857210 -0.142790
2 0.065639 1.065639 -0.934361
3 0.819160 1.819160 -0.180840
4 0.088212 1.088212 -0.911788
EDIT: Please be aware of the huge memory consumption and low speed: https://ys-l.github.io/posts/2015/08/28/how-not-to-use-pandas-apply/ !
This is what I've done in the past
df = pd.DataFrame({'textcol' : np.random.rand(5)})
df
textcol
0 0.626524
1 0.119967
2 0.803650
3 0.100880
4 0.017859
df.textcol.apply(lambda s: pd.Series({'feature1':s+1, 'feature2':s-1}))
feature1 feature2
0 1.626524 -0.373476
1 1.119967 -0.880033
2 1.803650 -0.196350
3 1.100880 -0.899120
4 1.017859 -0.982141
Editing for completeness
pd.concat([df, df.textcol.apply(lambda s: pd.Series({'feature1':s+1, 'feature2':s-1}))], axis=1)
textcol feature1 feature2
0 0.626524 1.626524 -0.373476
1 0.119967 1.119967 -0.880033
2 0.803650 1.803650 -0.196350
3 0.100880 1.100880 -0.899120
4 0.017859 1.017859 -0.982141
df[['col1', 'col2']] = df['col3'].apply(lambda x: pd.Series('val1', 'val2'))
This is the correct and easiest way to accomplish this for 95% of use cases:
>>> df = pd.DataFrame(zip(*[range(10)]), columns=['num'])
>>> df
num
0 0
1 1
2 2
3 3
4 4
5 5
>>> def example(x):
... x['p1'] = x['num']**2
... x['p2'] = x['num']**3
... x['p3'] = x['num']**4
... return x
>>> df = df.apply(example, axis=1)
>>> df
num p1 p2 p3
0 0 0 0 0
1 1 1 1 1
2 2 4 8 16
3 3 9 27 81
4 4 16 64 256
pd.Series({k:v})
and serialize the column assignment like in Ewan's answer?
Just use result_type="expand"
df = pd.DataFrame(np.random.randint(0,10,(10,2)), columns=["random", "a"])
df[["sq_a","cube_a"]] = df.apply(lambda x: [x.a**2, x.a**3], axis=1, result_type="expand")
For me this worked:
Input df
df = pd.DataFrame({'col x': [1,2,3]})
col x
0 1
1 2
2 3
Function
def f(x):
return pd.Series([x*x, x*x*x])
Create 2 new columns:
df[['square x', 'cube x']] = df['col x'].apply(f)
Output:
col x square x cube x
0 1 1 1
1 2 4 8
2 3 9 27
Summary: If you only want to create a few columns, use df[['new_col1','new_col2']] = df[['data1','data2']].apply( function_of_your_choosing(x), axis=1)
For this solution, the number of new columns you are creating must be equal to the number columns you use as input to the .apply() function. If you want to do something else, have a look at the other answers.
Details Let's say you have two-column dataframe. The first column is a person's height when they are 10; the second is said person's height when they are 20.
Suppose you need to calculate both the mean of each person's heights and sum of each person's heights. That's two values per each row.
You could do this via the following, soon-to-be-applied function:
def mean_and_sum(x):
"""
Calculates the mean and sum of two heights.
Parameters:
:x -- the values in the row this function is applied to. Could also work on a list or a tuple.
"""
sum=x[0]+x[1]
mean=sum/2
return [mean,sum]
You might use this function like so:
df[['height_at_age_10','height_at_age_20']].apply(mean_and_sum(x),axis=1)
(To be clear: this apply function takes in the values from each row in the subsetted dataframe and returns a list.)
However, if you do this:
df['Mean_&_Sum'] = df[['height_at_age_10','height_at_age_20']].apply(mean_and_sum(x),axis=1)
you'll create 1 new column that contains the [mean,sum] lists, which you'd presumably want to avoid, because that would require another Lambda/Apply.
Instead, you want to break out each value into its own column. To do this, you can create two columns at once:
df[['Mean','Sum']] = df[['height_at_age_10','height_at_age_20']]
.apply(mean_and_sum(x),axis=1)
df["mean"], df["sum"] = df[['height_at_age_10','height_at_age_20']] .apply(mean_and_sum(x),axis=1)
return pd.Series([mean,sum])
I've looked several ways of doing this and the method shown here (returning a pandas series) doesn't seem to be most efficient.
If we start with a largeish dataframe of random data:
# Setup a dataframe of random numbers and create a
df = pd.DataFrame(np.random.randn(10000,3),columns=list('ABC'))
df['D'] = df.apply(lambda r: ':'.join(map(str, (r.A, r.B, r.C))), axis=1)
columns = 'new_a', 'new_b', 'new_c'
The example shown here:
# Create the dataframe by returning a series
def method_b(v):
return pd.Series({k: v for k, v in zip(columns, v.split(':'))})
%timeit -n10 -r3 df.D.apply(method_b)
10 loops, best of 3: 2.77 s per loop
An alternative method:
# Create a dataframe from a series of tuples
def method_a(v):
return v.split(':')
%timeit -n10 -r3 pd.DataFrame(df.D.apply(method_a).tolist(), columns=columns)
10 loops, best of 3: 8.85 ms per loop
By my reckoning it's far more efficient to take a series of tuples and then convert that to a DataFrame. I'd be interested to hear people's thinking though if there's an error in my working.
The accepted solution is going to be extremely slow for lots of data. The solution with the greatest number of upvotes is a little difficult to read and also slow with numeric data. If each new column can be calculated independently of the others, I would just assign each of them directly without using apply
.
Example with fake character data
Create 100,000 strings in a DataFrame
df = pd.DataFrame(np.random.choice(['he jumped', 'she ran', 'they hiked'],
size=100000, replace=True),
columns=['words'])
df.head()
words
0 she ran
1 she ran
2 they hiked
3 they hiked
4 they hiked
Let's say we wanted to extract some text features as done in the original question. For instance, let's extract the first character, count the occurrence of the letter 'e' and capitalize the phrase.
df['first'] = df['words'].str[0]
df['count_e'] = df['words'].str.count('e')
df['cap'] = df['words'].str.capitalize()
df.head()
words first count_e cap
0 she ran s 1 She ran
1 she ran s 1 She ran
2 they hiked t 2 They hiked
3 they hiked t 2 They hiked
4 they hiked t 2 They hiked
Timings
%%timeit
df['first'] = df['words'].str[0]
df['count_e'] = df['words'].str.count('e')
df['cap'] = df['words'].str.capitalize()
127 ms ± 585 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
def extract_text_features(x):
return x[0], x.count('e'), x.capitalize()
%timeit df['first'], df['count_e'], df['cap'] = zip(*df['words'].apply(extract_text_features))
101 ms ± 2.96 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
Surprisingly, you can get better performance by looping through each value
%%timeit
a,b,c = [], [], []
for s in df['words']:
a.append(s[0]), b.append(s.count('e')), c.append(s.capitalize())
df['first'] = a
df['count_e'] = b
df['cap'] = c
79.1 ms ± 294 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
Another example with fake numeric data
Create 1 million random numbers and test the powers
function from above.
df = pd.DataFrame(np.random.rand(1000000), columns=['num'])
def powers(x):
return x, x**2, x**3, x**4, x**5, x**6
%%timeit
df['p1'], df['p2'], df['p3'], df['p4'], df['p5'], df['p6'] = \
zip(*df['num'].map(powers))
1.35 s ± 83.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
Assigning each column is 25x faster and very readable:
%%timeit
df['p1'] = df['num'] ** 1
df['p2'] = df['num'] ** 2
df['p3'] = df['num'] ** 3
df['p4'] = df['num'] ** 4
df['p5'] = df['num'] ** 5
df['p6'] = df['num'] ** 6
51.6 ms ± 1.9 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
I made a similar response with more details here on why apply
is typically not the way to go.
Have posted the same answer in two other similar questions. The way I prefer to do this is to wrap up the return values of the function in a series:
def f(x):
return pd.Series([x**2, x**3])
And then use apply as follows to create separate columns:
df[['x**2','x**3']] = df.apply(lambda row: f(row['x']), axis=1)
def extract_text_features(feature):
...
...
return pd.Series((feature1, feature2))
df[['NewFeature1', 'NewFeature1']] = df[['feature']].apply(extract_text_features, axis=1)
Here the a dataframe with a single feature is being converted to two new features. Give this a try too.
you can return the entire row instead of values:
df = df.apply(extract_text_features,axis = 1)
where the function returns the row
def extract_text_features(row):
row['new_col1'] = value1
row['new_col2'] = value2
return row
extract_text_features
to every column of the df, only to the text column df.textcol
I have a more complicated situation, the dataset has a nested structure:
import json
data = '{"TextID":{"0":"0038f0569e","1":"003eb6998d","2":"006da49ea0"},"Summary":{"0":{"Crisis_Level":["c"],"Type":["d"],"Special_Date":["a"]},"1":{"Crisis_Level":["d"],"Type":["a","d"],"Special_Date":["a"]},"2":{"Crisis_Level":["d"],"Type":["a"],"Special_Date":["a"]}}}'
df = pd.DataFrame.from_dict(json.loads(data))
print(df)
output:
TextID Summary
0 0038f0569e {'Crisis_Level': ['c'], 'Type': ['d'], 'Specia...
1 003eb6998d {'Crisis_Level': ['d'], 'Type': ['a', 'd'], 'S...
2 006da49ea0 {'Crisis_Level': ['d'], 'Type': ['a'], 'Specia...
The Summary
column contains dict objects, so I use apply
with from_dict
and stack
to extract each row of dict:
df2 = df.apply(
lambda x: pd.DataFrame.from_dict(x[1], orient='index').stack(), axis=1)
print(df2)
output:
Crisis_Level Special_Date Type
0 0 0 1
0 c a d NaN
1 d a a d
2 d a a NaN
Looks good, but missing the TextID
column. To get TextID
column back, I've tried three approach:
Modify apply to return multiple columns: df_tmp = df.copy() df_tmp[['TextID', 'Summary']] = df.apply( lambda x: pd.Series([x[0], pd.DataFrame.from_dict(x[1], orient='index').stack()]), axis=1) print(df_tmp) output: TextID Summary 0 0038f0569e Crisis_Level 0 c Type 0 d Spec... 1 003eb6998d Crisis_Level 0 d Type 0 a ... 2 006da49ea0 Crisis_Level 0 d Type 0 a Spec... But this is not what I want, the Summary structure are flatten. Use pd.concat: df_tmp2 = pd.concat([df['TextID'], df2], axis=1) print(df_tmp2) output: TextID (Crisis_Level, 0) (Special_Date, 0) (Type, 0) (Type, 1) 0 0038f0569e c a d NaN 1 003eb6998d d a a d 2 006da49ea0 d a a NaN Looks fine, the MultiIndex column structure are preserved as tuple. But check columns type: df_tmp2.columns output: Index(['TextID', ('Crisis_Level', 0), ('Special_Date', 0), ('Type', 0), ('Type', 1)], dtype='object') Just as a regular Index class, not MultiIndex class. use set_index: Turn all columns you want to preserve into row index, after some complicated apply function and then reset_index to get columns back: df_tmp3 = df.set_index('TextID') df_tmp3 = df_tmp3.apply( lambda x: pd.DataFrame.from_dict(x[0], orient='index').stack(), axis=1) df_tmp3 = df_tmp3.reset_index(level=0) print(df_tmp3) output: TextID Crisis_Level Special_Date Type 0 0 0 1 0 0038f0569e c a d NaN 1 003eb6998d d a a d 2 006da49ea0 d a a NaN Check the type of columns df_tmp3.columns output: MultiIndex(levels=[['Crisis_Level', 'Special_Date', 'Type', 'TextID'], [0, 1, '']], codes=[[3, 0, 1, 2, 2], [2, 0, 0, 0, 1]])
So, If your apply
function will return MultiIndex
columns, and you want to preserve it, you may want to try the third method.
Success story sharing
temp = list(zip(*df['num'].map(powers))); for i, c in enumerate(columns): df[c] = temp[c]
for i, c in enumerate(columns): df[c] = temp[i]
. Thanks to this, I really got the purpose ofenumerate
:Dzip(*df['col'].map(function))
is probably the way to go.