Table Of Contents

Search

Enter search terms or a module, class or function name.

Intro to Data Structures

We’ll start with a quick, non-comprehensive overview of the fundamental data structures in pandas to get you started. The fundamental behavior about data types, indexing, and axis labeling / alignment apply across all of the objects. To get started, import numpy and load pandas into your namespace:

In [1]: import numpy as np

# will use a lot in examples
In [2]: randn = np.random.randn

In [3]: from pandas import *

Here is a basic tenet to keep in mind: data alignment is intrinsic. The link between labels and data will not be broken unless done so explicitly by you.

We’ll give a brief intro to the data structures, then consider all of the broad categories of functionality and methods in separate sections.

When using pandas, we recommend the following import convention:

import pandas as pd

Series

Warning

In 0.13.0 Series has internaly been refactored to no longer sub-class ndarray but instead subclass NDFrame, similarly to the rest of the pandas containers. This should be a transparent change with only very limited API implications (See the Internal Refactoring)

Series is a one-dimensional labeled array capable of holding any data type (integers, strings, floating point numbers, Python objects, etc.). The axis labels are collectively referred to as the index. The basic method to create a Series is to call:

>>> s = Series(data, index=index)

Here, data can be many different things:

  • a Python dict
  • an ndarray
  • a scalar value (like 5)

The passed index is a list of axis labels. Thus, this separates into a few cases depending on what data is:

From ndarray

If data is an ndarray, index must be the same length as data. If no index is passed, one will be created having values [0, ..., len(data) - 1].

In [4]: s = Series(randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [5]: s
Out[5]: 
a   -2.783
b    0.426
c   -0.650
d    1.146
e   -0.663
dtype: float64

In [6]: s.index
Out[6]: Index([u'a', u'b', u'c', u'd', u'e'], dtype='object')

In [7]: Series(randn(5))
Out[7]: 
0    0.294
1   -0.405
2    1.167
3    0.842
4    0.540
dtype: float64

Note

Starting in v0.8.0, pandas supports non-unique index values. If an operation that does not support duplicate index values is attempted, an exception will be raised at that time. The reason for being lazy is nearly all performance-based (there are many instances in computations, like parts of GroupBy, where the index is not used).

From dict

If data is a dict, if index is passed the values in data corresponding to the labels in the index will be pulled out. Otherwise, an index will be constructed from the sorted keys of the dict, if possible.

In [8]: d = {'a' : 0., 'b' : 1., 'c' : 2.}

In [9]: Series(d)
Out[9]: 
a    0
b    1
c    2
dtype: float64

In [10]: Series(d, index=['b', 'c', 'd', 'a'])
Out[10]: 
b     1
c     2
d   NaN
a     0
dtype: float64

Note

NaN (not a number) is the standard missing data marker used in pandas

From scalar value If data is a scalar value, an index must be provided. The value will be repeated to match the length of index

In [11]: Series(5., index=['a', 'b', 'c', 'd', 'e'])
Out[11]: 
a    5
b    5
c    5
d    5
e    5
dtype: float64

Series is ndarray-like

Series acts very similarly to a ndarray, and is a valid argument to most NumPy functions. However, things like slicing also slice the index.

In [12]: s[0]
Out[12]: -2.7827595933769942

In [13]: s[:3]
Out[13]: 
a   -2.783
b    0.426
c   -0.650
dtype: float64

In [14]: s[s > s.median()]
Out[14]: 
b    0.426
d    1.146
dtype: float64

In [15]: s[[4, 3, 1]]
Out[15]: 
e   -0.663
d    1.146
b    0.426
dtype: float64

In [16]: np.exp(s)
Out[16]: 
a    0.062
b    1.532
c    0.522
d    3.147
e    0.515
dtype: float64

We will address array-based indexing in a separate section.

Series is dict-like

A Series is like a fixed-size dict in that you can get and set values by index label:

In [17]: s['a']
Out[17]: -2.7827595933769942

In [18]: s['e'] = 12.

In [19]: s
Out[19]: 
a    -2.783
b     0.426
c    -0.650
d     1.146
e    12.000
dtype: float64

In [20]: 'e' in s
Out[20]: True

In [21]: 'f' in s
Out[21]: False

If a label is not contained, an exception is raised:

>>> s['f']
KeyError: 'f'

Using the get method, a missing label will return None or specified default:

In [22]: s.get('f')

In [23]: s.get('f', np.nan)
Out[23]: nan

See also the section on attribute access.

Vectorized operations and label alignment with Series

When doing data analysis, as with raw NumPy arrays looping through Series value-by-value is usually not necessary. Series can be also be passed into most NumPy methods expecting an ndarray.

In [24]: s + s
Out[24]: 
a    -5.566
b     0.853
c    -1.301
d     2.293
e    24.000
dtype: float64

In [25]: s * 2
Out[25]: 
a    -5.566
b     0.853
c    -1.301
d     2.293
e    24.000
dtype: float64

In [26]: np.exp(s)
Out[26]: 
a         0.062
b         1.532
c         0.522
d         3.147
e    162754.791
dtype: float64

A key difference between Series and ndarray is that operations between Series automatically align the data based on label. Thus, you can write computations without giving consideration to whether the Series involved have the same labels.

In [27]: s[1:] + s[:-1]
Out[27]: 
a      NaN
b    0.853
c   -1.301
d    2.293
e      NaN
dtype: float64

The result of an operation between unaligned Series will have the union of the indexes involved. If a label is not found in one Series or the other, the result will be marked as missing NaN. Being able to write code without doing any explicit data alignment grants immense freedom and flexibility in interactive data analysis and research. The integrated data alignment features of the pandas data structures set pandas apart from the majority of related tools for working with labeled data.

Note

In general, we chose to make the default result of operations between differently indexed objects yield the union of the indexes in order to avoid loss of information. Having an index label, though the data is missing, is typically important information as part of a computation. You of course have the option of dropping labels with missing data via the dropna function.

Name attribute

Series can also have a name attribute:

In [28]: s = Series(np.random.randn(5), name='something')

In [29]: s
Out[29]: 
0    0.541
1   -1.175
2    0.129
3    0.043
4   -0.429
Name: something, dtype: float64

In [30]: s.name
Out[30]: 'something'

The Series name will be assigned automatically in many cases, in particular when taking 1D slices of DataFrame as you will see below.

DataFrame

DataFrame is a 2-dimensional labeled data structure with columns of potentially different types. You can think of it like a spreadsheet or SQL table, or a dict of Series objects. It is generally the most commonly used pandas object. Like Series, DataFrame accepts many different kinds of input:

  • Dict of 1D ndarrays, lists, dicts, or Series
  • 2-D numpy.ndarray
  • Structured or record ndarray
  • A Series
  • Another DataFrame

Along with the data, you can optionally pass index (row labels) and columns (column labels) arguments. If you pass an index and / or columns, you are guaranteeing the index and / or columns of the resulting DataFrame. Thus, a dict of Series plus a specific index will discard all data not matching up to the passed index.

If axis labels are not passed, they will be constructed from the input data based on common sense rules.

From dict of Series or dicts

The result index will be the union of the indexes of the various Series. If there are any nested dicts, these will be first converted to Series. If no columns are passed, the columns will be the sorted list of dict keys.

In [31]: d = {'one' : Series([1., 2., 3.], index=['a', 'b', 'c']),
   ....:      'two' : Series([1., 2., 3., 4.], index=['a', 'b', 'c', 'd'])}
   ....: 

In [32]: df = DataFrame(d)

In [33]: df
Out[33]: 
   one  two
a    1    1
b    2    2
c    3    3
d  NaN    4

In [34]: DataFrame(d, index=['d', 'b', 'a'])
Out[34]: 
   one  two
d  NaN    4
b    2    2
a    1    1

In [35]: DataFrame(d, index=['d', 'b', 'a'], columns=['two', 'three'])
Out[35]: 
   two three
d    4   NaN
b    2   NaN
a    1   NaN

The row and column labels can be accessed respectively by accessing the index and columns attributes:

Note

When a particular set of columns is passed along with a dict of data, the passed columns override the keys in the dict.

In [36]: df.index
Out[36]: Index([u'a', u'b', u'c', u'd'], dtype='object')

In [37]: df.columns
Out[37]: Index([u'one', u'two'], dtype='object')

From dict of ndarrays / lists

The ndarrays must all be the same length. If an index is passed, it must clearly also be the same length as the arrays. If no index is passed, the result will be range(n), where n is the array length.

In [38]: d = {'one' : [1., 2., 3., 4.],
   ....:      'two' : [4., 3., 2., 1.]}
   ....: 

In [39]: DataFrame(d)
Out[39]: 
   one  two
0    1    4
1    2    3
2    3    2
3    4    1

In [40]: DataFrame(d, index=['a', 'b', 'c', 'd'])
Out[40]: 
   one  two
a    1    4
b    2    3
c    3    2
d    4    1

From structured or record array

This case is handled identically to a dict of arrays.

In [41]: data = np.zeros((2,),dtype=[('A', 'i4'),('B', 'f4'),('C', 'a10')])

In [42]: data[:] = [(1,2.,'Hello'),(2,3.,"World")]

In [43]: DataFrame(data)
Out[43]: 
   A  B      C
0  1  2  Hello
1  2  3  World

In [44]: DataFrame(data, index=['first', 'second'])
Out[44]: 
        A  B      C
first   1  2  Hello
second  2  3  World

In [45]: DataFrame(data, columns=['C', 'A', 'B'])
Out[45]: 
       C  A  B
0  Hello  1  2
1  World  2  3

Note

DataFrame is not intended to work exactly like a 2-dimensional NumPy ndarray.

From a list of dicts

In [46]: data2 = [{'a': 1, 'b': 2}, {'a': 5, 'b': 10, 'c': 20}]

In [47]: DataFrame(data2)
Out[47]: 
   a   b   c
0  1   2 NaN
1  5  10  20

In [48]: DataFrame(data2, index=['first', 'second'])
Out[48]: 
        a   b   c
first   1   2 NaN
second  5  10  20

In [49]: DataFrame(data2, columns=['a', 'b'])
Out[49]: 
   a   b
0  1   2
1  5  10

From a dict of tuples

You can automatically create a multi-indexed frame by passing a tuples dictionary

In [50]: DataFrame({('a', 'b'): {('A', 'B'): 1, ('A', 'C'): 2},
   ....:            ('a', 'a'): {('A', 'C'): 3, ('A', 'B'): 4},
   ....:            ('a', 'c'): {('A', 'B'): 5, ('A', 'C'): 6},
   ....:            ('b', 'a'): {('A', 'C'): 7, ('A', 'B'): 8},
   ....:            ('b', 'b'): {('A', 'D'): 9, ('A', 'B'): 10}})
   ....: 
Out[50]: 
      a           b    
      a   b   c   a   b
A B   4   1   5   8  10
  C   3   2   6   7 NaN
  D NaN NaN NaN NaN   9

From a Series

The result will be a DataFrame with the same index as the input Series, and with one column whose name is the original name of the Series (only if no other column name provided).

Missing Data

Much more will be said on this topic in the Missing data section. To construct a DataFrame with missing data, use np.nan for those values which are missing. Alternatively, you may pass a numpy.MaskedArray as the data argument to the DataFrame constructor, and its masked entries will be considered missing.

Alternate Constructors

DataFrame.from_dict

DataFrame.from_dict takes a dict of dicts or a dict of array-like sequences and returns a DataFrame. It operates like the DataFrame constructor except for the orient parameter which is 'columns' by default, but which can be set to 'index' in order to use the dict keys as row labels.

DataFrame.from_records

DataFrame.from_records takes a list of tuples or an ndarray with structured dtype. Works analogously to the normal DataFrame constructor, except that index maybe be a specific field of the structured dtype to use as the index. For example:

In [51]: data
Out[51]: 
array([(1, 2.0, 'Hello'), (2, 3.0, 'World')], 
      dtype=[('A', '<i4'), ('B', '<f4'), ('C', 'S10')])

In [52]: DataFrame.from_records(data, index='C')
Out[52]: 
       A  B
C          
Hello  1  2
World  2  3

DataFrame.from_items

DataFrame.from_items works analogously to the form of the dict constructor that takes a sequence of (key, value) pairs, where the keys are column (or row, in the case of orient='index') names, and the value are the column values (or row values). This can be useful for constructing a DataFrame with the columns in a particular order without having to pass an explicit list of columns:

In [53]: DataFrame.from_items([('A', [1, 2, 3]), ('B', [4, 5, 6])])
Out[53]: 
   A  B
0  1  4
1  2  5
2  3  6

If you pass orient='index', the keys will be the row labels. But in this case you must also pass the desired column names:

In [54]: DataFrame.from_items([('A', [1, 2, 3]), ('B', [4, 5, 6])],
   ....:                      orient='index', columns=['one', 'two', 'three'])
   ....: 
Out[54]: 
   one  two  three
A    1    2      3
B    4    5      6

Column selection, addition, deletion

You can treat a DataFrame semantically like a dict of like-indexed Series objects. Getting, setting, and deleting columns works with the same syntax as the analogous dict operations:

In [55]: df['one']
Out[55]: 
a     1
b     2
c     3
d   NaN
Name: one, dtype: float64

In [56]: df['three'] = df['one'] * df['two']

In [57]: df['flag'] = df['one'] > 2

In [58]: df
Out[58]: 
   one  two  three   flag
a    1    1      1  False
b    2    2      4  False
c    3    3      9   True
d  NaN    4    NaN  False

Columns can be deleted or popped like with a dict:

In [59]: del df['two']

In [60]: three = df.pop('three')

In [61]: df
Out[61]: 
   one   flag
a    1  False
b    2  False
c    3   True
d  NaN  False

When inserting a scalar value, it will naturally be propagated to fill the column:

In [62]: df['foo'] = 'bar'

In [63]: df
Out[63]: 
   one   flag  foo
a    1  False  bar
b    2  False  bar
c    3   True  bar
d  NaN  False  bar

When inserting a Series that does not have the same index as the DataFrame, it will be conformed to the DataFrame’s index:

In [64]: df['one_trunc'] = df['one'][:2]

In [65]: df
Out[65]: 
   one   flag  foo  one_trunc
a    1  False  bar          1
b    2  False  bar          2
c    3   True  bar        NaN
d  NaN  False  bar        NaN

You can insert raw ndarrays but their length must match the length of the DataFrame’s index.

By default, columns get inserted at the end. The insert function is available to insert at a particular location in the columns:

In [66]: df.insert(1, 'bar', df['one'])

In [67]: df
Out[67]: 
   one  bar   flag  foo  one_trunc
a    1    1  False  bar          1
b    2    2  False  bar          2
c    3    3   True  bar        NaN
d  NaN  NaN  False  bar        NaN

Assigning New Columns in Method Chains

New in version 0.16.0.

Inspired by dplyr’s mutate verb, DataFrame has an assign() method that allows you to easily create new columns that are potentially derived from existing columns.

In [68]: iris = read_csv('data/iris.data')

In [69]: iris.head()
Out[69]: 
   SepalLength  SepalWidth  PetalLength  PetalWidth         Name
0          5.1         3.5          1.4         0.2  Iris-setosa
1          4.9         3.0          1.4         0.2  Iris-setosa
2          4.7         3.2          1.3         0.2  Iris-setosa
3          4.6         3.1          1.5         0.2  Iris-setosa
4          5.0         3.6          1.4         0.2  Iris-setosa

In [70]: (iris.assign(sepal_ratio = iris['SepalWidth'] / iris['SepalLength'])
   ....:      .head())
   ....: 
Out[70]: 
   SepalLength  SepalWidth  PetalLength  PetalWidth         Name  sepal_ratio
0          5.1         3.5          1.4         0.2  Iris-setosa        0.686
1          4.9         3.0          1.4         0.2  Iris-setosa        0.612
2          4.7         3.2          1.3         0.2  Iris-setosa        0.681
3          4.6         3.1          1.5         0.2  Iris-setosa        0.674
4          5.0         3.6          1.4         0.2  Iris-setosa        0.720

Above was an example of inserting a precomputed value. We can also pass in a function of one argument to be evalutated on the DataFrame being assigned to.

In [71]: iris.assign(sepal_ratio = lambda x: (x['SepalWidth'] /
   ....:                                      x['SepalLength'])).head()
   ....: 
Out[71]: 
   SepalLength  SepalWidth  PetalLength  PetalWidth         Name  sepal_ratio
0          5.1         3.5          1.4         0.2  Iris-setosa        0.686
1          4.9         3.0          1.4         0.2  Iris-setosa        0.612
2          4.7         3.2          1.3         0.2  Iris-setosa        0.681
3          4.6         3.1          1.5         0.2  Iris-setosa        0.674
4          5.0         3.6          1.4         0.2  Iris-setosa        0.720

assign always returns a copy of the data, leaving the original DataFrame untouched.

Passing a callable, as opposed to an actual value to be inserted, is useful when you don’t have a reference to the DataFrame at hand. This is common when using assign in chains of operations. For example, we can limit the DataFrame to just those observations with a Sepal Length greater than 5, calculate the ratio, and plot:

In [72]: (iris.query('SepalLength > 5')
   ....:      .assign(SepalRatio = lambda x: x.SepalWidth / x.SepalLength,
   ....:              PetalRatio = lambda x: x.PetalWidth / x.PetalLength)
   ....:      .plot(kind='scatter', x='SepalRatio', y='PetalRatio'))
   ....: 
Out[72]: <matplotlib.axes._subplots.AxesSubplot at 0xad87126c>
_images/basics_assign.png

Since a function is passed in, the function is computed on the DataFrame being assigned to. Importantly, this is the DataFrame that’s been filtered to those rows with sepal length greater than 5. The filtering happens first, and then the ratio calculations. This is an example where we didn’t have a reference to the filtered DataFrame available.

The function signature for assign is simply **kwargs. The keys are the column names for the new fields, and the values are either a value to be inserted (for example, a Series or NumPy array), or a function of one argument to be called on the DataFrame. A copy of the original DataFrame is returned, with the new values inserted.

Warning

Since the function signature of assign is **kwargs, a dictionary, the order of the new columns in the resulting DataFrame cannot be guaranteed.

All expressions are computed first, and then assigned. So you can’t refer to another column being assigned in the same call to assign. For example:

In [73]: # Don't do this, bad reference to `C`
        df.assign(C = lambda x: x['A'] + x['B'],
                  D = lambda x: x['A'] + x['C'])
In [2]: # Instead, break it into two assigns
        (df.assign(C = lambda x: x['A'] + x['B'])
           .assign(D = lambda x: x['A'] + x['C']))

Indexing / Selection

The basics of indexing are as follows:

Operation Syntax Result
Select column df[col] Series
Select row by label df.loc[label] Series
Select row by integer location df.iloc[loc] Series
Slice rows df[5:10] DataFrame
Select rows by boolean vector df[bool_vec] DataFrame

Row selection, for example, returns a Series whose index is the columns of the DataFrame:

In [74]: df.loc['b']
Out[74]: 
one              2
bar              2
flag         False
foo            bar
one_trunc        2
Name: b, dtype: object

In [75]: df.iloc[2]
Out[75]: 
one             3
bar             3
flag         True
foo           bar
one_trunc     NaN
Name: c, dtype: object

For a more exhaustive treatment of more sophisticated label-based indexing and slicing, see the section on indexing. We will address the fundamentals of reindexing / conforming to new sets of labels in the section on reindexing.

Data alignment and arithmetic

Data alignment between DataFrame objects automatically align on both the columns and the index (row labels). Again, the resulting object will have the union of the column and row labels.

In [76]: df = DataFrame(randn(10, 4), columns=['A', 'B', 'C', 'D'])

In [77]: df2 = DataFrame(randn(7, 3), columns=['A', 'B', 'C'])

In [78]: df + df2
Out[78]: 
       A      B      C   D
0 -1.916 -0.986 -2.421 NaN
1  0.965  1.677  0.330 NaN
2 -1.662  2.197 -1.917 NaN
3 -0.189  0.765 -0.001 NaN
4 -1.076  0.397 -1.177 NaN
5  2.810 -0.179 -0.570 NaN
6 -1.227  0.196  0.531 NaN
7    NaN    NaN    NaN NaN
8    NaN    NaN    NaN NaN
9    NaN    NaN    NaN NaN

When doing an operation between DataFrame and Series, the default behavior is to align the Series index on the DataFrame columns, thus broadcasting row-wise. For example:

In [79]: df - df.iloc[0]
Out[79]: 
       A      B      C      D
0  0.000  0.000  0.000  0.000
1  2.386  1.358  1.223 -2.107
2  2.105  1.700  1.327 -0.689
3  1.874  2.718  2.382 -0.760
4  2.199  0.966  0.826  0.093
5  4.997  1.197  1.330 -0.285
6  1.263  0.578  1.071 -0.525
7  3.463  0.632  1.063 -0.443
8  2.680  3.163  1.298 -1.818
9  1.304  0.196  3.590 -0.867

In the special case of working with time series data, if the Series is a TimeSeries (which it will be automatically if the index contains datetime objects), and the DataFrame index also contains dates, the broadcasting will be column-wise:

In [80]: index = date_range('1/1/2000', periods=8)

In [81]: df = DataFrame(randn(8, 3), index=index, columns=list('ABC'))

In [82]: df
Out[82]: 
                A      B      C
2000-01-01  0.063 -0.028  0.444
2000-01-02 -0.269 -1.578  1.850
2000-01-03  0.638 -0.557 -0.071
2000-01-04 -0.511  0.156 -1.076
2000-01-05  1.664 -0.438 -0.077
2000-01-06  0.029  0.179  1.740
2000-01-07 -0.729 -0.898 -0.314
2000-01-08 -0.048 -0.876  0.169

In [83]: type(df['A'])
Out[83]: pandas.core.series.Series

In [84]: df - df['A']
Out[84]: 
            A      B      C
2000-01-01  0 -0.091  0.381
2000-01-02  0 -1.309  2.119
2000-01-03  0 -1.195 -0.709
2000-01-04  0  0.668 -0.564
2000-01-05  0 -2.101 -1.741
2000-01-06  0  0.150  1.711
2000-01-07  0 -0.169  0.415
2000-01-08  0 -0.828  0.217

Warning

df - df['A']

is now deprecated and will be removed in a future release. The preferred way to replicate this behavior is

df.sub(df['A'], axis=0)

For explicit control over the matching and broadcasting behavior, see the section on flexible binary operations.

Operations with scalars are just as you would expect:

In [85]: df * 5 + 2
Out[85]: 
                 A      B       C
2000-01-01   2.314  1.858   4.218
2000-01-02   0.656 -5.888  11.251
2000-01-03   5.190 -0.783   1.644
2000-01-04  -0.557  2.781  -3.378
2000-01-05  10.318 -0.189   1.613
2000-01-06   2.146  2.895  10.700
2000-01-07  -1.645 -2.490   0.429
2000-01-08   1.760 -2.378   2.846

In [86]: 1 / df
Out[86]: 
                 A       B       C
2000-01-01  15.948 -35.193   2.255
2000-01-02  -3.721  -0.634   0.540
2000-01-03   1.567  -1.797 -14.039
2000-01-04  -1.955   6.398  -0.930
2000-01-05   0.601  -2.285 -12.936
2000-01-06  34.257   5.586   0.575
2000-01-07  -1.372  -1.114  -3.183
2000-01-08 -20.802  -1.142   5.913

In [87]: df ** 4
Out[87]: 
                    A          B          C
2000-01-01  1.546e-05  6.519e-07  3.871e-02
2000-01-02  5.219e-03  6.195e+00  1.172e+01
2000-01-03  1.657e-01  9.598e-02  2.574e-05
2000-01-04  6.841e-02  5.966e-04  1.339e+00
2000-01-05  7.660e+00  3.671e-02  3.571e-05
2000-01-06  7.261e-07  1.027e-03  9.168e+00
2000-01-07  2.825e-01  6.503e-01  9.747e-03
2000-01-08  5.341e-06  5.878e-01  8.178e-04

Boolean operators work as well:

In [88]: df1 = DataFrame({'a' : [1, 0, 1], 'b' : [0, 1, 1] }, dtype=bool)

In [89]: df2 = DataFrame({'a' : [0, 1, 1], 'b' : [1, 1, 0] }, dtype=bool)

In [90]: df1 & df2
Out[90]: 
       a      b
0  False  False
1  False   True
2   True  False

In [91]: df1 | df2
Out[91]: 
      a     b
0  True  True
1  True  True
2  True  True

In [92]: df1 ^ df2
Out[92]: 
       a      b
0   True   True
1   True  False
2  False   True

In [93]: -df1
Out[93]: 
       a      b
0  False   True
1   True  False
2  False  False

Transposing

To transpose, access the T attribute (also the transpose function), similar to an ndarray:

# only show the first 5 rows
In [94]: df[:5].T
Out[94]: 
   2000-01-01  2000-01-02  2000-01-03  2000-01-04  2000-01-05
A       0.063      -0.269       0.638      -0.511       1.664
B      -0.028      -1.578      -0.557       0.156      -0.438
C       0.444       1.850      -0.071      -1.076      -0.077

DataFrame interoperability with NumPy functions

Elementwise NumPy ufuncs (log, exp, sqrt, ...) and various other NumPy functions can be used with no issues on DataFrame, assuming the data within are numeric:

In [95]: np.exp(df)
Out[95]: 
                A      B      C
2000-01-01  1.065  0.972  1.558
2000-01-02  0.764  0.206  6.361
2000-01-03  1.893  0.573  0.931
2000-01-04  0.600  1.169  0.341
2000-01-05  5.278  0.646  0.926
2000-01-06  1.030  1.196  5.698
2000-01-07  0.482  0.407  0.730
2000-01-08  0.953  0.417  1.184

In [96]: np.asarray(df)
Out[96]: 
array([[ 0.0627, -0.0284,  0.4436],
       [-0.2688, -1.5776,  1.8502],
       [ 0.6381, -0.5566, -0.0712],
       [-0.5114,  0.1563, -1.0756],
       [ 1.6636, -0.4377, -0.0773],
       [ 0.0292,  0.179 ,  1.7401],
       [-0.729 , -0.898 , -0.3142],
       [-0.0481, -0.8756,  0.1691]])

The dot method on DataFrame implements matrix multiplication:

In [97]: df.T.dot(df)
Out[97]: 
       A      B      C
A  4.047 -0.039  0.178
B -0.039  4.621 -2.581
C  0.178 -2.581  7.943

Similarly, the dot method on Series implements dot product:

In [98]: s1 = Series(np.arange(5,10))

In [99]: s1.dot(s1)
Out[99]: 255

DataFrame is not intended to be a drop-in replacement for ndarray as its indexing semantics are quite different in places from a matrix.

Console display

Very large DataFrames will be truncated to display them in the console. You can also get a summary using info(). (Here I am reading a CSV version of the baseball dataset from the plyr R package):

In [100]: baseball = read_csv('data/baseball.csv')

In [101]: print(baseball)
       id     player  year  stint  ...  hbp sh  sf  gidp
0   88641  womacto01  2006      2  ...    0  3   0     0
1   88643  schilcu01  2006      1  ...    0  0   0     0
..    ...        ...   ...    ...  ...   .. ..  ..   ...
98  89533   aloumo01  2007      1  ...    2  0   3    13
99  89534  alomasa02  2007      1  ...    0  0   0     0

[100 rows x 23 columns]

In [102]: baseball.info()
<class 'pandas.core.frame.DataFrame'>
Int64Index: 100 entries, 0 to 99
Data columns (total 23 columns):
id        100 non-null int64
player    100 non-null object
year      100 non-null int64
stint     100 non-null int64
team      100 non-null object
lg        100 non-null object
g         100 non-null int64
ab        100 non-null int64
r         100 non-null int64
h         100 non-null int64
X2b       100 non-null int64
X3b       100 non-null int64
hr        100 non-null int64
rbi       100 non-null float64
sb        100 non-null float64
cs        100 non-null float64
bb        100 non-null int64
so        100 non-null float64
ibb       100 non-null float64
hbp       100 non-null float64
sh        100 non-null float64
sf        100 non-null float64
gidp      100 non-null float64
dtypes: float64(9), int64(11), object(3)
memory usage: 17.6+ KB

However, using to_string will return a string representation of the DataFrame in tabular form, though it won’t always fit the console width:

In [103]: print(baseball.iloc[-20:, :12].to_string())
       id     player  year  stint team  lg    g   ab   r    h  X2b  X3b
80  89474  finlest01  2007      1  COL  NL   43   94   9   17    3    0
81  89480  embreal01  2007      1  OAK  AL    4    0   0    0    0    0
82  89481  edmonji01  2007      1  SLN  NL  117  365  39   92   15    2
83  89482  easleda01  2007      1  NYN  NL   76  193  24   54    6    0
84  89489  delgaca01  2007      1  NYN  NL  139  538  71  139   30    0
85  89493  cormirh01  2007      1  CIN  NL    6    0   0    0    0    0
86  89494  coninje01  2007      2  NYN  NL   21   41   2    8    2    0
87  89495  coninje01  2007      1  CIN  NL   80  215  23   57   11    1
88  89497  clemero02  2007      1  NYA  AL    2    2   0    1    0    0
89  89498  claytro01  2007      2  BOS  AL    8    6   1    0    0    0
90  89499  claytro01  2007      1  TOR  AL   69  189  23   48   14    0
91  89501  cirilje01  2007      2  ARI  NL   28   40   6    8    4    0
92  89502  cirilje01  2007      1  MIN  AL   50  153  18   40    9    2
93  89521  bondsba01  2007      1  SFN  NL  126  340  75   94   14    0
94  89523  biggicr01  2007      1  HOU  NL  141  517  68  130   31    3
95  89525  benitar01  2007      2  FLO  NL   34    0   0    0    0    0
96  89526  benitar01  2007      1  SFN  NL   19    0   0    0    0    0
97  89530  ausmubr01  2007      1  HOU  NL  117  349  38   82   16    3
98  89533   aloumo01  2007      1  NYN  NL   87  328  51  112   19    1
99  89534  alomasa02  2007      1  NYN  NL    8   22   1    3    1    0

New since 0.10.0, wide DataFrames will now be printed across multiple rows by default:

In [104]: DataFrame(randn(3, 12))
Out[104]: 
         0         1         2         3         4         5         6   \
0  1.225021 -0.528620  0.448676  0.619107 -1.199110 -0.949097  2.169523   
1 -1.753617  0.992384 -0.505601 -0.599848  0.133585  0.008836 -1.767710   
2 -0.461585 -1.321106  1.745476  1.445100  0.991037 -0.860733 -0.870661   

         7         8         9         10        11  
0  0.302230  0.919516  0.657436  0.262574 -0.804798  
1  0.700112 -0.020773 -0.302481  0.347869  0.179123  
2 -0.117845 -0.046266  2.095649 -0.524324 -0.610555  

You can change how much to print on a single row by setting the display.width option:

In [105]: set_option('display.width', 40) # default is 80

In [106]: DataFrame(randn(3, 12))
Out[106]: 
         0         1         2   \
0 -1.280951  1.472585 -1.001914   
1  0.130529 -1.603771 -0.128830   
2 -1.084566 -0.515272  1.367586   

         3         4         5   \
0  1.044770 -0.050668 -0.013289   
1 -1.869301 -0.232977 -0.139801   
2  0.963500  0.224105 -0.020051   

         6         7         8   \
0 -0.291893  2.029038 -1.117195   
1 -1.083341 -0.357234 -0.818199   
2  0.524663  0.351081 -1.574209   

         9         10        11  
0  1.598577 -0.397325  0.151653  
1 -0.886885  1.238885 -1.639274  
2 -0.486856 -0.545888 -0.927076  

You can also disable this feature via the expand_frame_repr option. This will print the table in one block.

DataFrame column attribute access and IPython completion

If a DataFrame column label is a valid Python variable name, the column can be accessed like attributes:

In [107]: df = DataFrame({'foo1' : np.random.randn(5),
   .....:                 'foo2' : np.random.randn(5)})
   .....: 

In [108]: df
Out[108]: 
       foo1      foo2
0  0.909160  1.360298
1 -0.667763 -1.603624
2 -0.101656 -1.648929
3  1.189682  0.145121
4 -0.090648 -2.536359

In [109]: df.foo1
Out[109]: 
0    0.909160
1   -0.667763
2   -0.101656
3    1.189682
4   -0.090648
Name: foo1, dtype: float64

The columns are also connected to the IPython completion mechanism so they can be tab-completed:

In [5]: df.fo<TAB>
df.foo1  df.foo2

Panel

Panel is a somewhat less-used, but still important container for 3-dimensional data. The term panel data is derived from econometrics and is partially responsible for the name pandas: pan(el)-da(ta)-s. The names for the 3 axes are intended to give some semantic meaning to describing operations involving panel data and, in particular, econometric analysis of panel data. However, for the strict purposes of slicing and dicing a collection of DataFrame objects, you may find the axis names slightly arbitrary:

  • items: axis 0, each item corresponds to a DataFrame contained inside
  • major_axis: axis 1, it is the index (rows) of each of the DataFrames
  • minor_axis: axis 2, it is the columns of each of the DataFrames

Construction of Panels works about like you would expect:

From 3D ndarray with optional axis labels

In [110]: wp = Panel(randn(2, 5, 4), items=['Item1', 'Item2'],
   .....:            major_axis=date_range('1/1/2000', periods=5),
   .....:            minor_axis=['A', 'B', 'C', 'D'])
   .....: 

In [111]: wp
Out[111]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 5 (major_axis) x 4 (minor_axis)
Items axis: Item1 to Item2
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00
Minor_axis axis: A to D

From dict of DataFrame objects

In [112]: data = {'Item1' : DataFrame(randn(4, 3)),
   .....:         'Item2' : DataFrame(randn(4, 2))}
   .....: 

In [113]: Panel(data)
Out[113]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 4 (major_axis) x 3 (minor_axis)
Items axis: Item1 to Item2
Major_axis axis: 0 to 3
Minor_axis axis: 0 to 2

Note that the values in the dict need only be convertible to DataFrame. Thus, they can be any of the other valid inputs to DataFrame as per above.

One helpful factory method is Panel.from_dict, which takes a dictionary of DataFrames as above, and the following named parameters:

Parameter Default Description
intersect False drops elements whose indices do not align
orient items use minor to use DataFrames’ columns as panel items

For example, compare to the construction above:

In [114]: Panel.from_dict(data, orient='minor')
Out[114]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 3 (items) x 4 (major_axis) x 2 (minor_axis)
Items axis: 0 to 2
Major_axis axis: 0 to 3
Minor_axis axis: Item1 to Item2

Orient is especially useful for mixed-type DataFrames. If you pass a dict of DataFrame objects with mixed-type columns, all of the data will get upcasted to dtype=object unless you pass orient='minor':

In [115]: df = DataFrame({'a': ['foo', 'bar', 'baz'],
   .....:                 'b': np.random.randn(3)})
   .....: 

In [116]: df
Out[116]: 
     a         b
0  foo -1.264356
1  bar -0.497629
2  baz  1.789719

In [117]: data = {'item1': df, 'item2': df}

In [118]: panel = Panel.from_dict(data, orient='minor')

In [119]: panel['a']
Out[119]: 
  item1 item2
0   foo   foo
1   bar   bar
2   baz   baz

In [120]: panel['b']
Out[120]: 
      item1     item2
0 -1.264356 -1.264356
1 -0.497629 -0.497629
2  1.789719  1.789719

In [121]: panel['b'].dtypes
Out[121]: 
item1    float64
item2    float64
dtype: object

Note

Unfortunately Panel, being less commonly used than Series and DataFrame, has been slightly neglected feature-wise. A number of methods and options available in DataFrame are not available in Panel. This will get worked on, of course, in future releases. And faster if you join me in working on the codebase.

From DataFrame using to_panel method

This method was introduced in v0.7 to replace LongPanel.to_long, and converts a DataFrame with a two-level index to a Panel.

In [122]: midx = MultiIndex(levels=[['one', 'two'], ['x','y']], labels=[[1,1,0,0],[1,0,1,0]])

In [123]: df = DataFrame({'A' : [1, 2, 3, 4], 'B': [5, 6, 7, 8]}, index=midx)

In [124]: df.to_panel()
Out[124]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 2 (major_axis) x 2 (minor_axis)
Items axis: A to B
Major_axis axis: one to two
Minor_axis axis: x to y

Item selection / addition / deletion

Similar to DataFrame functioning as a dict of Series, Panel is like a dict of DataFrames:

In [125]: wp['Item1']
Out[125]: 
                   A         B         C         D
2000-01-01  0.835993 -0.621868 -0.173710 -0.174326
2000-01-02 -0.354356  2.090183 -0.736019 -1.250412
2000-01-03 -0.581326 -0.244477  0.917119  0.611695
2000-01-04 -1.576078 -0.528562 -0.704643 -0.481453
2000-01-05  1.085093 -1.229749  2.295679 -1.016910

In [126]: wp['Item3'] = wp['Item1'] / wp['Item2']

The API for insertion and deletion is the same as for DataFrame. And as with DataFrame, if the item is a valid python identifier, you can access it as an attribute and tab-complete it in IPython.

Transposing

A Panel can be rearranged using its transpose method (which does not make a copy by default unless the data are heterogeneous):

In [127]: wp.transpose(2, 0, 1)
Out[127]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 4 (items) x 3 (major_axis) x 5 (minor_axis)
Items axis: A to D
Major_axis axis: Item1 to Item3
Minor_axis axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00

Indexing / Selection

Operation Syntax Result
Select item wp[item] DataFrame
Get slice at major_axis label wp.major_xs(val) DataFrame
Get slice at minor_axis label wp.minor_xs(val) DataFrame

For example, using the earlier example data, we could do:

In [128]: wp['Item1']
Out[128]: 
                   A         B         C         D
2000-01-01  0.835993 -0.621868 -0.173710 -0.174326
2000-01-02 -0.354356  2.090183 -0.736019 -1.250412
2000-01-03 -0.581326 -0.244477  0.917119  0.611695
2000-01-04 -1.576078 -0.528562 -0.704643 -0.481453
2000-01-05  1.085093 -1.229749  2.295679 -1.016910

In [129]: wp.major_xs(wp.major_axis[2])
Out[129]: 
      Item1     Item2     Item3
A -0.581326 -1.271582  0.457167
B -0.244477 -0.861256  0.283861
C  0.917119 -0.597879 -1.533955
D  0.611695 -0.118700 -5.153265

In [130]: wp.minor_axis
Out[130]: Index([u'A', u'B', u'C', u'D'], dtype='object')

In [131]: wp.minor_xs('C')
Out[131]: 
               Item1     Item2      Item3
2000-01-01 -0.173710  2.381645  -0.072937
2000-01-02 -0.736019 -2.413161   0.305002
2000-01-03  0.917119 -0.597879  -1.533955
2000-01-04 -0.704643 -1.536019   0.458746
2000-01-05  2.295679  0.181524  12.646732

Squeezing

Another way to change the dimensionality of an object is to squeeze a 1-len object, similar to wp['Item1']

In [132]: wp.reindex(items=['Item1']).squeeze()
Out[132]: 
                   A         B         C         D
2000-01-01  0.835993 -0.621868 -0.173710 -0.174326
2000-01-02 -0.354356  2.090183 -0.736019 -1.250412
2000-01-03 -0.581326 -0.244477  0.917119  0.611695
2000-01-04 -1.576078 -0.528562 -0.704643 -0.481453
2000-01-05  1.085093 -1.229749  2.295679 -1.016910

In [133]: wp.reindex(items=['Item1'],minor=['B']).squeeze()
Out[133]: 
2000-01-01   -0.621868
2000-01-02    2.090183
2000-01-03   -0.244477
2000-01-04   -0.528562
2000-01-05   -1.229749
Freq: D, Name: B, dtype: float64

Conversion to DataFrame

A Panel can be represented in 2D form as a hierarchically indexed DataFrame. See the section hierarchical indexing for more on this. To convert a Panel to a DataFrame, use the to_frame method:

In [134]: panel = Panel(np.random.randn(3, 5, 4), items=['one', 'two', 'three'],
   .....:               major_axis=date_range('1/1/2000', periods=5),
   .....:               minor_axis=['a', 'b', 'c', 'd'])
   .....: 

In [135]: panel.to_frame()
Out[135]: 
                       one       two     three
major      minor                              
2000-01-01 a      0.445900 -1.286198 -1.023189
           b     -0.574496 -0.407154  0.591682
           c      0.872979  0.068084 -0.008919
           d      0.297255 -2.157051 -0.415572
2000-01-02 a     -1.022617 -0.443982 -0.772683
           b      1.091870 -0.881639 -0.516197
           c      1.831444  0.851834  0.626655
           d      1.271808 -1.352515  0.269623
2000-01-03 a     -0.472876  0.228761  1.709250
           b     -0.279340  0.416858 -0.830728
           c      0.495966  0.301709 -0.290244
           d      0.367858  0.569010 -1.588782
2000-01-04 a     -1.530917 -0.047619  0.639406
           b     -0.285890  0.413370  1.055533
           c      0.943062  0.573056 -0.260898
           d      1.361752 -0.154419 -0.289725
2000-01-05 a      0.210373  0.987044  0.279621
           b     -1.945608  0.063191  0.454423
           c      2.532409  0.439086 -0.065750
           d      0.373819  1.657475  1.465709

Panel4D (Experimental)

Panel4D is a 4-Dimensional named container very much like a Panel, but having 4 named dimensions. It is intended as a test bed for more N-Dimensional named containers.

  • labels: axis 0, each item corresponds to a Panel contained inside
  • items: axis 1, each item corresponds to a DataFrame contained inside
  • major_axis: axis 2, it is the index (rows) of each of the DataFrames
  • minor_axis: axis 3, it is the columns of each of the DataFrames

Panel4D is a sub-class of Panel, so most methods that work on Panels are applicable to Panel4D. The following methods are disabled:

  • join , to_frame , to_excel , to_sparse , groupby

Construction of Panel4D works in a very similar manner to a Panel

From 4D ndarray with optional axis labels

In [136]: p4d = Panel4D(randn(2, 2, 5, 4),
   .....:            labels=['Label1','Label2'],
   .....:            items=['Item1', 'Item2'],
   .....:            major_axis=date_range('1/1/2000', periods=5),
   .....:            minor_axis=['A', 'B', 'C', 'D'])
   .....: 

In [137]: p4d
Out[137]: 
<class 'pandas.core.panelnd.Panel4D'>
Dimensions: 2 (labels) x 2 (items) x 5 (major_axis) x 4 (minor_axis)
Labels axis: Label1 to Label2
Items axis: Item1 to Item2
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00
Minor_axis axis: A to D

From dict of Panel objects

In [138]: data = { 'Label1' : Panel({ 'Item1' : DataFrame(randn(4, 3)) }),
   .....:          'Label2' : Panel({ 'Item2' : DataFrame(randn(4, 2)) }) }
   .....: 

In [139]: Panel4D(data)
Out[139]: 
<class 'pandas.core.panelnd.Panel4D'>
Dimensions: 2 (labels) x 2 (items) x 4 (major_axis) x 3 (minor_axis)
Labels axis: Label1 to Label2
Items axis: Item1 to Item2
Major_axis axis: 0 to 3
Minor_axis axis: 0 to 2

Note that the values in the dict need only be convertible to Panels. Thus, they can be any of the other valid inputs to Panel as per above.

Slicing

Slicing works in a similar manner to a Panel. [] slices the first dimension. .ix allows you to slice arbitrarily and get back lower dimensional objects

In [140]: p4d['Label1']
Out[140]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 5 (major_axis) x 4 (minor_axis)
Items axis: Item1 to Item2
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00
Minor_axis axis: A to D

4D -> Panel

In [141]: p4d.ix[:,:,:,'A']
Out[141]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 2 (major_axis) x 5 (minor_axis)
Items axis: Label1 to Label2
Major_axis axis: Item1 to Item2
Minor_axis axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00

4D -> DataFrame

In [142]: p4d.ix[:,:,0,'A']
Out[142]: 
         Label1    Label2
Item1  1.127489  0.015494
Item2 -1.650400  0.130533

4D -> Series

In [143]: p4d.ix[:,0,0,'A']
Out[143]: 
Label1    1.127489
Label2    0.015494
Name: A, dtype: float64

Transposing

A Panel4D can be rearranged using its transpose method (which does not make a copy by default unless the data are heterogeneous):

In [144]: p4d.transpose(3, 2, 1, 0)
Out[144]: 
<class 'pandas.core.panelnd.Panel4D'>
Dimensions: 4 (labels) x 5 (items) x 2 (major_axis) x 2 (minor_axis)
Labels axis: A to D
Items axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00
Major_axis axis: Item1 to Item2
Minor_axis axis: Label1 to Label2

PanelND (Experimental)

PanelND is a module with a set of factory functions to enable a user to construct N-dimensional named containers like Panel4D, with a custom set of axis labels. Thus a domain-specific container can easily be created.

The following creates a Panel5D. A new panel type object must be sliceable into a lower dimensional object. Here we slice to a Panel4D.

In [145]: from pandas.core import panelnd

In [146]: Panel5D = panelnd.create_nd_panel_factory(
   .....:     klass_name   = 'Panel5D',
   .....:     orders  = [ 'cool', 'labels','items','major_axis','minor_axis'],
   .....:     slices  = { 'labels' : 'labels', 'items' : 'items',
   .....:                     'major_axis' : 'major_axis', 'minor_axis' : 'minor_axis' },
   .....:     slicer  = Panel4D,
   .....:     aliases = { 'major' : 'major_axis', 'minor' : 'minor_axis' },
   .....:     stat_axis    = 2)
   .....: 

In [147]: p5d = Panel5D(dict(C1 = p4d))

In [148]: p5d
Out[148]: 
<class 'pandas.core.panelnd.Panel5D'>
Dimensions: 1 (cool) x 2 (labels) x 2 (items) x 5 (major_axis) x 4 (minor_axis)
Cool axis: C1 to C1
Labels axis: Label1 to Label2
Items axis: Item1 to Item2
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00
Minor_axis axis: A to D

# print a slice of our 5D
In [149]: p5d.ix['C1',:,:,0:3,:]
Out[149]: 
<class 'pandas.core.panelnd.Panel4D'>
Dimensions: 2 (labels) x 2 (items) x 3 (major_axis) x 4 (minor_axis)
Labels axis: Label1 to Label2
Items axis: Item1 to Item2
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-03 00:00:00
Minor_axis axis: A to D

# transpose it
In [150]: p5d.transpose(1,2,3,4,0)
Out[150]: 
<class 'pandas.core.panelnd.Panel5D'>
Dimensions: 2 (cool) x 2 (labels) x 5 (items) x 4 (major_axis) x 1 (minor_axis)
Cool axis: Label1 to Label2
Labels axis: Item1 to Item2
Items axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00
Major_axis axis: A to D
Minor_axis axis: C1 to C1

# look at the shape & dim
In [151]: p5d.shape
Out[151]: (1, 2, 2, 5, 4)

In [152]: p5d.ndim
Out[152]: 5