Store numpy.array in cells of a Pandas.DataFrame

Question

I have a dataframe in which I would like to store 'raw' numpy.array:

df['COL_ARRAY'] = df.apply(lambda r: np.array(do_something_with_r), axis=1)

but it seems that pandas tries to 'unpack' the numpy.array.

Is there a workaround? Other than using a wrapper (see edit below)?

I tried reduce=False with no success.

EDIT

This works, but I have to use the 'dummy' Data class to wrap around the array, which is unsatisfactory and not very elegant.

class Data:
    def __init__(self, v):
        self.v = v

meas = pd.read_excel(DATA_FILE)
meas['DATA'] = meas.apply(
    lambda r: Data(np.array(pd.read_csv(r['filename'])))),
    axis=1
)

score 68 · Accepted Answer · edited Mar 15 '18 at 21:30

68

Use a wrapper around the numpy array i.e. pass the numpy array as list

a = np.array([5, 6, 7, 8])
df = pd.DataFrame({"a": [a]})

Output:

             a
0  [5, 6, 7, 8]

Or you can use apply(np.array) by creating the tuples i.e. if you have a dataframe

df = pd.DataFrame({'id': [1, 2, 3, 4],
                   'a': ['on', 'on', 'off', 'off'],
                   'b': ['on', 'off', 'on', 'off']})

df['new'] = df.apply(lambda r: tuple(r), axis=1).apply(np.array)

Output :

     a    b  id            new
0   on   on   1    [on, on, 1]
1   on  off   2   [on, off, 2]
2  off   on   3   [off, on, 3]
3  off  off   4  [off, off, 4]

df['new'][0]

Output :

array(['on', 'on', '1'], dtype='<U2')

edited Mar 15 '18 at 21:30

Engineero

12,340
5
53
75

answered Aug 07 '17 at 13:46

Bharath M Shetty

30,075
6
57
108

That works, but then I'd rather use a dummy class instead of a list. – Cedric H. Aug 07 '17 at 13:48
Does that work if instead of `tuple(r)` you do something like `np.array([[1,2],[3,4]])`, ie. a 2-dim array? – Cedric H. Aug 07 '17 at 14:20
`tuple(r)` works with 2D array too. Did you mean replacing tuple(r) with 2D np.array? – Bharath M Shetty Aug 07 '17 at 14:27
Yes. I mean, I understand your solution, and it works, but what if I want to have a 2D np.array in the `new` column (and not a 1D array as shown)? – Cedric H. Aug 07 '17 at 14:33
Can you add the expected output in your question? All the elements in the row should be of numpy array if you want to create a new 2D array. My solution works in that case. If its mixed type you have to use if else first to make it a numpy array. – Bharath M Shetty Aug 07 '17 at 14:37

score 29 · Answer 2 · answered Feb 05 '20 at 19:10

29

If you first set a column to have type object, you can insert an array without any wrapping:

df = pd.DataFrame(columns=[1])
df[1] = df[1].astype(object)
df.loc[1, 1] = np.array([5, 6, 7, 8])
df

Output:

    1
1   [5, 6, 7, 8]

answered Feb 05 '20 at 19:10

David Wasserman

551
5
10

score 6 · Answer 3 · answered Aug 07 '17 at 19:25

6

You can wrap the Data Frame data args in square brackets to maintain the np.array in each cell:

one_d_array = np.array([1,2,3])
two_d_array = one_d_array*one_d_array[:,np.newaxis]
two_d_array

array([[1, 2, 3],
       [2, 4, 6],
       [3, 6, 9]])


pd.DataFrame([
    [one_d_array],
    [two_d_array] ])

                                   0
0                          [1, 2, 3]
1  [[1, 2, 3], [2, 4, 6], [3, 6, 9]]

answered Aug 07 '17 at 19:25

user1717828

7,122
8
34
59

Those become `list`s no? – WestCoastProjects Nov 17 '18 at 19:26
1

@javadba I set the above output to `df` and see `print(type(df.iloc[0,0]))` --> __ This is Pandas 0.23.0. Are you seeing something different with another version? – user1717828 Nov 17 '18 at 19:48
1

You are correct actually - one of the more highly upvoted answers made it sound like the ndarray would be converted to a list. I just tested this. `without` the brackets we get `rows` in the `df` and `with` brackets we get `ndarray`s as you say – WestCoastProjects Nov 17 '18 at 20:20

yuzhen_3301 · Answer 4 · 2019-06-13T03:21:10.410

1

Suppose you have a DataFrame ds and it has a column named as 'class'. If ds['class'] contains strings or numbers, and you want to change them with numpy.ndarrays or lists, the following code would help. In the code, class2vector is a numpy.ndarray or list and ds_class is a filter condition.

ds['class'] = ds['class'].map(lambda x: class2vector if (isinstance(x, str) and (x == ds_class)) else x)

edited Jun 13 '19 at 03:21

answered Jun 13 '19 at 02:03

yuzhen_3301

19
3

score 1 · Answer 5 · answered Sep 23 '21 at 06:06

1

choose eval buildin function is easy to use and easy to read.

# First ensure use object store str
df['col2'] = self.df['col2'].astype(object)
# read
arr_obj = eval(df.at[df[df.col_1=='xyz'].index[0], 'col2']))
# write
df.at[df[df.col_1=='xyz'].index[0], 'col2'] = str(arr_obj)

real store display perfect human readable value:

col_1,  col_2
xyz,    "['aaa', 'bbb', 'ccc', 'ddd']"

answered Sep 23 '21 at 06:06

Kenneth

21
2

Using eval() is usually not a good idea. Transforming freely between strings and code is insecure and confuses the IDE/linter/type-checker/other people reading your code. Consider using the other methods mentioned in this post. – vvolhejn Apr 11 '23 at 08:00

score 1 · Answer 6 · answered Jun 08 '23 at 07:58

1

Preset the type of your column to object, this will allow you to store a NumPy array as-is:

df['COL_ARRAY'] = pd.Series(dtype='object')

df['COL_ARRAY'] = df.apply(lambda r: np.array(do_something_with_r), axis=1)

answered Jun 08 '23 at 07:58

Begoodpy

1,018
3
8
20

score 0 · Answer 7 · answered Jan 18 '20 at 09:02

Just wrap what you want to store in a cell to a list object through first apply, and extract it by index 0of that list through second apply:

import pandas as pd
import numpy as np

df = pd.DataFrame({'id': [1, 2, 3, 4],
                   'a': ['on', 'on', 'off', 'off'],
                   'b': ['on', 'off', 'on', 'off']})


df['new'] = df.apply(lambda x: [np.array(x)], axis=1).apply(lambda x: x[0])

df

output:

    id  a       b       new
0   1   on      on      [1, on, on]
1   2   on      off     [2, on, off]
2   3   off     on      [3, off, on]
3   4   off     off     [4, off, off]

score 0 · Answer 8 · answered Oct 08 '22 at 17:24

0

Here goes my 2 cents contribution (tested on Python 3.7):

import pandas as pd
import numpy as np

dataArray = np.array([0.0, 1.0, 2.0])
df = pd.DataFrame()
df['User Col A'] = [1]
df['Array'] = [dataArray]

answered Oct 08 '22 at 17:24

Marcelo

65
2
10

score 0 · Answer 9 · answered Feb 06 '23 at 05:32

If you only want some of the columns you could do something lie this. Taking the example of @allenyllee,

df = pd.DataFrame({'id': [1, 2, 3, 4],
                   'a': ['on', 'on', 'off', 'off'],
                   'b': ['on', 'off', 'on', 'off']})

df['new'] = df[['a','b']].apply(lambda x: np.array(x), axis=1)

which outputs

   id    a    b         new
0   1   on   on    [on, on]
1   2   on  off   [on, off]
2   3  off   on   [off, on]
3   4  off  off  [off, off]

you can also change the order of [['a', 'b']] if you need a specific order.

Store numpy.array in cells of a Pandas.DataFrame

9 Answers9

Linked