Table Of Contents

Search

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

Essential Basic Functionality

Here we discuss a lot of the essential functionality common to the pandas data structures. Here’s how to create some of the objects used in the examples from the previous section:

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

In [2]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [3]: df = pd.DataFrame(np.random.randn(8, 3), index=index,
   ...:                   columns=['A', 'B', 'C'])
   ...: 

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

Head and Tail

To view a small sample of a Series or DataFrame object, use the head() and tail() methods. The default number of elements to display is five, but you may pass a custom number.

In [5]: long_series = pd.Series(np.random.randn(1000))

In [6]: long_series.head()
Out[6]: 
0   -0.305384
1   -0.479195
2    0.095031
3   -0.270099
4   -0.707140
dtype: float64

In [7]: long_series.tail(3)
Out[7]: 
997    0.588446
998    0.026465
999   -1.728222
dtype: float64

Attributes and the raw ndarray(s)

pandas objects have a number of attributes enabling you to access the metadata

  • shape: gives the axis dimensions of the object, consistent with ndarray
  • Axis labels
    • Series: index (only axis)
    • DataFrame: index (rows) and columns
    • Panel: items, major_axis, and minor_axis

Note, these attributes can be safely assigned to!

In [8]: df[:2]
Out[8]: 
                   A         B         C
2000-01-01  0.187483 -1.933946  0.377312
2000-01-02  0.734122  2.141616 -0.011225

In [9]: df.columns = [x.lower() for x in df.columns]

In [10]: df
Out[10]: 
                   a         b         c
2000-01-01  0.187483 -1.933946  0.377312
2000-01-02  0.734122  2.141616 -0.011225
2000-01-03  0.048869 -1.360687 -0.479010
2000-01-04 -0.859661 -0.231595 -0.527750
2000-01-05 -1.296337  0.150680  0.123836
2000-01-06  0.571764  1.555563 -0.823761
2000-01-07  0.535420 -1.032853  1.469725
2000-01-08  1.304124  1.449735  0.203109

To get the actual data inside a data structure, one need only access the values property:

In [11]: s.values
Out[11]: array([ 0.1122,  0.8717, -0.8161, -0.7849,  1.0307])

In [12]: df.values
Out[12]: 
array([[ 0.1875, -1.9339,  0.3773],
       [ 0.7341,  2.1416, -0.0112],
       [ 0.0489, -1.3607, -0.479 ],
       [-0.8597, -0.2316, -0.5278],
       [-1.2963,  0.1507,  0.1238],
       [ 0.5718,  1.5556, -0.8238],
       [ 0.5354, -1.0329,  1.4697],
       [ 1.3041,  1.4497,  0.2031]])

In [13]: wp.values
Out[13]: 
array([[[-1.032 ,  0.9698, -0.9627,  1.3821],
        [-0.9388,  0.6691, -0.4336, -0.2736],
        [ 0.6804, -0.3084, -0.2761, -1.8212],
        [-1.9936, -1.9274, -2.0279,  1.625 ],
        [ 0.5511,  3.0593,  0.4553, -0.0307]],

       [[ 0.9357,  1.0612, -2.1079,  0.1999],
        [ 0.3236, -0.6416, -0.5875,  0.0539],
        [ 0.1949, -0.382 ,  0.3186,  2.0891],
        [-0.7283, -0.0903, -0.7482,  1.3189],
        [-2.0298,  0.7927,  0.461 , -0.5427]]])

If a DataFrame or Panel contains homogeneously-typed data, the ndarray can actually be modified in-place, and the changes will be reflected in the data structure. For heterogeneous data (e.g. some of the DataFrame’s columns are not all the same dtype), this will not be the case. The values attribute itself, unlike the axis labels, cannot be assigned to.

Note

When working with heterogeneous data, the dtype of the resulting ndarray will be chosen to accommodate all of the data involved. For example, if strings are involved, the result will be of object dtype. If there are only floats and integers, the resulting array will be of float dtype.

Accelerated operations

pandas has support for accelerating certain types of binary numerical and boolean operations using the numexpr library (starting in 0.11.0) and the bottleneck libraries.

These libraries are especially useful when dealing with large data sets, and provide large speedups. numexpr uses smart chunking, caching, and multiple cores. bottleneck is a set of specialized cython routines that are especially fast when dealing with arrays that have nans.

Here is a sample (using 100 column x 100,000 row DataFrames):

Operation 0.11.0 (ms) Prior Version (ms) Ratio to Prior
df1 > df2 13.32 125.35 0.1063
df1 * df2 21.71 36.63 0.5928
df1 + df2 22.04 36.50 0.6039

You are highly encouraged to install both libraries. See the section Recommended Dependencies for more installation info.

Flexible binary operations

With binary operations between pandas data structures, there are two key points of interest:

  • Broadcasting behavior between higher- (e.g. DataFrame) and lower-dimensional (e.g. Series) objects.
  • Missing data in computations

We will demonstrate how to manage these issues independently, though they can be handled simultaneously.

Matching / broadcasting behavior

DataFrame has the methods add(), sub(), mul(), div() and related functions radd(), rsub(), ... for carrying out binary operations. For broadcasting behavior, Series input is of primary interest. Using these functions, you can use to either match on the index or columns via the axis keyword:

In [14]: df = pd.DataFrame({'one' : pd.Series(np.random.randn(3), index=['a', 'b', 'c']),
   ....:                    'two' : pd.Series(np.random.randn(4), index=['a', 'b', 'c', 'd']),
   ....:                    'three' : pd.Series(np.random.randn(3), index=['b', 'c', 'd'])})
   ....: 

In [15]: df
Out[15]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789
d       NaN  1.124472 -1.101558

In [16]: row = df.ix[1]

In [17]: column = df['two']

In [18]: df.sub(row, axis='columns')
Out[18]: 
        one     three       two
a -0.487650       NaN -1.487837
b  0.000000  0.000000  0.000000
c  0.150512  0.639504 -1.585038
d       NaN  1.301762 -2.237808

In [19]: df.sub(row, axis=1)
Out[19]: 
        one     three       two
a -0.487650       NaN -1.487837
b  0.000000  0.000000  0.000000
c  0.150512  0.639504 -1.585038
d       NaN  1.301762 -2.237808

In [20]: df.sub(column, axis='index')
Out[20]: 
        one     three  two
a -0.274957       NaN  0.0
b -1.275144 -1.313539  0.0
c  0.460406  0.911003  0.0
d       NaN  2.226031  0.0

In [21]: df.sub(column, axis=0)
Out[21]: 
        one     three  two
a -0.274957       NaN  0.0
b -1.275144 -1.313539  0.0
c  0.460406  0.911003  0.0
d       NaN  2.226031  0.0

Furthermore you can align a level of a multi-indexed DataFrame with a Series.

In [22]: dfmi = df.copy()

In [23]: dfmi.index = pd.MultiIndex.from_tuples([(1,'a'),(1,'b'),(1,'c'),(2,'a')],
   ....:                                        names=['first','second'])
   ....: 

In [24]: dfmi.sub(column, axis=0, level='second')
Out[24]: 
                   one     three       two
first second                              
1     a      -0.274957       NaN  0.000000
      b      -1.275144 -1.313539  0.000000
      c       0.460406  0.911003  0.000000
2     a            NaN  1.476060 -0.749971

With Panel, describing the matching behavior is a bit more difficult, so the arithmetic methods instead (and perhaps confusingly?) give you the option to specify the broadcast axis. For example, suppose we wished to demean the data over a particular axis. This can be accomplished by taking the mean over an axis and broadcasting over the same axis:

In [25]: major_mean = wp.mean(axis='major')

In [26]: major_mean
Out[26]: 
      Item1     Item2
A -0.546569 -0.260774
B  0.492478  0.147993
C -0.649010 -0.532794
D  0.176307  0.623812

In [27]: wp.sub(major_mean, axis='major')
Out[27]: 
<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

And similarly for axis="items" and axis="minor".

Note

I could be convinced to make the axis argument in the DataFrame methods match the broadcasting behavior of Panel. Though it would require a transition period so users can change their code...

Missing data / operations with fill values

In Series and DataFrame (though not yet in Panel), the arithmetic functions have the option of inputting a fill_value, namely a value to substitute when at most one of the values at a location are missing. For example, when adding two DataFrame objects, you may wish to treat NaN as 0 unless both DataFrames are missing that value, in which case the result will be NaN (you can later replace NaN with some other value using fillna if you wish).

In [28]: df
Out[28]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789
d       NaN  1.124472 -1.101558

In [29]: df2
Out[29]: 
        one     three       two
a -0.626544  1.000000 -0.351587
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789
d       NaN  1.124472 -1.101558

In [30]: df + df2
Out[30]: 
        one     three       two
a -1.253088       NaN -0.703174
b -0.277789 -0.354579  2.272499
c  0.023235  0.924429 -0.897577
d       NaN  2.248945 -2.203116

In [31]: df.add(df2, fill_value=0)
Out[31]: 
        one     three       two
a -1.253088  1.000000 -0.703174
b -0.277789 -0.354579  2.272499
c  0.023235  0.924429 -0.897577
d       NaN  2.248945 -2.203116

Flexible Comparisons

Starting in v0.8, pandas introduced binary comparison methods eq, ne, lt, gt, le, and ge to Series and DataFrame whose behavior is analogous to the binary arithmetic operations described above:

In [32]: df.gt(df2)
Out[32]: 
     one  three    two
a  False  False  False
b  False  False  False
c  False  False  False
d  False  False  False

In [33]: df2.ne(df)
Out[33]: 
     one  three    two
a  False   True  False
b  False  False  False
c  False  False  False
d   True  False  False

These operations produce a pandas object the same type as the left-hand-side input that if of dtype bool. These boolean objects can be used in indexing operations, see here

Boolean Reductions

You can apply the reductions: empty, any(), all(), and bool() to provide a way to summarize a boolean result.

In [34]: (df > 0).all()
Out[34]: 
one      False
three    False
two      False
dtype: bool

In [35]: (df > 0).any()
Out[35]: 
one      True
three    True
two      True
dtype: bool

You can reduce to a final boolean value.

In [36]: (df > 0).any().any()
Out[36]: True

You can test if a pandas object is empty, via the empty property.

In [37]: df.empty
Out[37]: False

In [38]: pd.DataFrame(columns=list('ABC')).empty
Out[38]: True

To evaluate single-element pandas objects in a boolean context, use the method bool():

In [39]: pd.Series([True]).bool()
Out[39]: True

In [40]: pd.Series([False]).bool()
Out[40]: False

In [41]: pd.DataFrame([[True]]).bool()
Out[41]: True

In [42]: pd.DataFrame([[False]]).bool()
Out[42]: False

Warning

You might be tempted to do the following:

>>> if df:
     ...

Or

>>> df and df2

These both will raise as you are trying to compare multiple values.

ValueError: The truth value of an array is ambiguous. Use a.empty, a.any() or a.all().

See gotchas for a more detailed discussion.

Comparing if objects are equivalent

Often you may find there is more than one way to compute the same result. As a simple example, consider df+df and df*2. To test that these two computations produce the same result, given the tools shown above, you might imagine using (df+df == df*2).all(). But in fact, this expression is False:

In [43]: df+df == df*2
Out[43]: 
     one  three   two
a   True  False  True
b   True   True  True
c   True   True  True
d  False   True  True

In [44]: (df+df == df*2).all()
Out[44]: 
one      False
three    False
two       True
dtype: bool

Notice that the boolean DataFrame df+df == df*2 contains some False values! That is because NaNs do not compare as equals:

In [45]: np.nan == np.nan
Out[45]: False

So, as of v0.13.1, NDFrames (such as Series, DataFrames, and Panels) have an equals() method for testing equality, with NaNs in corresponding locations treated as equal.

In [46]: (df+df).equals(df*2)
Out[46]: True

Note that the Series or DataFrame index needs to be in the same order for equality to be True:

In [47]: df1 = pd.DataFrame({'col':['foo', 0, np.nan]})

In [48]: df2 = pd.DataFrame({'col':[np.nan, 0, 'foo']}, index=[2,1,0])

In [49]: df1.equals(df2)
Out[49]: False

In [50]: df1.equals(df2.sort_index())
Out[50]: True

Comparing array-like objects

You can conveniently do element-wise comparisons when comparing a pandas data structure with a scalar value:

In [51]: pd.Series(['foo', 'bar', 'baz']) == 'foo'
Out[51]: 
0     True
1    False
2    False
dtype: bool

In [52]: pd.Index(['foo', 'bar', 'baz']) == 'foo'
Out[52]: array([ True, False, False], dtype=bool)

Pandas also handles element-wise comparisons between different array-like objects of the same length:

In [53]: pd.Series(['foo', 'bar', 'baz']) == pd.Index(['foo', 'bar', 'qux'])
Out[53]: 
0     True
1     True
2    False
dtype: bool

In [54]: pd.Series(['foo', 'bar', 'baz']) == np.array(['foo', 'bar', 'qux'])
Out[54]: 
0     True
1     True
2    False
dtype: bool

Trying to compare Index or Series objects of different lengths will raise a ValueError:

In [55]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo', 'bar'])
ValueError: Series lengths must match to compare

In [56]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo'])
ValueError: Series lengths must match to compare

Note that this is different from the numpy behavior where a comparison can be broadcast:

In [55]: np.array([1, 2, 3]) == np.array([2])
Out[55]: array([False,  True, False], dtype=bool)

or it can return False if broadcasting can not be done:

In [56]: np.array([1, 2, 3]) == np.array([1, 2])
Out[56]: False

Combining overlapping data sets

A problem occasionally arising is the combination of two similar data sets where values in one are preferred over the other. An example would be two data series representing a particular economic indicator where one is considered to be of “higher quality”. However, the lower quality series might extend further back in history or have more complete data coverage. As such, we would like to combine two DataFrame objects where missing values in one DataFrame are conditionally filled with like-labeled values from the other DataFrame. The function implementing this operation is combine_first(), which we illustrate:

In [57]: df1 = pd.DataFrame({'A' : [1., np.nan, 3., 5., np.nan],
   ....:                     'B' : [np.nan, 2., 3., np.nan, 6.]})
   ....: 

In [58]: df2 = pd.DataFrame({'A' : [5., 2., 4., np.nan, 3., 7.],
   ....:                     'B' : [np.nan, np.nan, 3., 4., 6., 8.]})
   ....: 

In [59]: df1
Out[59]: 
     A    B
0  1.0  NaN
1  NaN  2.0
2  3.0  3.0
3  5.0  NaN
4  NaN  6.0

In [60]: df2
Out[60]: 
     A    B
0  5.0  NaN
1  2.0  NaN
2  4.0  3.0
3  NaN  4.0
4  3.0  6.0
5  7.0  8.0

In [61]: df1.combine_first(df2)
Out[61]: 
     A    B
0  1.0  NaN
1  2.0  2.0
2  3.0  3.0
3  5.0  4.0
4  3.0  6.0
5  7.0  8.0

General DataFrame Combine

The combine_first() method above calls the more general DataFrame method combine(). This method takes another DataFrame and a combiner function, aligns the input DataFrame and then passes the combiner function pairs of Series (i.e., columns whose names are the same).

So, for instance, to reproduce combine_first() as above:

In [62]: combiner = lambda x, y: np.where(pd.isnull(x), y, x)

In [63]: df1.combine(df2, combiner)
Out[63]: 
     A    B
0  1.0  NaN
1  2.0  2.0
2  3.0  3.0
3  5.0  4.0
4  3.0  6.0
5  7.0  8.0

Descriptive statistics

A large number of methods for computing descriptive statistics and other related operations on Series, DataFrame, and Panel. Most of these are aggregations (hence producing a lower-dimensional result) like sum(), mean(), and quantile(), but some of them, like cumsum() and cumprod(), produce an object of the same size. Generally speaking, these methods take an axis argument, just like ndarray.{sum, std, ...}, but the axis can be specified by name or integer:

  • Series: no axis argument needed
  • DataFrame: “index” (axis=0, default), “columns” (axis=1)
  • Panel: “items” (axis=0), “major” (axis=1, default), “minor” (axis=2)

For example:

In [64]: df
Out[64]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789
d       NaN  1.124472 -1.101558

In [65]: df.mean(0)
Out[65]: 
one     -0.251274
three    0.469799
two     -0.191421
dtype: float64

In [66]: df.mean(1)
Out[66]: 
a   -0.489066
b    0.273355
c    0.008348
d    0.011457
dtype: float64

All such methods have a skipna option signaling whether to exclude missing data (True by default):

In [67]: df.sum(0, skipna=False)
Out[67]: 
one           NaN
three         NaN
two     -0.765684
dtype: float64

In [68]: df.sum(axis=1, skipna=True)
Out[68]: 
a   -0.978131
b    0.820066
c    0.025044
d    0.022914
dtype: float64

Combined with the broadcasting / arithmetic behavior, one can describe various statistical procedures, like standardization (rendering data zero mean and standard deviation 1), very concisely:

In [69]: ts_stand = (df - df.mean()) / df.std()

In [70]: ts_stand.std()
Out[70]: 
one      1.0
three    1.0
two      1.0
dtype: float64

In [71]: xs_stand = df.sub(df.mean(1), axis=0).div(df.std(1), axis=0)

In [72]: xs_stand.std(1)
Out[72]: 
a    1.0
b    1.0
c    1.0
d    1.0
dtype: float64

Note that methods like cumsum() and cumprod() preserve the location of NA values:

In [73]: df.cumsum()
Out[73]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.765438 -0.177289  0.784662
c -0.753821  0.284925  0.335874
d       NaN  1.409398 -0.765684

Here is a quick reference summary table of common functions. Each also takes an optional level parameter which applies only if the object has a hierarchical index.

Function Description
count Number of non-null observations
sum Sum of values
mean Mean of values
mad Mean absolute deviation
median Arithmetic median of values
min Minimum
max Maximum
mode Mode
abs Absolute Value
prod Product of values
std Bessel-corrected sample standard deviation
var Unbiased variance
sem Standard error of the mean
skew Sample skewness (3rd moment)
kurt Sample kurtosis (4th moment)
quantile Sample quantile (value at %)
cumsum Cumulative sum
cumprod Cumulative product
cummax Cumulative maximum
cummin Cumulative minimum

Note that by chance some NumPy methods, like mean, std, and sum, will exclude NAs on Series input by default:

In [74]: np.mean(df['one'])
Out[74]: -0.25127365175839511

In [75]: np.mean(df['one'].values)
Out[75]: nan

Series also has a method nunique() which will return the number of unique non-null values:

In [76]: series = pd.Series(np.random.randn(500))

In [77]: series[20:500] = np.nan

In [78]: series[10:20]  = 5

In [79]: series.nunique()
Out[79]: 11

Summarizing data: describe

There is a convenient describe() function which computes a variety of summary statistics about a Series or the columns of a DataFrame (excluding NAs of course):

In [80]: series = pd.Series(np.random.randn(1000))

In [81]: series[::2] = np.nan

In [82]: series.describe()
Out[82]: 
count    500.000000
mean      -0.039663
std        1.069371
min       -3.463789
25%             NaN
50%             NaN
75%             NaN
max        3.120271
dtype: float64

In [83]: frame = pd.DataFrame(np.random.randn(1000, 5), columns=['a', 'b', 'c', 'd', 'e'])

In [84]: frame.ix[::2] = np.nan

In [85]: frame.describe()
Out[85]: 
                a           b           c           d           e
count  500.000000  500.000000  500.000000  500.000000  500.000000
mean     0.000954   -0.044014    0.075936   -0.003679    0.020751
std      1.005133    0.974882    0.967432    1.004732    0.963812
min     -3.010899   -2.782760   -3.401252   -2.944925   -3.794127
25%           NaN         NaN         NaN         NaN         NaN
50%           NaN         NaN         NaN         NaN         NaN
75%           NaN         NaN         NaN         NaN         NaN
max      3.007143    2.627688    2.702490    2.850852    3.072117

You can select specific percentiles to include in the output:

In [86]: series.describe(percentiles=[.05, .25, .75, .95])
Out[86]: 
count    500.000000
mean      -0.039663
std        1.069371
min       -3.463789
5%              NaN
25%             NaN
50%             NaN
75%             NaN
95%             NaN
max        3.120271
dtype: float64

By default, the median is always included.

For a non-numerical Series object, describe() will give a simple summary of the number of unique values and most frequently occurring values:

In [87]: s = pd.Series(['a', 'a', 'b', 'b', 'a', 'a', np.nan, 'c', 'd', 'a'])

In [88]: s.describe()
Out[88]: 
count     9
unique    4
top       a
freq      5
dtype: object

Note that on a mixed-type DataFrame object, describe() will restrict the summary to include only numerical columns or, if none are, only categorical columns:

In [89]: frame = pd.DataFrame({'a': ['Yes', 'Yes', 'No', 'No'], 'b': range(4)})

In [90]: frame.describe()
Out[90]: 
              b
count  4.000000
mean   1.500000
std    1.290994
min    0.000000
25%    0.750000
50%    1.500000
75%    2.250000
max    3.000000

This behaviour can be controlled by providing a list of types as include/exclude arguments. The special value all can also be used:

In [91]: frame.describe(include=['object'])
Out[91]: 
         a
count    4
unique   2
top     No
freq     2

In [92]: frame.describe(include=['number'])
Out[92]: 
              b
count  4.000000
mean   1.500000
std    1.290994
min    0.000000
25%    0.750000
50%    1.500000
75%    2.250000
max    3.000000

In [93]: frame.describe(include='all')
Out[93]: 
          a         b
count     4  4.000000
unique    2       NaN
top      No       NaN
freq      2       NaN
mean    NaN  1.500000
std     NaN  1.290994
min     NaN  0.000000
25%     NaN  0.750000
50%     NaN  1.500000
75%     NaN  2.250000
max     NaN  3.000000

That feature relies on select_dtypes. Refer to there for details about accepted inputs.

Index of Min/Max Values

The idxmin() and idxmax() functions on Series and DataFrame compute the index labels with the minimum and maximum corresponding values:

In [94]: s1 = pd.Series(np.random.randn(5))

In [95]: s1
Out[95]: 
0   -0.872725
1    1.522411
2    0.080594
3   -1.676067
4    0.435804
dtype: float64

In [96]: s1.idxmin(), s1.idxmax()
Out[96]: (3, 1)

In [97]: df1 = pd.DataFrame(np.random.randn(5,3), columns=['A','B','C'])

In [98]: df1
Out[98]: 
          A         B         C
0  0.445734 -1.649461  0.169660
1  1.246181  0.131682 -2.001988
2 -1.273023  0.870502  0.214583
3  0.088452 -0.173364  1.207466
4  0.546121  0.409515 -0.310515

In [99]: df1.idxmin(axis=0)
Out[99]: 
A    2
B    0
C    1
dtype: int64

In [100]: df1.idxmax(axis=1)
Out[100]: 
0    A
1    A
2    B
3    C
4    A
dtype: object

When there are multiple rows (or columns) matching the minimum or maximum value, idxmin() and idxmax() return the first matching index:

In [101]: df3 = pd.DataFrame([2, 1, 1, 3, np.nan], columns=['A'], index=list('edcba'))

In [102]: df3
Out[102]: 
     A
e  2.0
d  1.0
c  1.0
b  3.0
a  NaN

In [103]: df3['A'].idxmin()
Out[103]: 'd'

Note

idxmin and idxmax are called argmin and argmax in NumPy.

Value counts (histogramming) / Mode

The value_counts() Series method and top-level function computes a histogram of a 1D array of values. It can also be used as a function on regular arrays:

In [104]: data = np.random.randint(0, 7, size=50)

In [105]: data
Out[105]: 
array([5, 3, 2, 2, 1, 4, 0, 4, 0, 2, 0, 6, 4, 1, 6, 3, 3, 0, 2, 1, 0, 5, 5,
       3, 6, 1, 5, 6, 2, 0, 0, 6, 3, 3, 5, 0, 4, 3, 3, 3, 0, 6, 1, 3, 5, 5,
       0, 4, 0, 6])

In [106]: s = pd.Series(data)

In [107]: s.value_counts()
Out[107]: 
0    11
3    10
6     7
5     7
4     5
2     5
1     5
dtype: int64

In [108]: pd.value_counts(data)
Out[108]: 
0    11
3    10
6     7
5     7
4     5
2     5
1     5
dtype: int64

Similarly, you can get the most frequently occurring value(s) (the mode) of the values in a Series or DataFrame:

In [109]: s5 = pd.Series([1, 1, 3, 3, 3, 5, 5, 7, 7, 7])

In [110]: s5.mode()
Out[110]: 
0    3
1    7
dtype: int64

In [111]: df5 = pd.DataFrame({"A": np.random.randint(0, 7, size=50),
   .....:                     "B": np.random.randint(-10, 15, size=50)})
   .....: 

In [112]: df5.mode()
Out[112]: 
   A  B
0  1 -5

Discretization and quantiling

Continuous values can be discretized using the cut() (bins based on values) and qcut() (bins based on sample quantiles) functions:

In [113]: arr = np.random.randn(20)

In [114]: factor = pd.cut(arr, 4)

In [115]: factor
Out[115]: 
[(-0.645, 0.336], (-2.61, -1.626], (-1.626, -0.645], (-1.626, -0.645], (-1.626, -0.645], ..., (0.336, 1.316], (0.336, 1.316], (0.336, 1.316], (0.336, 1.316], (-2.61, -1.626]]
Length: 20
Categories (4, object): [(-2.61, -1.626] < (-1.626, -0.645] < (-0.645, 0.336] < (0.336, 1.316]]

In [116]: factor = pd.cut(arr, [-5, -1, 0, 1, 5])

In [117]: factor
Out[117]: 
[(-1, 0], (-5, -1], (-1, 0], (-5, -1], (-1, 0], ..., (0, 1], (1, 5], (0, 1], (0, 1], (-5, -1]]
Length: 20
Categories (4, object): [(-5, -1] < (-1, 0] < (0, 1] < (1, 5]]

qcut() computes sample quantiles. For example, we could slice up some normally distributed data into equal-size quartiles like so:

In [118]: arr = np.random.randn(30)

In [119]: factor = pd.qcut(arr, [0, .25, .5, .75, 1])

In [120]: factor
Out[120]: 
[(-0.139, 1.00736], (1.00736, 1.976], (1.00736, 1.976], [-1.0705, -0.439], [-1.0705, -0.439], ..., (1.00736, 1.976], [-1.0705, -0.439], (-0.439, -0.139], (-0.439, -0.139], (-0.439, -0.139]]
Length: 30
Categories (4, object): [[-1.0705, -0.439] < (-0.439, -0.139] < (-0.139, 1.00736] < (1.00736, 1.976]]

In [121]: pd.value_counts(factor)
Out[121]: 
(1.00736, 1.976]     8
[-1.0705, -0.439]    8
(-0.139, 1.00736]    7
(-0.439, -0.139]     7
dtype: int64

We can also pass infinite values to define the bins:

In [122]: arr = np.random.randn(20)

In [123]: factor = pd.cut(arr, [-np.inf, 0, np.inf])

In [124]: factor
Out[124]: 
[(-inf, 0], (0, inf], (0, inf], (0, inf], (-inf, 0], ..., (-inf, 0], (0, inf], (-inf, 0], (-inf, 0], (0, inf]]
Length: 20
Categories (2, object): [(-inf, 0] < (0, inf]]

Function application

To apply your own or another library’s functions to pandas objects, you should be aware of the three methods below. The appropriate method to use depends on whether your function expects to operate on an entire DataFrame or Series, row- or column-wise, or elementwise.

  1. Tablewise Function Application: pipe()
  2. Row or Column-wise Function Application: apply()
  3. Elementwise function application: applymap()

Tablewise Function Application

New in version 0.16.2.

DataFrames and Series can of course just be passed into functions. However, if the function needs to be called in a chain, consider using the pipe() method. Compare the following

# f, g, and h are functions taking and returning ``DataFrames``
>>> f(g(h(df), arg1=1), arg2=2, arg3=3)

with the equivalent

>>> (df.pipe(h)
       .pipe(g, arg1=1)
       .pipe(f, arg2=2, arg3=3)
    )

Pandas encourages the second style, which is known as method chaining. pipe makes it easy to use your own or another library’s functions in method chains, alongside pandas’ methods.

In the example above, the functions f, g, and h each expected the DataFrame as the first positional argument. What if the function you wish to apply takes its data as, say, the second argument? In this case, provide pipe with a tuple of (callable, data_keyword). .pipe will route the DataFrame to the argument specified in the tuple.

For example, we can fit a regression using statsmodels. Their API expects a formula first and a DataFrame as the second argument, data. We pass in the function, keyword pair (sm.poisson, 'data') to pipe:

In [125]: import statsmodels.formula.api as sm

In [126]: bb = pd.read_csv('data/baseball.csv', index_col='id')

In [127]: (bb.query('h > 0')
   .....:    .assign(ln_h = lambda df: np.log(df.h))
   .....:    .pipe((sm.poisson, 'data'), 'hr ~ ln_h + year + g + C(lg)')
   .....:    .fit()
   .....:    .summary()
   .....: )
   .....: 
Optimization terminated successfully.
         Current function value: 2.116284
         Iterations 24
Out[127]: 
<class 'statsmodels.iolib.summary.Summary'>
"""
                          Poisson Regression Results                          
==============================================================================
Dep. Variable:                     hr   No. Observations:                   68
Model:                        Poisson   Df Residuals:                       63
Method:                           MLE   Df Model:                            4
Date:                Tue, 03 May 2016   Pseudo R-squ.:                  0.6878
Time:                        09:03:28   Log-Likelihood:                -143.91
converged:                       True   LL-Null:                       -460.91
                                        LLR p-value:                6.774e-136
===============================================================================
                  coef    std err          z      P>|z|      [95.0% Conf. Int.]
-------------------------------------------------------------------------------
Intercept   -1267.3636    457.867     -2.768      0.006     -2164.767  -369.960
C(lg)[T.NL]    -0.2057      0.101     -2.044      0.041        -0.403    -0.008
ln_h            0.9280      0.191      4.866      0.000         0.554     1.302
year            0.6301      0.228      2.762      0.006         0.183     1.077
g               0.0099      0.004      2.754      0.006         0.003     0.017
===============================================================================
"""

The pipe method is inspired by unix pipes and more recently dplyr and magrittr, which have introduced the popular (%>%) (read pipe) operator for R. The implementation of pipe here is quite clean and feels right at home in python. We encourage you to view the source code (pd.DataFrame.pipe?? in IPython).

Row or Column-wise Function Application

Arbitrary functions can be applied along the axes of a DataFrame or Panel using the apply() method, which, like the descriptive statistics methods, take an optional axis argument:

In [128]: df.apply(np.mean)
Out[128]: 
one     -0.251274
three    0.469799
two     -0.191421
dtype: float64

In [129]: df.apply(np.mean, axis=1)
Out[129]: 
a   -0.489066
b    0.273355
c    0.008348
d    0.011457
dtype: float64

In [130]: df.apply(lambda x: x.max() - x.min())
Out[130]: 
one      0.638161
three    1.301762
two      2.237808
dtype: float64

In [131]: df.apply(np.cumsum)
Out[131]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.765438 -0.177289  0.784662
c -0.753821  0.284925  0.335874
d       NaN  1.409398 -0.765684

In [132]: df.apply(np.exp)
Out[132]: 
        one     three       two
a  0.534436       NaN  0.703570
b  0.870320  0.837537  3.115063
c  1.011685  1.587586  0.638401
d       NaN  3.078592  0.332353

Depending on the return type of the function passed to apply(), the result will either be of lower dimension or the same dimension.

apply() combined with some cleverness can be used to answer many questions about a data set. For example, suppose we wanted to extract the date where the maximum value for each column occurred:

In [133]: tsdf = pd.DataFrame(np.random.randn(1000, 3), columns=['A', 'B', 'C'],
   .....:                     index=pd.date_range('1/1/2000', periods=1000))
   .....: 

In [134]: tsdf.apply(lambda x: x.idxmax())
Out[134]: 
A   2001-04-27
B   2002-06-02
C   2000-04-02
dtype: datetime64[ns]

You may also pass additional arguments and keyword arguments to the apply() method. For instance, consider the following function you would like to apply:

def subtract_and_divide(x, sub, divide=1):
    return (x - sub) / divide

You may then apply this function as follows:

df.apply(subtract_and_divide, args=(5,), divide=3)

Another useful feature is the ability to pass Series methods to carry out some Series operation on each column or row:

In [135]: tsdf
Out[135]: 
                   A         B         C
2000-01-01  1.796883 -0.930690  3.542846
2000-01-02 -1.242888 -0.695279 -1.000884
2000-01-03 -0.720299  0.546303 -0.082042
2000-01-04       NaN       NaN       NaN
2000-01-05       NaN       NaN       NaN
2000-01-06       NaN       NaN       NaN
2000-01-07       NaN       NaN       NaN
2000-01-08 -0.527402  0.933507  0.129646
2000-01-09 -0.338903 -1.265452 -1.969004
2000-01-10  0.532566  0.341548  0.150493

In [136]: tsdf.apply(pd.Series.interpolate)
Out[136]: 
                   A         B         C
2000-01-01  1.796883 -0.930690  3.542846
2000-01-02 -1.242888 -0.695279 -1.000884
2000-01-03 -0.720299  0.546303 -0.082042
2000-01-04 -0.681720  0.623743 -0.039704
2000-01-05 -0.643140  0.701184  0.002633
2000-01-06 -0.604561  0.778625  0.044971
2000-01-07 -0.565982  0.856066  0.087309
2000-01-08 -0.527402  0.933507  0.129646
2000-01-09 -0.338903 -1.265452 -1.969004
2000-01-10  0.532566  0.341548  0.150493

Finally, apply() takes an argument raw which is False by default, which converts each row or column into a Series before applying the function. When set to True, the passed function will instead receive an ndarray object, which has positive performance implications if you do not need the indexing functionality.

See also

The section on GroupBy demonstrates related, flexible functionality for grouping by some criterion, applying, and combining the results into a Series, DataFrame, etc.

Applying elementwise Python functions

Since not all functions can be vectorized (accept NumPy arrays and return another array or value), the methods applymap() on DataFrame and analogously map() on Series accept any Python function taking a single value and returning a single value. For example:

In [137]: df4
Out[137]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789
d       NaN  1.124472 -1.101558

In [138]: f = lambda x: len(str(x))

In [139]: df4['one'].map(f)
Out[139]: 
a    14
b    15
c    15
d     3
Name: one, dtype: int64

In [140]: df4.applymap(f)
Out[140]: 
   one  three  two
a   14      3   15
b   15     15   11
c   15     14   15
d    3     13   14

Series.map() has an additional feature which is that it can be used to easily “link” or “map” values defined by a secondary series. This is closely related to merging/joining functionality:

In [141]: s = pd.Series(['six', 'seven', 'six', 'seven', 'six'],
   .....:               index=['a', 'b', 'c', 'd', 'e'])
   .....: 

In [142]: t = pd.Series({'six' : 6., 'seven' : 7.})

In [143]: s
Out[143]: 
a      six
b    seven
c      six
d    seven
e      six
dtype: object

In [144]: s.map(t)
Out[144]: 
a    6.0
b    7.0
c    6.0
d    7.0
e    6.0
dtype: float64

Applying with a Panel

Applying with a Panel will pass a Series to the applied function. If the applied function returns a Series, the result of the application will be a Panel. If the applied function reduces to a scalar, the result of the application will be a DataFrame.

Note

Prior to 0.13.1 apply on a Panel would only work on ufuncs (e.g. np.sum/np.max).

In [145]: import pandas.util.testing as tm

In [146]: panel = tm.makePanel(5)

In [147]: panel
Out[147]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 3 (items) x 5 (major_axis) x 4 (minor_axis)
Items axis: ItemA to ItemC
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00
Minor_axis axis: A to D

In [148]: panel['ItemA']
Out[148]: 
                   A         B         C         D
2000-01-03  0.330418  1.893177  0.801111  0.528154
2000-01-04  1.761200  0.170247  0.445614 -0.029371
2000-01-05  0.567133 -0.916844  1.453046 -0.631117
2000-01-06 -0.251020  0.835024  2.430373 -0.172441
2000-01-07  1.020099  1.259919  0.653093 -1.020485

A transformational apply.

In [149]: result = panel.apply(lambda x: x*2, axis='items')

In [150]: result
Out[150]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 3 (items) x 5 (major_axis) x 4 (minor_axis)
Items axis: ItemA to ItemC
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00
Minor_axis axis: A to D

In [151]: result['ItemA']
Out[151]: 
                   A         B         C         D
2000-01-03  0.660836  3.786354  1.602222  1.056308
2000-01-04  3.522400  0.340494  0.891228 -0.058742
2000-01-05  1.134266 -1.833689  2.906092 -1.262234
2000-01-06 -0.502039  1.670047  4.860747 -0.344882
2000-01-07  2.040199  2.519838  1.306185 -2.040969

A reduction operation.

In [152]: panel.apply(lambda x: x.dtype, axis='items')
Out[152]: 
                  A        B        C        D
2000-01-03  float64  float64  float64  float64
2000-01-04  float64  float64  float64  float64
2000-01-05  float64  float64  float64  float64
2000-01-06  float64  float64  float64  float64
2000-01-07  float64  float64  float64  float64

A similar reduction type operation

In [153]: panel.apply(lambda x: x.sum(), axis='major_axis')
Out[153]: 
      ItemA     ItemB     ItemC
A  3.427831 -2.581431  0.840809
B  3.241522 -1.409935 -1.114512
C  5.783237  0.319672 -0.431906
D -1.325260 -2.914834  0.857043

This last reduction is equivalent to

In [154]: panel.sum('major_axis')
Out[154]: 
      ItemA     ItemB     ItemC
A  3.427831 -2.581431  0.840809
B  3.241522 -1.409935 -1.114512
C  5.783237  0.319672 -0.431906
D -1.325260 -2.914834  0.857043

A transformation operation that returns a Panel, but is computing the z-score across the major_axis.

In [155]: result = panel.apply(
   .....:            lambda x: (x-x.mean())/x.std(),
   .....:            axis='major_axis')
   .....: 

In [156]: result
Out[156]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 3 (items) x 5 (major_axis) x 4 (minor_axis)
Items axis: ItemA to ItemC
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00
Minor_axis axis: A to D

In [157]: result['ItemA']
Out[157]: 
                   A         B         C         D
2000-01-03 -0.469761  1.156225 -0.441347  1.341731
2000-01-04  1.422763 -0.444015 -0.882647  0.398661
2000-01-05 -0.156654 -1.453694  0.367936 -0.619210
2000-01-06 -1.238841  0.173423  1.581149  0.156654
2000-01-07  0.442494  0.568061 -0.625091 -1.277837

Apply can also accept multiple axes in the axis argument. This will pass a DataFrame of the cross-section to the applied function.

In [158]: f = lambda x: ((x.T-x.mean(1))/x.std(1)).T

In [159]: result = panel.apply(f, axis = ['items','major_axis'])

In [160]: result
Out[160]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 4 (items) x 5 (major_axis) x 3 (minor_axis)
Items axis: A to D
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00
Minor_axis axis: ItemA to ItemC

In [161]: result.loc[:,:,'ItemA']
Out[161]: 
                   A         B         C         D
2000-01-03  0.864236  1.132969  0.557316  0.575106
2000-01-04  0.795745  0.652527  0.534808 -0.070674
2000-01-05 -0.310864  0.558627  1.086688 -1.051477
2000-01-06 -0.001065  0.832460  0.846006  0.043602
2000-01-07  1.128946  1.152469 -0.218186 -0.891680

This is equivalent to the following

In [162]: result = pd.Panel(dict([ (ax, f(panel.loc[:,:,ax]))
   .....:                         for ax in panel.minor_axis ]))
   .....: 

In [163]: result
Out[163]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 4 (items) x 5 (major_axis) x 3 (minor_axis)
Items axis: A to D
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00
Minor_axis axis: ItemA to ItemC

In [164]: result.loc[:,:,'ItemA']
Out[164]: 
                   A         B         C         D
2000-01-03  0.864236  1.132969  0.557316  0.575106
2000-01-04  0.795745  0.652527  0.534808 -0.070674
2000-01-05 -0.310864  0.558627  1.086688 -1.051477
2000-01-06 -0.001065  0.832460  0.846006  0.043602
2000-01-07  1.128946  1.152469 -0.218186 -0.891680

Reindexing and altering labels

reindex() is the fundamental data alignment method in pandas. It is used to implement nearly all other features relying on label-alignment functionality. To reindex means to conform the data to match a given set of labels along a particular axis. This accomplishes several things:

  • Reorders the existing data to match a new set of labels
  • Inserts missing value (NA) markers in label locations where no data for that label existed
  • If specified, fill data for missing labels using logic (highly relevant to working with time series data)

Here is a simple example:

In [165]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [166]: s
Out[166]: 
a   -1.010924
b   -0.672504
c   -1.139222
d    0.354653
e    0.563622
dtype: float64

In [167]: s.reindex(['e', 'b', 'f', 'd'])
Out[167]: 
e    0.563622
b   -0.672504
f         NaN
d    0.354653
dtype: float64

Here, the f label was not contained in the Series and hence appears as NaN in the result.

With a DataFrame, you can simultaneously reindex the index and columns:

In [168]: df
Out[168]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789
d       NaN  1.124472 -1.101558

In [169]: df.reindex(index=['c', 'f', 'b'], columns=['three', 'two', 'one'])
Out[169]: 
      three       two       one
c  0.462215 -0.448789  0.011617
f       NaN       NaN       NaN
b -0.177289  1.136249 -0.138894

For convenience, you may utilize the reindex_axis() method, which takes the labels and a keyword axis parameter.

Note that the Index objects containing the actual axis labels can be shared between objects. So if we have a Series and a DataFrame, the following can be done:

In [170]: rs = s.reindex(df.index)

In [171]: rs
Out[171]: 
a   -1.010924
b   -0.672504
c   -1.139222
d    0.354653
dtype: float64

In [172]: rs.index is df.index
Out[172]: True

This means that the reindexed Series’s index is the same Python object as the DataFrame’s index.

See also

MultiIndex / Advanced Indexing is an even more concise way of doing reindexing.

Note

When writing performance-sensitive code, there is a good reason to spend some time becoming a reindexing ninja: many operations are faster on pre-aligned data. Adding two unaligned DataFrames internally triggers a reindexing step. For exploratory analysis you will hardly notice the difference (because reindex has been heavily optimized), but when CPU cycles matter sprinkling a few explicit reindex calls here and there can have an impact.

Reindexing to align with another object

You may wish to take an object and reindex its axes to be labeled the same as another object. While the syntax for this is straightforward albeit verbose, it is a common enough operation that the reindex_like() method is available to make this simpler:

In [173]: df2
Out[173]: 
        one       two
a -0.626544 -0.351587
b -0.138894  1.136249
c  0.011617 -0.448789

In [174]: df3
Out[174]: 
        one       two
a -0.375270 -0.463545
b  0.112379  1.024292
c  0.262891 -0.560746

In [175]: df.reindex_like(df2)
Out[175]: 
        one       two
a -0.626544 -0.351587
b -0.138894  1.136249
c  0.011617 -0.448789

Aligning objects with each other with align

The align() method is the fastest way to simultaneously align two objects. It supports a join argument (related to joining and merging):

  • join='outer': take the union of the indexes (default)
  • join='left': use the calling object’s index
  • join='right': use the passed object’s index
  • join='inner': intersect the indexes

It returns a tuple with both of the reindexed Series:

In [176]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e'])

In [177]: s1 = s[:4]

In [178]: s2 = s[1:]

In [179]: s1.align(s2)
Out[179]: 
(a   -0.365106
 b    1.092702
 c   -1.481449
 d    1.781190
 e         NaN
 dtype: float64, a         NaN
 b    1.092702
 c   -1.481449
 d    1.781190
 e   -0.031543
 dtype: float64)

In [180]: s1.align(s2, join='inner')
Out[180]: 
(b    1.092702
 c   -1.481449
 d    1.781190
 dtype: float64, b    1.092702
 c   -1.481449
 d    1.781190
 dtype: float64)

In [181]: s1.align(s2, join='left')
Out[181]: 
(a   -0.365106
 b    1.092702
 c   -1.481449
 d    1.781190
 dtype: float64, a         NaN
 b    1.092702
 c   -1.481449
 d    1.781190
 dtype: float64)

For DataFrames, the join method will be applied to both the index and the columns by default:

In [182]: df.align(df2, join='inner')
Out[182]: 
(        one       two
 a -0.626544 -0.351587
 b -0.138894  1.136249
 c  0.011617 -0.448789,         one       two
 a -0.626544 -0.351587
 b -0.138894  1.136249
 c  0.011617 -0.448789)

You can also pass an axis option to only align on the specified axis:

In [183]: df.align(df2, join='inner', axis=0)
Out[183]: 
(        one     three       two
 a -0.626544       NaN -0.351587
 b -0.138894 -0.177289  1.136249
 c  0.011617  0.462215 -0.448789,         one       two
 a -0.626544 -0.351587
 b -0.138894  1.136249
 c  0.011617 -0.448789)

If you pass a Series to DataFrame.align(), you can choose to align both objects either on the DataFrame’s index or columns using the axis argument:

In [184]: df.align(df2.ix[0], axis=1)
Out[184]: 
(        one     three       two
 a -0.626544       NaN -0.351587
 b -0.138894 -0.177289  1.136249
 c  0.011617  0.462215 -0.448789
 d       NaN  1.124472 -1.101558, one     -0.626544
 three         NaN
 two     -0.351587
 Name: a, dtype: float64)

Filling while reindexing

reindex() takes an optional parameter method which is a filling method chosen from the following table:

Method Action
pad / ffill Fill values forward
bfill / backfill Fill values backward
nearest Fill from the nearest index value

We illustrate these fill methods on a simple Series:

In [185]: rng = pd.date_range('1/3/2000', periods=8)

In [186]: ts = pd.Series(np.random.randn(8), index=rng)

In [187]: ts2 = ts[[0, 3, 6]]

In [188]: ts
Out[188]: 
2000-01-03    0.480993
2000-01-04    0.604244
2000-01-05   -0.487265
2000-01-06    1.990533
2000-01-07    0.327007
2000-01-08    1.053639
2000-01-09   -2.927808
2000-01-10    0.082065
Freq: D, dtype: float64

In [189]: ts2
Out[189]: 
2000-01-03    0.480993
2000-01-06    1.990533
2000-01-09   -2.927808
dtype: float64

In [190]: ts2.reindex(ts.index)
Out[190]: 
2000-01-03    0.480993
2000-01-04         NaN
2000-01-05         NaN
2000-01-06    1.990533
2000-01-07         NaN
2000-01-08         NaN
2000-01-09   -2.927808
2000-01-10         NaN
Freq: D, dtype: float64

In [191]: ts2.reindex(ts.index, method='ffill')
Out[191]: 
2000-01-03    0.480993
2000-01-04    0.480993
2000-01-05    0.480993
2000-01-06    1.990533
2000-01-07    1.990533
2000-01-08    1.990533
2000-01-09   -2.927808
2000-01-10   -2.927808
Freq: D, dtype: float64

In [192]: ts2.reindex(ts.index, method='bfill')
Out[192]: 
2000-01-03    0.480993
2000-01-04    1.990533
2000-01-05    1.990533
2000-01-06    1.990533
2000-01-07   -2.927808
2000-01-08   -2.927808
2000-01-09   -2.927808
2000-01-10         NaN
Freq: D, dtype: float64

In [193]: ts2.reindex(ts.index, method='nearest')
Out[193]: 
2000-01-03    0.480993
2000-01-04    0.480993
2000-01-05    1.990533
2000-01-06    1.990533
2000-01-07    1.990533
2000-01-08   -2.927808
2000-01-09   -2.927808
2000-01-10   -2.927808
Freq: D, dtype: float64

These methods require that the indexes are ordered increasing or decreasing.

Note that the same result could have been achieved using fillna (except for method='nearest') or interpolate:

In [194]: ts2.reindex(ts.index).fillna(method='ffill')
Out[194]: 
2000-01-03    0.480993
2000-01-04    0.480993
2000-01-05    0.480993
2000-01-06    1.990533
2000-01-07    1.990533
2000-01-08    1.990533
2000-01-09   -2.927808
2000-01-10   -2.927808
Freq: D, dtype: float64

reindex() will raise a ValueError if the index is not monotonic increasing or decreasing. fillna() and interpolate() will not make any checks on the order of the index.

Limits on filling while reindexing

The limit and tolerance arguments provide additional control over filling while reindexing. Limit specifies the maximum count of consecutive matches:

In [195]: ts2.reindex(ts.index, method='ffill', limit=1)
Out[195]: 
2000-01-03    0.480993
2000-01-04    0.480993
2000-01-05         NaN
2000-01-06    1.990533
2000-01-07    1.990533
2000-01-08         NaN
2000-01-09   -2.927808
2000-01-10   -2.927808
Freq: D, dtype: float64

In contrast, tolerance specifies the maximum distance between the index and indexer values:

In [196]: ts2.reindex(ts.index, method='ffill', tolerance='1 day')
Out[196]: 
2000-01-03    0.480993
2000-01-04    0.480993
2000-01-05         NaN
2000-01-06    1.990533
2000-01-07    1.990533
2000-01-08         NaN
2000-01-09   -2.927808
2000-01-10   -2.927808
Freq: D, dtype: float64

Notice that when used on a DatetimeIndex, TimedeltaIndex or PeriodIndex, tolerance will coerced into a Timedelta if possible. This allows you to specify tolerance with appropriate strings.

Dropping labels from an axis

A method closely related to reindex is the drop() function. It removes a set of labels from an axis:

In [197]: df
Out[197]: 
        one     three       two
a -0.626544       NaN -0.351587
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789
d       NaN  1.124472 -1.101558

In [198]: df.drop(['a', 'd'], axis=0)
Out[198]: 
        one     three       two
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789

In [199]: df.drop(['one'], axis=1)
Out[199]: 
      three       two
a       NaN -0.351587
b -0.177289  1.136249
c  0.462215 -0.448789
d  1.124472 -1.101558

Note that the following also works, but is a bit less obvious / clean:

In [200]: df.reindex(df.index.difference(['a', 'd']))
Out[200]: 
        one     three       two
b -0.138894 -0.177289  1.136249
c  0.011617  0.462215 -0.448789

Renaming / mapping labels

The rename() method allows you to relabel an axis based on some mapping (a dict or Series) or an arbitrary function.

In [201]: s
Out[201]: 
a   -0.365106
b    1.092702
c   -1.481449
d    1.781190
e   -0.031543
dtype: float64

In [202]: s.rename(str.upper)
Out[202]: 
A   -0.365106
B    1.092702
C   -1.481449
D    1.781190
E   -0.031543
dtype: float64

If you pass a function, it must return a value when called with any of the labels (and must produce a set of unique values). But if you pass a dict or Series, it need only contain a subset of the labels as keys:

In [203]: df.rename(columns={'one' : 'foo', 'two' : 'bar'},
   .....:           index={'a' : 'apple', 'b' : 'banana', 'd' : 'durian'})
   .....: 
Out[203]: 
             foo     three       bar
apple  -0.626544       NaN -0.351587
banana -0.138894 -0.177289  1.136249
c       0.011617  0.462215 -0.448789
durian       NaN  1.124472 -1.101558

The rename() method also provides an inplace named parameter that is by default False and copies the underlying data. Pass inplace=True to rename the data in place.

New in version 0.18.0.

Finally, rename() also accepts a scalar or list-like for altering the Series.name attribute.

In [204]: s.rename("scalar-name")
Out[204]: 
a   -0.365106
b    1.092702
c   -1.481449
d    1.781190
e   -0.031543
Name: scalar-name, dtype: float64

The Panel class has a related rename_axis() class which can rename any of its three axes.

Iteration

The behavior of basic iteration over pandas objects depends on the type. When iterating over a Series, it is regarded as array-like, and basic iteration produces the values. Other data structures, like DataFrame and Panel, follow the dict-like convention of iterating over the “keys” of the objects.

In short, basic iteration (for i in object) produces:

  • Series: values
  • DataFrame: column labels
  • Panel: item labels

Thus, for example, iterating over a DataFrame gives you the column names:

In [205]: df = pd.DataFrame({'col1' : np.random.randn(3), 'col2' : np.random.randn(3)},
   .....:                   index=['a', 'b', 'c'])
   .....: 

In [206]: for col in df:
   .....:     print(col)
   .....: 
col1
col2

Pandas objects also have the dict-like iteritems() method to iterate over the (key, value) pairs.

To iterate over the rows of a DataFrame, you can use the following methods:

  • iterrows(): Iterate over the rows of a DataFrame as (index, Series) pairs. This converts the rows to Series objects, which can change the dtypes and has some performance implications.
  • itertuples(): Iterate over the rows of a DataFrame as namedtuples of the values. This is a lot faster than iterrows(), and is in most cases preferable to use to iterate over the values of a DataFrame.

Warning

Iterating through pandas objects is generally slow. In many cases, iterating manually over the rows is not needed and can be avoided with one of the following approaches:

  • Look for a vectorized solution: many operations can be performed using built-in methods or numpy functions, (boolean) indexing, ...
  • When you have a function that cannot work on the full DataFrame/Series at once, it is better to use apply() instead of iterating over the values. See the docs on function application.
  • If you need to do iterative manipulations on the values but performance is important, consider writing the inner loop using e.g. cython or numba. See the enhancing performance section for some examples of this approach.

Warning

You should never modify something you are iterating over. This is not guaranteed to work in all cases. Depending on the data types, the iterator returns a copy and not a view, and writing to it will have no effect!

For example, in the following case setting the value has no effect:

In [207]: df = pd.DataFrame({'a': [1, 2, 3], 'b': ['a', 'b', 'c']})

In [208]: for index, row in df.iterrows():
   .....:     row['a'] = 10
   .....: 

In [209]: df
Out[209]: 
   a  b
0  1  a
1  2  b
2  3  c

iteritems

Consistent with the dict-like interface, iteritems() iterates through key-value pairs:

  • Series: (index, scalar value) pairs
  • DataFrame: (column, Series) pairs
  • Panel: (item, DataFrame) pairs

For example:

In [210]: for item, frame in wp.iteritems():
   .....:     print(item)
   .....:     print(frame)
   .....: 
Item1
                   A         B         C         D
2000-01-01 -1.032011  0.969818 -0.962723  1.382083
2000-01-02 -0.938794  0.669142 -0.433567 -0.273610
2000-01-03  0.680433 -0.308450 -0.276099 -1.821168
2000-01-04 -1.993606 -1.927385 -2.027924  1.624972
2000-01-05  0.551135  3.059267  0.455264 -0.030740
Item2
                   A         B         C         D
2000-01-01  0.935716  1.061192 -2.107852  0.199905
2000-01-02  0.323586 -0.641630 -0.587514  0.053897
2000-01-03  0.194889 -0.381994  0.318587  2.089075
2000-01-04 -0.728293 -0.090255 -0.748199  1.318931
2000-01-05 -2.029766  0.792652  0.461007 -0.542749

iterrows

iterrows() allows you to iterate through the rows of a DataFrame as Series objects. It returns an iterator yielding each index value along with a Series containing the data in each row:

In [211]: for row_index, row in df.iterrows():
   .....:     print('%s\n%s' % (row_index, row))
   .....: 
0
a    1
b    a
Name: 0, dtype: object
1
a    2
b    b
Name: 1, dtype: object
2
a    3
b    c
Name: 2, dtype: object

Note

Because iterrows() returns a Series for each row, it does not preserve dtypes across the rows (dtypes are preserved across columns for DataFrames). For example,

In [212]: df_orig = pd.DataFrame([[1, 1.5]], columns=['int', 'float'])

In [213]: df_orig.dtypes
Out[213]: 
int        int64
float    float64
dtype: object

In [214]: row = next(df_orig.iterrows())[1]

In [215]: row
Out[215]: 
int      1.0
float    1.5
Name: 0, dtype: float64

All values in row, returned as a Series, are now upcasted to floats, also the original integer value in column x:

In [216]: row['int'].dtype
Out[216]: dtype('float64')

In [217]: df_orig['int'].dtype
Out[217]: dtype('int64')

To preserve dtypes while iterating over the rows, it is better to use itertuples() which returns namedtuples of the values and which is generally much faster as iterrows.

For instance, a contrived way to transpose the DataFrame would be:

In [218]: df2 = pd.DataFrame({'x': [1, 2, 3], 'y': [4, 5, 6]})

In [219]: print(df2)
   x  y
0  1  4
1  2  5
2  3  6

In [220]: print(df2.T)
   0  1  2
x  1  2  3
y  4  5  6

In [221]: df2_t = pd.DataFrame(dict((idx,values) for idx, values in df2.iterrows()))

In [222]: print(df2_t)
   0  1  2
x  1  2  3
y  4  5  6

itertuples

The itertuples() method will return an iterator yielding a namedtuple for each row in the DataFrame. The first element of the tuple will be the row’s corresponding index value, while the remaining values are the row values.

For instance,

In [223]: for row in df.itertuples():
   .....:     print(row)
   .....: 
Pandas(Index=0, a=1, b='a')
Pandas(Index=1, a=2, b='b')
Pandas(Index=2, a=3, b='c')

This method does not convert the row to a Series object but just returns the values inside a namedtuple. Therefore, itertuples() preserves the data type of the values and is generally faster as iterrows().

Note

The column names will be renamed to positional names if they are invalid Python identifiers, repeated, or start with an underscore. With a large number of columns (>255), regular tuples are returned.

.dt accessor

Series has an accessor to succinctly return datetime like properties for the values of the Series, if it is a datetime/period like Series. This will return a Series, indexed like the existing Series.

# datetime
In [224]: s = pd.Series(pd.date_range('20130101 09:10:12', periods=4))

In [225]: s
Out[225]: 
0   2013-01-01 09:10:12
1   2013-01-02 09:10:12
2   2013-01-03 09:10:12
3   2013-01-04 09:10:12
dtype: datetime64[ns]

In [226]: s.dt.hour
Out[226]: 
0    9
1    9
2    9
3    9
dtype: int64

In [227]: s.dt.second
Out[227]: 
0    12
1    12
2    12
3    12
dtype: int64

In [228]: s.dt.day
Out[228]: 
0    1
1    2
2    3
3    4
dtype: int64

This enables nice expressions like this:

In [229]: s[s.dt.day==2]
Out[229]: 
1   2013-01-02 09:10:12
dtype: datetime64[ns]

You can easily produces tz aware transformations:

In [230]: stz = s.dt.tz_localize('US/Eastern')

In [231]: stz
Out[231]: 
0   2013-01-01 09:10:12-05:00
1   2013-01-02 09:10:12-05:00
2   2013-01-03 09:10:12-05:00
3   2013-01-04 09:10:12-05:00
dtype: datetime64[ns, US/Eastern]

In [232]: stz.dt.tz
Out[232]: <DstTzInfo 'US/Eastern' LMT-1 day, 19:04:00 STD>

You can also chain these types of operations:

In [233]: s.dt.tz_localize('UTC').dt.tz_convert('US/Eastern')
Out[233]: 
0   2013-01-01 04:10:12-05:00
1   2013-01-02 04:10:12-05:00
2   2013-01-03 04:10:12-05:00
3   2013-01-04 04:10:12-05:00
dtype: datetime64[ns, US/Eastern]

You can also format datetime values as strings with Series.dt.strftime() which supports the same format as the standard strftime().

# DatetimeIndex
In [234]: s = pd.Series(pd.date_range('20130101', periods=4))

In [235]: s
Out[235]: 
0   2013-01-01
1   2013-01-02
2   2013-01-03
3   2013-01-04
dtype: datetime64[ns]

In [236]: s.dt.strftime('%Y/%m/%d')
Out[236]: 
0    2013/01/01
1    2013/01/02
2    2013/01/03
3    2013/01/04
dtype: object
# PeriodIndex
In [237]: s = pd.Series(pd.period_range('20130101', periods=4))

In [238]: s
Out[238]: 
0   2013-01-01
1   2013-01-02
2   2013-01-03
3   2013-01-04
dtype: object

In [239]: s.dt.strftime('%Y/%m/%d')
Out[239]: 
0    2013/01/01
1    2013/01/02
2    2013/01/03
3    2013/01/04
dtype: object

The .dt accessor works for period and timedelta dtypes.

# period
In [240]: s = pd.Series(pd.period_range('20130101', periods=4, freq='D'))

In [241]: s
Out[241]: 
0   2013-01-01
1   2013-01-02
2   2013-01-03
3   2013-01-04
dtype: object

In [242]: s.dt.year
Out[242]: 
0    2013
1    2013
2    2013
3    2013
dtype: int64

In [243]: s.dt.day
Out[243]: 
0    1
1    2
2    3
3    4
dtype: int64
# timedelta
In [244]: s = pd.Series(pd.timedelta_range('1 day 00:00:05', periods=4, freq='s'))

In [245]: s
Out[245]: 
0   1 days 00:00:05
1   1 days 00:00:06
2   1 days 00:00:07
3   1 days 00:00:08
dtype: timedelta64[ns]

In [246]: s.dt.days
Out[246]: 
0    1
1    1
2    1
3    1
dtype: int64

In [247]: s.dt.seconds
Out[247]: 
0    5
1    6
2    7
3    8
dtype: int64

In [248]: s.dt.components
Out[248]: 
   days  hours  minutes  seconds  milliseconds  microseconds  nanoseconds
0     1      0        0        5             0             0            0
1     1      0        0        6             0             0            0
2     1      0        0        7             0             0            0
3     1      0        0        8             0             0            0

Note

Series.dt will raise a TypeError if you access with a non-datetimelike values

Vectorized string methods

Series is equipped with a set of string processing methods that make it easy to operate on each element of the array. Perhaps most importantly, these methods exclude missing/NA values automatically. These are accessed via the Series’s str attribute and generally have names matching the equivalent (scalar) built-in string methods. For example:

In [249]: s = pd.Series(['A', 'B', 'C', 'Aaba', 'Baca', np.nan, 'CABA', 'dog', 'cat'])

In [250]: s.str.lower()
Out[250]: 
0       a
1       b
2       c
3    aaba
4    baca
5     NaN
6    caba
7     dog
8     cat
dtype: object

Powerful pattern-matching methods are provided as well, but note that pattern-matching generally uses regular expressions by default (and in some cases always uses them).

Please see Vectorized String Methods for a complete description.

Sorting

Warning

The sorting API is substantially changed in 0.17.0, see here for these changes. In particular, all sorting methods now return a new object by default, and DO NOT operate in-place (except by passing inplace=True).

There are two obvious kinds of sorting that you may be interested in: sorting by label and sorting by actual values.

By Index

The primary method for sorting axis labels (indexes) are the Series.sort_index() and the DataFrame.sort_index() methods.

In [251]: unsorted_df = df.reindex(index=['a', 'd', 'c', 'b'],
   .....:                          columns=['three', 'two', 'one'])
   .....: 

# DataFrame
In [252]: unsorted_df.sort_index()
Out[252]: 
   three  two  one
a    NaN  NaN  NaN
b    NaN  NaN  NaN
c    NaN  NaN  NaN
d    NaN  NaN  NaN

In [253]: unsorted_df.sort_index(ascending=False)
Out[253]: 
   three  two  one
d    NaN  NaN  NaN
c    NaN  NaN  NaN
b    NaN  NaN  NaN
a    NaN  NaN  NaN

In [254]: unsorted_df.sort_index(axis=1)
Out[254]: 
   one  three  two
a  NaN    NaN  NaN
d  NaN    NaN  NaN
c  NaN    NaN  NaN
b  NaN    NaN  NaN

# Series
In [255]: unsorted_df['three'].sort_index()
Out[255]: 
a   NaN
b   NaN
c   NaN
d   NaN
Name: three, dtype: float64

By Values

The Series.sort_values() and DataFrame.sort_values() are the entry points for value sorting (that is the values in a column or row). DataFrame.sort_values() can accept an optional by argument for axis=0 which will use an arbitrary vector or a column name of the DataFrame to determine the sort order:

In [256]: df1 = pd.DataFrame({'one':[2,1,1,1],'two':[1,3,2,4],'three':[5,4,3,2]})

In [257]: df1.sort_values(by='two')
Out[257]: 
   one  three  two
0    2      5    1
2    1      3    2
1    1      4    3
3    1      2    4

The by argument can take a list of column names, e.g.:

In [258]: df1[['one', 'two', 'three']].sort_values(by=['one','two'])
Out[258]: 
   one  two  three
2    1    2      3
1    1    3      4
3    1    4      2
0    2    1      5

These methods have special treatment of NA values via the na_position argument:

In [259]: s[2] = np.nan

In [260]: s.sort_values()
Out[260]: 
0       A
3    Aaba
1       B
4    Baca
6    CABA
8     cat
7     dog
2     NaN
5     NaN
dtype: object

In [261]: s.sort_values(na_position='first')
Out[261]: 
2     NaN
5     NaN
0       A
3    Aaba
1       B
4    Baca
6    CABA
8     cat
7     dog
dtype: object

searchsorted

Series has the searchsorted() method, which works similar to numpy.ndarray.searchsorted().

In [262]: ser = pd.Series([1, 2, 3])

In [263]: ser.searchsorted([0, 3])
Out[263]: array([0, 2])

In [264]: ser.searchsorted([0, 4])
Out[264]: array([0, 3])

In [265]: ser.searchsorted([1, 3], side='right')
Out[265]: array([1, 3])

In [266]: ser.searchsorted([1, 3], side='left')
Out[266]: array([0, 2])

In [267]: ser = pd.Series([3, 1, 2])

In [268]: ser.searchsorted([0, 3], sorter=np.argsort(ser))
Out[268]: array([0, 2])

smallest / largest values

New in version 0.14.0.

Series has the nsmallest() and nlargest() methods which return the smallest or largest n values. For a large Series this can be much faster than sorting the entire Series and calling head(n) on the result.

In [269]: s = pd.Series(np.random.permutation(10))

In [270]: s
Out[270]: 
0    9
1    8
2    5
3    3
4    6
5    7
6    0
7    2
8    4
9    1
dtype: int64

In [271]: s.sort_values()
Out[271]: 
6    0
9    1
7    2
3    3
8    4
2    5
4    6
5    7
1    8
0    9
dtype: int64

In [272]: s.nsmallest(3)
Out[272]: 
6    0
9    1
7    2
dtype: int64

In [273]: s.nlargest(3)
Out[273]: 
0    9
1    8
5    7
dtype: int64

New in version 0.17.0.

DataFrame also has the nlargest and nsmallest methods.

In [274]: df = pd.DataFrame({'a': [-2, -1, 1, 10, 8, 11, -1],
   .....:                    'b': list('abdceff'),
   .....:                    'c': [1.0, 2.0, 4.0, 3.2, np.nan, 3.0, 4.0]})
   .....: 

In [275]: df.nlargest(3, 'a')
Out[275]: 
    a  b    c
5  11  f  3.0
3  10  c  3.2
4   8  e  NaN

In [276]: df.nlargest(5, ['a', 'c'])
Out[276]: 
    a  b    c
5  11  f  3.0
3  10  c  3.2
4   8  e  NaN
2   1  d  4.0
1  -1  b  2.0

In [277]: df.nsmallest(3, 'a')
Out[277]: 
   a  b    c
0 -2  a  1.0
1 -1  b  2.0
6 -1  f  4.0

In [278]: df.nsmallest(5, ['a', 'c'])
Out[278]: 
   a  b    c
0 -2  a  1.0
1 -1  b  2.0
6 -1  f  4.0
2  1  d  4.0
4  8  e  NaN

Sorting by a multi-index column

You must be explicit about sorting when the column is a multi-index, and fully specify all levels to by.

In [279]: df1.columns = pd.MultiIndex.from_tuples([('a','one'),('a','two'),('b','three')])

In [280]: df1.sort_values(by=('a','two'))
Out[280]: 
    a         b
  one two three
3   1   2     4
2   1   3     2
1   1   4     3
0   2   5     1

Copying

The copy() method on pandas objects copies the underlying data (though not the axis indexes, since they are immutable) and returns a new object. Note that it is seldom necessary to copy objects. For example, there are only a handful of ways to alter a DataFrame in-place:

  • Inserting, deleting, or modifying a column
  • Assigning to the index or columns attributes
  • For homogeneous data, directly modifying the values via the values attribute or advanced indexing

To be clear, no pandas methods have the side effect of modifying your data; almost all methods return new objects, leaving the original object untouched. If data is modified, it is because you did so explicitly.

dtypes

The main types stored in pandas objects are float, int, bool, datetime64[ns] and datetime64[ns, tz] (in >= 0.17.0), timedelta[ns], category (in >= 0.15.0), and object. In addition these dtypes have item sizes, e.g. int64 and int32. See Series with TZ for more detail on datetime64[ns, tz] dtypes.

A convenient dtypes attribute for DataFrames returns a Series with the data type of each column.

In [281]: dft = pd.DataFrame(dict(A = np.random.rand(3),
   .....:                         B = 1,
   .....:                         C = 'foo',
   .....:                         D = pd.Timestamp('20010102'),
   .....:                         E = pd.Series([1.0]*3).astype('float32'),
   .....:                                     F = False,
   .....:                                     G = pd.Series([1]*3,dtype='int8')))
   .....: 

In [282]: dft
Out[282]: 
          A  B    C          D    E      F  G
0  0.954940  1  foo 2001-01-02  1.0  False  1
1  0.318163  1  foo 2001-01-02  1.0  False  1
2  0.985803  1  foo 2001-01-02  1.0  False  1

In [283]: dft.dtypes
Out[283]: 
A           float64
B             int64
C            object
D    datetime64[ns]
E           float32
F              bool
G              int8
dtype: object

On a Series use the dtype attribute.

In [284]: dft['A'].dtype
Out[284]: dtype('float64')

If a pandas object contains data multiple dtypes IN A SINGLE COLUMN, the dtype of the column will be chosen to accommodate all of the data types (object is the most general).

# these ints are coerced to floats
In [285]: pd.Series([1, 2, 3, 4, 5, 6.])
Out[285]: 
0    1.0
1    2.0
2    3.0
3    4.0
4    5.0
5    6.0
dtype: float64

# string data forces an ``object`` dtype
In [286]: pd.Series([1, 2, 3, 6., 'foo'])
Out[286]: 
0      1
1      2
2      3
3      6
4    foo
dtype: object

The method get_dtype_counts() will return the number of columns of each type in a DataFrame:

In [287]: dft.get_dtype_counts()
Out[287]: 
bool              1
datetime64[ns]    1
float32           1
float64           1
int64             1
int8              1
object            1
dtype: int64

Numeric dtypes will propagate and can coexist in DataFrames (starting in v0.11.0). If a dtype is passed (either directly via the dtype keyword, a passed ndarray, or a passed Series, then it will be preserved in DataFrame operations. Furthermore, different numeric dtypes will NOT be combined. The following example will give you a taste.

In [288]: df1 = pd.DataFrame(np.random.randn(8, 1), columns=['A'], dtype='float32')

In [289]: df1
Out[289]: 
          A
0  0.647650
1  0.822993
2  1.778703
3 -1.543048
4 -0.123256
5  2.239740
6 -0.143778
7 -2.885090

In [290]: df1.dtypes
Out[290]: 
A    float32
dtype: object

In [291]: df2 = pd.DataFrame(dict( A = pd.Series(np.random.randn(8), dtype='float16'),
   .....:                         B = pd.Series(np.random.randn(8)),
   .....:                         C = pd.Series(np.array(np.random.randn(8), dtype='uint8')) ))
   .....: 

In [292]: df2
Out[292]: 
          A         B    C
0  0.027588  0.296947    0
1 -1.150391  0.007045  255
2  0.246460  0.707877    1
3 -0.455078  0.950661    0
4 -1.507812  0.087527    0
5 -0.502441 -0.339212    0
6  0.528809 -0.278698    0
7  0.590332  1.775379    0

In [293]: df2.dtypes
Out[293]: 
A    float16
B    float64
C      uint8
dtype: object

defaults

By default integer types are int64 and float types are float64, REGARDLESS of platform (32-bit or 64-bit). The following will all result in int64 dtypes.

In [294]: pd.DataFrame([1, 2], columns=['a']).dtypes
Out[294]: 
a    int64
dtype: object

In [295]: pd.DataFrame({'a': [1, 2]}).dtypes
Out[295]: 
a    int64
dtype: object

In [296]: pd.DataFrame({'a': 1 }, index=list(range(2))).dtypes
Out[296]: 
a    int64
dtype: object

Numpy, however will choose platform-dependent types when creating arrays. The following WILL result in int32 on 32-bit platform.

In [297]: frame = pd.DataFrame(np.array([1, 2]))

upcasting

Types can potentially be upcasted when combined with other types, meaning they are promoted from the current type (say int to float)

In [298]: df3 = df1.reindex_like(df2).fillna(value=0.0) + df2

In [299]: df3
Out[299]: 
          A         B      C
0  0.675238  0.296947    0.0
1 -0.327398  0.007045  255.0
2  2.025163  0.707877    1.0
3 -1.998126  0.950661    0.0
4 -1.631068  0.087527    0.0
5  1.737299 -0.339212    0.0
6  0.385030 -0.278698    0.0
7 -2.294758  1.775379    0.0

In [300]: df3.dtypes
Out[300]: 
A    float32
B    float64
C    float64
dtype: object

The values attribute on a DataFrame return the lower-common-denominator of the dtypes, meaning the dtype that can accommodate ALL of the types in the resulting homogeneous dtyped numpy array. This can force some upcasting.

In [301]: df3.values.dtype
Out[301]: dtype('float64')

astype

You can use the astype() method to explicitly convert dtypes from one to another. These will by default return a copy, even if the dtype was unchanged (pass copy=False to change this behavior). In addition, they will raise an exception if the astype operation is invalid.

Upcasting is always according to the numpy rules. If two different dtypes are involved in an operation, then the more general one will be used as the result of the operation.

In [302]: df3
Out[302]: 
          A         B      C
0  0.675238  0.296947    0.0
1 -0.327398  0.007045  255.0
2  2.025163  0.707877    1.0
3 -1.998126  0.950661    0.0
4 -1.631068  0.087527    0.0
5  1.737299 -0.339212    0.0
6  0.385030 -0.278698    0.0
7 -2.294758  1.775379    0.0

In [303]: df3.dtypes
Out[303]: 
A    float32
B    float64
C    float64
dtype: object

# conversion of dtypes
In [304]: df3.astype('float32').dtypes
Out[304]: 
A    float32
B    float32
C    float32
dtype: object

object conversion

convert_objects() is a method to try to force conversion of types from the object dtype to other types. To force conversion of specific types that are number like, e.g. could be a string that represents a number, pass convert_numeric=True. This will force strings and numbers alike to be numbers if possible, otherwise they will be set to np.nan.

In [305]: df3['D'] = '1.'

In [306]: df3['E'] = '1'

In [307]: df3.convert_objects(convert_numeric=True).dtypes
Out[307]: 
A    float32
B    float64
C    float64
D    float64
E      int64
dtype: object

# same, but specific dtype conversion
In [308]: df3['D'] = df3['D'].astype('float16')

In [309]: df3['E'] = df3['E'].astype('int32')

In [310]: df3.dtypes
Out[310]: 
A    float32
B    float64
C    float64
D    float16
E      int32
dtype: object

To force conversion to datetime64[ns], pass convert_dates='coerce'. This will convert any datetime-like object to dates, forcing other values to NaT. This might be useful if you are reading in data which is mostly dates, but occasionally has non-dates intermixed and you want to represent as missing.

In [311]: import datetime

In [312]: s = pd.Series([datetime.datetime(2001,1,1,0,0),
   .....:               'foo', 1.0, 1, pd.Timestamp('20010104'),
   .....:               '20010105'], dtype='O')
   .....: 

In [313]: s
Out[313]: 
0    2001-01-01 00:00:00
1                    foo
2                      1
3                      1
4    2001-01-04 00:00:00
5               20010105
dtype: object

In [314]: pd.to_datetime(s, errors='coerce')
Out[314]: 
0   2001-01-01
1          NaT
2          NaT
3          NaT
4   2001-01-04
5   2001-01-05
dtype: datetime64[ns]

In addition, convert_objects() will attempt the soft conversion of any object dtypes, meaning that if all the objects in a Series are of the same type, the Series will have that dtype.

gotchas

Performing selection operations on integer type data can easily upcast the data to floating. The dtype of the input data will be preserved in cases where nans are not introduced (starting in 0.11.0) See also integer na gotchas

In [315]: dfi = df3.astype('int32')

In [316]: dfi['E'] = 1

In [317]: dfi
Out[317]: 
   A  B    C  D  E
0  0  0    0  1  1
1  0  0  255  1  1
2  2  0    1  1  1
3 -1  0    0  1  1
4 -1  0    0  1  1
5  1  0    0  1  1
6  0  0    0  1  1
7 -2  1    0  1  1

In [318]: dfi.dtypes
Out[318]: 
A    int32
B    int32
C    int32
D    int32
E    int64
dtype: object

In [319]: casted = dfi[dfi>0]

In [320]: casted
Out[320]: 
     A    B      C  D  E
0  NaN  NaN    NaN  1  1
1  NaN  NaN  255.0  1  1
2  2.0  NaN    1.0  1  1
3  NaN  NaN    NaN  1  1
4  NaN  NaN    NaN  1  1
5  1.0  NaN    NaN  1  1
6  NaN  NaN    NaN  1  1
7  NaN  1.0    NaN  1  1

In [321]: casted.dtypes
Out[321]: 
A    float64
B    float64
C    float64
D      int32
E      int64
dtype: object

While float dtypes are unchanged.

In [322]: dfa = df3.copy()

In [323]: dfa['A'] = dfa['A'].astype('float32')

In [324]: dfa.dtypes
Out[324]: 
A    float32
B    float64
C    float64
D    float16
E      int32
dtype: object

In [325]: casted = dfa[df2>0]

In [326]: casted
Out[326]: 
          A         B      C   D   E
0  0.675238  0.296947    NaN NaN NaN
1       NaN  0.007045  255.0 NaN NaN
2  2.025163  0.707877    1.0 NaN NaN
3       NaN  0.950661    NaN NaN NaN
4       NaN  0.087527    NaN NaN NaN
5       NaN       NaN    NaN NaN NaN
6  0.385030       NaN    NaN NaN NaN
7 -2.294758  1.775379    NaN NaN NaN

In [327]: casted.dtypes
Out[327]: 
A    float32
B    float64
C    float64
D    float16
E    float64
dtype: object

Selecting columns based on dtype

New in version 0.14.1.

The select_dtypes() method implements subsetting of columns based on their dtype.

First, let’s create a DataFrame with a slew of different dtypes:

In [328]: df = pd.DataFrame({'string': list('abc'),
   .....:                    'int64': list(range(1, 4)),
   .....:                    'uint8': np.arange(3, 6).astype('u1'),
   .....:                    'float64': np.arange(4.0, 7.0),
   .....:                    'bool1': [True, False, True],
   .....:                    'bool2': [False, True, False],
   .....:                    'dates': pd.date_range('now', periods=3).values,
   .....:                    'category': pd.Series(list("ABC")).astype('category')})
   .....: 

In [329]: df['tdeltas'] = df.dates.diff()

In [330]: df['uint64'] = np.arange(3, 6).astype('u8')

In [331]: df['other_dates'] = pd.date_range('20130101', periods=3).values

In [332]: df['tz_aware_dates'] = pd.date_range('20130101', periods=3, tz='US/Eastern')

In [333]: df
Out[333]: 
   bool1  bool2 category                      dates  float64  int64 string  \
0   True  False        A 2016-05-03 09:03:31.005096      4.0      1      a   
1  False   True        B 2016-05-04 09:03:31.005096      5.0      2      b   
2   True  False        C 2016-05-05 09:03:31.005096      6.0      3      c   

   uint8  tdeltas  uint64 other_dates            tz_aware_dates  
0      3      NaT       3  2013-01-01 2013-01-01 00:00:00-05:00  
1      4   1 days       4  2013-01-02 2013-01-02 00:00:00-05:00  
2      5   1 days       5  2013-01-03 2013-01-03 00:00:00-05:00  

And the dtypes

In [334]: df.dtypes
Out[334]: 
bool1                                   bool
bool2                                   bool
category                            category
dates                         datetime64[ns]
float64                              float64
int64                                  int64
string                                object
uint8                                  uint8
tdeltas                      timedelta64[ns]
uint64                                uint64
other_dates                   datetime64[ns]
tz_aware_dates    datetime64[ns, US/Eastern]
dtype: object

select_dtypes() has two parameters include and exclude that allow you to say “give me the columns WITH these dtypes” (include) and/or “give the columns WITHOUT these dtypes” (exclude).

For example, to select bool columns

In [335]: df.select_dtypes(include=[bool])
Out[335]: 
   bool1  bool2
0   True  False
1  False   True
2   True  False

You can also pass the name of a dtype in the numpy dtype hierarchy:

In [336]: df.select_dtypes(include=['bool'])
Out[336]: 
   bool1  bool2
0   True  False
1  False   True
2   True  False

select_dtypes() also works with generic dtypes as well.

For example, to select all numeric and boolean columns while excluding unsigned integers

In [337]: df.select_dtypes(include=['number', 'bool'], exclude=['unsignedinteger'])
Out[337]: 
   bool1  bool2  float64  int64  tdeltas
0   True  False      4.0      1      NaT
1  False   True      5.0      2   1 days
2   True  False      6.0      3   1 days

To select string columns you must use the object dtype:

In [338]: df.select_dtypes(include=['object'])
Out[338]: 
  string
0      a
1      b
2      c

To see all the child dtypes of a generic dtype like numpy.number you can define a function that returns a tree of child dtypes:

In [339]: def subdtypes(dtype):
   .....:     subs = dtype.__subclasses__()
   .....:     if not subs:
   .....:         return dtype
   .....:     return [dtype, [subdtypes(dt) for dt in subs]]
   .....: 

All numpy dtypes are subclasses of numpy.generic:

In [340]: subdtypes(np.generic)
Out[340]: 
[numpy.generic,
 [[numpy.number,
   [[numpy.integer,
     [[numpy.signedinteger,
       [numpy.int8,
        numpy.int16,
        numpy.int32,
        numpy.int64,
        numpy.int64,
        numpy.timedelta64]],
      [numpy.unsignedinteger,
       [numpy.uint8,
        numpy.uint16,
        numpy.uint32,
        numpy.uint64,
        numpy.uint64]]]],
    [numpy.inexact,
     [[numpy.floating,
       [numpy.float16, numpy.float32, numpy.float64, numpy.float128]],
      [numpy.complexfloating,
       [numpy.complex64, numpy.complex128, numpy.complex256]]]]]],
  [numpy.flexible,
   [[numpy.character, [numpy.string_, numpy.unicode_]],
    [numpy.void, [numpy.record]]]],
  numpy.bool_,
  numpy.datetime64,
  numpy.object_]]

Note

Pandas also defines the types category, and datetime64[ns, tz], which are not integrated into the normal numpy hierarchy and wont show up with the above function.

Note

The include and exclude parameters must be non-string sequences.