get non numerical rows in a column pandas python

user_5516874 0
Viewed 36990

I checked this post: finding non-numeric rows in dataframe in pandas? but it doesn't really answer my question.

my sample data:

import pandas as pd


d = {
 'unit': ['UD', 'UD', 'UD', 'UD', 'UD','UD'],
 'N-D': [ 'Q1', 'Q2', 'Q3', 'Q4','Q5','Q6'],
 'num' : [ -1.48, 1.7, -6.18, 0.25, 'sum(d)', 0.25]

}
df = pd.DataFrame(d)

it looks like this:

  N-D   num   unit
0  Q1  -1.48   UD
1  Q2   1.70   UD
2  Q3  -6.18   UD
3  Q4   0.25   UD
4  Q5   sum(d) UD
5  Q6   0.25   UD

I want to filter out only the rows in column 'num' that are NON-NUMERIC. I want all of the columns for only the rows that contain non-numeric values for column 'num'.

desired output:

  N-D   num   unit
4  Q5   sum(d) UD

my attempts:

nonnumeric=df[~df.applymap(np.isreal).all(1)] #didn't work, it pulled out everything, besides i want the condition to check only column 'num'. 

nonnumeric=df['num'][~df.applymap(np.isreal).all(1)] #didn't work, it pulled out all the rows for column 'num' only.
pandas python
5 Answers

Old topic, but if the numbers have been converted to 'str', type(x) == str is not working.

Instead, it's better to use isnumeric() or isdigit().

df = df[df['num'].apply(lambda x: not x.isnumeric())]

I tested all three approaches on my own dataframe with 200k+ rows, assuming numbers have been converted to 'str' by pd.read_csv().

def f1():
    df[pd.to_numeric(df['num'], errors='coerce').isnull()]

def f2():
    df[~df.num.str.match('^\-?(\d*\.?\d+|\d+\.?\d*)$')]

def f3():
    df[df['num'].apply(lambda x: not x.isnumeric())]

I got following execution times by running each function 10 times.

timeit.timeit(f1, number=10)
1.04128568888882

timeit.timeit(f2, number=10)
1.959099448888992

timeit.timeit(f3, number=10)
0.48741375999998127

Conculsion: fastest method is isnumeric(), slowest is regular expression method.

=========================================

Edit: As @set92 commented, isnumeric() works for integer only. So the fastest applicable function is pd.to_numeric() to have a universal solutions works for any type of numerical values.

It is possible to define a isfloat() function in python; but it will be slower than internal functions, especially for big DataFrames.

tmp=['4.0','4','4.5','1','test']*200000
df=pd.DataFrame(data=tmp,columns=['num'])


def f1():
    df[pd.to_numeric(df['num'], errors='coerce').isnull()]

def f2():
    df[df['num'].apply(lambda x: not isfloat(x))] 

def f3():
    df[~df.num.str.match('^\-?(\d*\.?\d+|\d+\.?\d*)$')]


print('to_numeric:',timeit.timeit(f1, number=10))
print('isfloat:',timeit.timeit(f2, number=10))
print('regular exp:',timeit.timeit(f3, number=10))

Results:

to_numeric: 8.303612694763615
isfloat: 9.972200270603594
regular exp: 11.420604273894583
user_105375710

There are many ways to detect non-numeric values in the column of pandas DataFrame, here is one.

df[~df['num'].map(lambda x:x.isnumeric())]
user_34120330
Related