Essential basic functionality#

Here we discuss a lot of the essential functionality common to the pandas data structures. To begin, let’s create some example objects like we did in the 10 minutes to pandas 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"])

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 [4]: long_series = pd.Series(np.random.randn(1000))
In [5]: long_series.head()
Out[5]: 
0 -1.157892
1 -1.344312
2 0.844885
3 1.075770
4 -0.109050
dtype: float64
In [6]: long_series.tail(3)
Out[6]: 
997 -0.289388
998 -1.020544
999 0.589993
dtype: float64

Attributes and underlying data#

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

Note, these attributes can be safely assigned to!

In [7]: df[:2]
Out[7]: 
 A B C
2000年01月01日 -0.173215 0.119209 -1.044236
2000年01月02日 -0.861849 -2.104569 -0.494929
In [8]: df.columns = [x.lower() for x in df.columns]
In [9]: df
Out[9]: 
 a b c
2000年01月01日 -0.173215 0.119209 -1.044236
2000年01月02日 -0.861849 -2.104569 -0.494929
2000年01月03日 1.071804 0.721555 -0.706771
2000年01月04日 -1.039575 0.271860 -0.424972
2000年01月05日 0.567020 0.276232 -1.087401
2000年01月06日 -0.673690 0.113648 -1.478427
2000年01月07日 0.524988 0.404705 0.577046
2000年01月08日 -1.715002 -1.039268 -0.370647

pandas objects (Index, Series, DataFrame) can be thought of as containers for arrays, which hold the actual data and do the actual computation. For many types, the underlying array is a numpy.ndarray. However, pandas and 3rd party libraries may extend NumPy’s type system to add support for custom arrays (see dtypes).

To get the actual data inside a Index or Series, use the .array property

In [10]: s.array
Out[10]: 
<NumpyExtensionArray>
[ 0.4691122999071863, -0.2828633443286633, -1.5090585031735124,
 -1.1356323710171934, 1.2121120250208506]
Length: 5, dtype: float64
In [11]: s.index.array
Out[11]: 
<NumpyExtensionArray>
['a', 'b', 'c', 'd', 'e']
Length: 5, dtype: object

array will always be an ExtensionArray. The exact details of what an ExtensionArray is and why pandas uses them are a bit beyond the scope of this introduction. See dtypes for more.

If you know you need a NumPy array, use to_numpy() or numpy.asarray().

In [12]: s.to_numpy()
Out[12]: array([ 0.4691, -0.2829, -1.5091, -1.1356, 1.2121])
In [13]: np.asarray(s)
Out[13]: array([ 0.4691, -0.2829, -1.5091, -1.1356, 1.2121])

When the Series or Index is backed by an ExtensionArray, to_numpy() may involve copying data and coercing values. See dtypes for more.

to_numpy() gives some control over the dtype of the resulting numpy.ndarray. For example, consider datetimes with timezones. NumPy doesn’t have a dtype to represent timezone-aware datetimes, so there are two possibly useful representations:

1. An object-dtype numpy.ndarray with Timestamp objects, each with the correct tz

2. A datetime64[ns] -dtype numpy.ndarray, where the values have been converted to UTC and the timezone discarded

Timezones may be preserved with dtype=object

In [14]: ser = pd.Series(pd.date_range("2000", periods=2, tz="CET"))
In [15]: ser.to_numpy(dtype=object)
Out[15]: 
array([Timestamp('2000年01月01日 00:00:00+0100', tz='CET'),
 Timestamp('2000年01月02日 00:00:00+0100', tz='CET')], dtype=object)

Or thrown away with dtype='datetime64[ns]'

In [16]: ser.to_numpy(dtype="datetime64[ns]")
Out[16]: 
array(['1999年12月31日T23:00:00.000000000', '2000年01月01日T23:00:00.000000000'],
 dtype='datetime64[ns]')

Getting the "raw data" inside a DataFrame is possibly a bit more complex. When your DataFrame only has a single data type for all the columns, DataFrame.to_numpy() will return the underlying data:

In [17]: df.to_numpy()
Out[17]: 
array([[-0.1732, 0.1192, -1.0442],
 [-0.8618, -2.1046, -0.4949],
 [ 1.0718, 0.7216, -0.7068],
 [-1.0396, 0.2719, -0.425 ],
 [ 0.567 , 0.2762, -1.0874],
 [-0.6737, 0.1136, -1.4784],
 [ 0.525 , 0.4047, 0.577 ],
 [-1.715 , -1.0393, -0.3706]])

If a DataFrame 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.

In the past, pandas recommended Series.values or DataFrame.values for extracting the data from a Series or DataFrame. You’ll still find references to these in old code bases and online. Going forward, we recommend avoiding .values and using .array or .to_numpy(). .values has the following drawbacks:

1. When your Series contains an extension type, it’s unclear whether Series.values returns a NumPy array or the extension array. Series.array will always return an ExtensionArray, and will never copy data. Series.to_numpy() will always return a NumPy array, potentially at the cost of copying / coercing values.

2. When your DataFrame contains a mixture of data types, DataFrame.values may involve copying data and coercing values to a common dtype, a relatively expensive operation. DataFrame.to_numpy(), being a method, makes it clearer that the returned NumPy array may not be a view on the same data in the DataFrame.

Accelerated operations#

pandas has support for accelerating certain types of binary numerical and boolean operations using the numexpr library 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.

These are both enabled to be used by default, you can control this by setting the options:

pd.set_option("compute.use_bottleneck", False)
pd.set_option("compute.use_numexpr", False)

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 [18]: 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 [19]: df
Out[19]: 
 one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [20]: row = df.iloc[1]
In [21]: column = df["two"]
In [22]: df.sub(row, axis="columns")
Out[22]: 
 one two three
a 1.051928 -0.139606 NaN
b 0.000000 0.000000 0.000000
c 0.352192 -0.433754 1.277825
d NaN -1.632779 -0.562782
In [23]: df.sub(row, axis=1)
Out[23]: 
 one two three
a 1.051928 -0.139606 NaN
b 0.000000 0.000000 0.000000
c 0.352192 -0.433754 1.277825
d NaN -1.632779 -0.562782
In [24]: df.sub(column, axis="index")
Out[24]: 
 one two three
a -0.377535 0.0 NaN
b -1.569069 0.0 -1.962513
c -0.783123 0.0 -0.250933
d NaN 0.0 -0.892516
In [25]: df.sub(column, axis=0)
Out[25]: 
 one two three
a -0.377535 0.0 NaN
b -1.569069 0.0 -1.962513
c -0.783123 0.0 -0.250933
d NaN 0.0 -0.892516

Furthermore you can align a level of a MultiIndexed DataFrame with a Series.

In [26]: dfmi = df.copy()
In [27]: dfmi.index = pd.MultiIndex.from_tuples(
 ....:  [(1, "a"), (1, "b"), (1, "c"), (2, "a")], names=["first", "second"]
 ....: )
 ....: 
In [28]: dfmi.sub(column, axis=0, level="second")
Out[28]: 
 one two three
first second 
1 a -0.377535 0.000000 NaN
 b -1.569069 0.000000 -1.962513
 c -0.783123 0.000000 -0.250933
2 a NaN -1.493173 -2.385688

Series and Index also support the divmod() builtin. This function takes the floor division and modulo operation at the same time returning a two-tuple of the same type as the left hand side. For example:

In [29]: s = pd.Series(np.arange(10))
In [30]: s
Out[30]: 
0 0
1 1
2 2
3 3
4 4
5 5
6 6
7 7
8 8
9 9
dtype: int64
In [31]: div, rem = divmod(s, 3)
In [32]: div
Out[32]: 
0 0
1 0
2 0
3 1
4 1
5 1
6 2
7 2
8 2
9 3
dtype: int64
In [33]: rem
Out[33]: 
0 0
1 1
2 2
3 0
4 1
5 2
6 0
7 1
8 2
9 0
dtype: int64
In [34]: idx = pd.Index(np.arange(10))
In [35]: idx
Out[35]: Index([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype='int64')
In [36]: div, rem = divmod(idx, 3)
In [37]: div
Out[37]: Index([0, 0, 0, 1, 1, 1, 2, 2, 2, 3], dtype='int64')
In [38]: rem
Out[38]: Index([0, 1, 2, 0, 1, 2, 0, 1, 2, 0], dtype='int64')

We can also do elementwise divmod():

In [39]: div, rem = divmod(s, [2, 2, 3, 3, 4, 4, 5, 5, 6, 6])
In [40]: div
Out[40]: 
0 0
1 0
2 0
3 1
4 1
5 1
6 1
7 1
8 1
9 1
dtype: int64
In [41]: rem
Out[41]: 
0 0
1 1
2 2
3 0
4 0
5 1
6 1
7 2
8 2
9 3
dtype: int64

Missing data / operations with fill values#

In Series and DataFrame, 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 [42]: df2 = df.copy()
In [43]: df2.loc["a", "three"] = 1.0
In [44]: df
Out[44]: 
 one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [45]: df2
Out[45]: 
 one two three
a 1.394981 1.772517 1.000000
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [46]: df + df2
Out[46]: 
 one two three
a 2.789963 3.545034 NaN
b 0.686107 3.824246 -0.100780
c 1.390491 2.956737 2.454870
d NaN 0.558688 -1.226343
In [47]: df.add(df2, fill_value=0)
Out[47]: 
 one two three
a 2.789963 3.545034 1.000000
b 0.686107 3.824246 -0.100780
c 1.390491 2.956737 2.454870
d NaN 0.558688 -1.226343

Flexible comparisons#

Series and DataFrame have the binary comparison methods eq, ne, lt, gt, le, and ge whose behavior is analogous to the binary arithmetic operations described above:

In [48]: df.gt(df2)
Out[48]: 
 one two three
a False False False
b False False False
c False False False
d False False False
In [49]: df2.ne(df)
Out[49]: 
 one two three
a False False True
b False False False
c False False False
d True False False

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

Boolean reductions#

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

In [50]: (df > 0).all()
Out[50]: 
one False
two True
three False
dtype: bool
In [51]: (df > 0).any()
Out[51]: 
one True
two True
three True
dtype: bool

You can reduce to a final boolean value.

In [52]: (df > 0).any().any()
Out[52]: True

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

In [53]: df.empty
Out[53]: False
In [54]: pd.DataFrame(columns=list("ABC")).empty
Out[54]: True

Warning

Asserting the truthiness of a pandas object will raise an error, as the testing of the emptiness or values is ambiguous.

In [55]: if df:
 ....:  print(True)
 ....: 
---------------------------------------------------------------------------
ValueErrorTraceback (most recent call last)
<ipython-input-55-318d08b2571a> in ?()
----> 1 if df:
2 print(True)
~/work/pandas/pandas/pandas/core/generic.py in ?(self)
1575 @final
1576 def__nonzero__(self) -> NoReturn:
-> 1577 raise ValueError(
1578 f"The truth value of a {type(self).__name__} is ambiguous. "
1579 "Use a.empty, a.bool(), a.item(), a.any() or a.all()."
1580 )
ValueError: The truth value of a DataFrame is ambiguous. Use a.empty, a.bool(), a.item(), a.any() or a.all().

In [56]: df and df2
---------------------------------------------------------------------------
ValueErrorTraceback (most recent call last)
<ipython-input-56-b241b64bb471> in ?()
----> 1 df and df2
~/work/pandas/pandas/pandas/core/generic.py in ?(self)
1575 @final
1576 def__nonzero__(self) -> NoReturn:
-> 1577 raise ValueError(
1578 f"The truth value of a {type(self).__name__} is ambiguous. "
1579 "Use a.empty, a.bool(), a.item(), a.any() or a.all()."
1580 )
ValueError: The truth value of a DataFrame is ambiguous. Use a.empty, a.bool(), a.item(), a.any() or a.all().

See gotchas for a more detailed discussion.

Comparing if objects are equivalent#

Often you may find that 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 [57]: df + df == df * 2
Out[57]: 
 one two three
a True True False
b True True True
c True True True
d False True True
In [58]: (df + df == df * 2).all()
Out[58]: 
one False
two True
three False
dtype: bool

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

In [59]: np.nan == np.nan
Out[59]: False

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

In [60]: (df + df).equals(df * 2)
Out[60]: True

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

In [61]: df1 = pd.DataFrame({"col": ["foo", 0, np.nan]})
In [62]: df2 = pd.DataFrame({"col": [np.nan, 0, "foo"]}, index=[2, 1, 0])
In [63]: df1.equals(df2)
Out[63]: False
In [64]: df1.equals(df2.sort_index())
Out[64]: True

Comparing array-like objects#

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

In [65]: pd.Series(["foo", "bar", "baz"]) == "foo"
Out[65]: 
0 True
1 False
2 False
dtype: bool
In [66]: pd.Index(["foo", "bar", "baz"]) == "foo"
Out[66]: array([ True, False, False])

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

In [67]: pd.Series(["foo", "bar", "baz"]) == pd.Index(["foo", "bar", "qux"])
Out[67]: 
0 True
1 True
2 False
dtype: bool
In [68]: pd.Series(["foo", "bar", "baz"]) == np.array(["foo", "bar", "qux"])
Out[68]: 
0 True
1 True
2 False
dtype: bool

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

In [69]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo', 'bar'])
---------------------------------------------------------------------------
ValueErrorTraceback (most recent call last)
Cell In[69], line 1
----> 1 pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo', 'bar'])
File ~/work/pandas/pandas/pandas/core/ops/common.py:76, in _unpack_zerodim_and_defer.<locals>.new_method(self, other)
72 return NotImplemented
74 other = item_from_zerodim(other)
---> 76 return method(self, other)
File ~/work/pandas/pandas/pandas/core/arraylike.py:40, in OpsMixin.__eq__(self, other)
38 @unpack_zerodim_and_defer("__eq__")
39 def__eq__(self, other):
---> 40 return self._cmp_method(other, operator.eq)
File ~/work/pandas/pandas/pandas/core/series.py:6125, in Series._cmp_method(self, other, op)
6122 res_name = ops.get_op_result_name(self, other)
6124 if isinstance(other, Series) and not self._indexed_same(other):
-> 6125 raise ValueError("Can only compare identically-labeled Series objects")
6127 lvalues = self._values
6128 rvalues = extract_array(other, extract_numpy=True, extract_range=True)
ValueError: Can only compare identically-labeled Series objects
In [70]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo'])
---------------------------------------------------------------------------
ValueErrorTraceback (most recent call last)
Cell In[70], line 1
----> 1 pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo'])
File ~/work/pandas/pandas/pandas/core/ops/common.py:76, in _unpack_zerodim_and_defer.<locals>.new_method(self, other)
72 return NotImplemented
74 other = item_from_zerodim(other)
---> 76 return method(self, other)
File ~/work/pandas/pandas/pandas/core/arraylike.py:40, in OpsMixin.__eq__(self, other)
38 @unpack_zerodim_and_defer("__eq__")
39 def__eq__(self, other):
---> 40 return self._cmp_method(other, operator.eq)
File ~/work/pandas/pandas/pandas/core/series.py:6125, in Series._cmp_method(self, other, op)
6122 res_name = ops.get_op_result_name(self, other)
6124 if isinstance(other, Series) and not self._indexed_same(other):
-> 6125 raise ValueError("Can only compare identically-labeled Series objects")
6127 lvalues = self._values
6128 rvalues = extract_array(other, extract_numpy=True, extract_range=True)
ValueError: Can only compare identically-labeled Series objects

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 [71]: df1 = pd.DataFrame(
 ....:  {"A": [1.0, np.nan, 3.0, 5.0, np.nan], "B": [np.nan, 2.0, 3.0, np.nan, 6.0]}
 ....: )
 ....: 
In [72]: df2 = pd.DataFrame(
 ....:  {
 ....:  "A": [5.0, 2.0, 4.0, np.nan, 3.0, 7.0],
 ....:  "B": [np.nan, np.nan, 3.0, 4.0, 6.0, 8.0],
 ....:  }
 ....: )
 ....: 
In [73]: df1
Out[73]: 
 A B
0 1.0 NaN
1 NaN 2.0
2 3.0 3.0
3 5.0 NaN
4 NaN 6.0
In [74]: df2
Out[74]: 
 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 [75]: df1.combine_first(df2)
Out[75]: 
 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.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 [76]: defcombiner(x, y):
 ....:  return np.where(pd.isna(x), y, x)
 ....: 
In [77]: df1.combine(df2, combiner)
Out[77]: 
 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#

There exists a large number of methods for computing descriptive statistics and other related operations on Series, DataFrame. 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)

For example:

In [78]: df
Out[78]: 
 one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [79]: df.mean(0)
Out[79]: 
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [80]: df.mean(1)
Out[80]: 
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64

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

In [81]: df.sum(0, skipna=False)
Out[81]: 
one NaN
two 5.442353
three NaN
dtype: float64
In [82]: df.sum(axis=1, skipna=True)
Out[82]: 
a 3.167498
b 2.204786
c 3.401050
d -0.333828
dtype: float64

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

In [83]: ts_stand = (df - df.mean()) / df.std()
In [84]: ts_stand.std()
Out[84]: 
one 1.0
two 1.0
three 1.0
dtype: float64
In [85]: xs_stand = df.sub(df.mean(1), axis=0).div(df.std(1), axis=0)
In [86]: xs_stand.std(1)
Out[86]: 
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 NaN values. This is somewhat different from expanding() and rolling() since NaN behavior is furthermore dictated by a min_periods parameter.

In [87]: df.cumsum()
Out[87]: 
 one two three
a 1.394981 1.772517 NaN
b 1.738035 3.684640 -0.050390
c 2.433281 5.163008 1.177045
d NaN 5.442353 0.563873

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-NA observations
sum	Sum of values
mean	Mean of values
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 [88]: np.mean(df["one"])
Out[88]: 0.8110935116651192
In [89]: np.mean(df["one"].to_numpy())
Out[89]: nan

Series.nunique() will return the number of unique non-NA values in a Series:

In [90]: series = pd.Series(np.random.randn(500))
In [91]: series[20:500] = np.nan
In [92]: series[10:20] = 5
In [93]: series.nunique()
Out[93]: 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 [94]: series = pd.Series(np.random.randn(1000))
In [95]: series[::2] = np.nan
In [96]: series.describe()
Out[96]: 
count 500.000000
mean -0.021292
std 1.015906
min -2.683763
25% -0.699070
50% -0.069718
75% 0.714483
max 3.160915
dtype: float64
In [97]: frame = pd.DataFrame(np.random.randn(1000, 5), columns=["a", "b", "c", "d", "e"])
In [98]: frame.iloc[::2] = np.nan
In [99]: frame.describe()
Out[99]: 
 a b c d e
count 500.000000 500.000000 500.000000 500.000000 500.000000
mean 0.033387 0.030045 -0.043719 -0.051686 0.005979
std 1.017152 0.978743 1.025270 1.015988 1.006695
min -3.000951 -2.637901 -3.303099 -3.159200 -3.188821
25% -0.647623 -0.576449 -0.712369 -0.691338 -0.691115
50% 0.047578 -0.021499 -0.023888 -0.032652 -0.025363
75% 0.729907 0.775880 0.618896 0.670047 0.649748
max 2.740139 2.752332 3.004229 2.728702 3.240991

You can select specific percentiles to include in the output:

In [100]: series.describe(percentiles=[0.05, 0.25, 0.75, 0.95])
Out[100]: 
count 500.000000
mean -0.021292
std 1.015906
min -2.683763
5% -1.645423
25% -0.699070
50% -0.069718
75% 0.714483
95% 1.711409
max 3.160915
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 [101]: s = pd.Series(["a", "a", "b", "b", "a", "a", np.nan, "c", "d", "a"])
In [102]: s.describe()
Out[102]: 
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 [103]: frame = pd.DataFrame({"a": ["Yes", "Yes", "No", "No"], "b": range(4)})
In [104]: frame.describe()
Out[104]: 
 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 behavior can be controlled by providing a list of types as include/exclude arguments. The special value all can also be used:

In [105]: frame.describe(include=["object"])
Out[105]: 
 a
count 4
unique 2
top Yes
freq 2
In [106]: frame.describe(include=["number"])
Out[106]: 
 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 [107]: frame.describe(include="all")
Out[107]: 
 a b
count 4 4.000000
unique 2 NaN
top Yes 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 [108]: s1 = pd.Series(np.random.randn(5))
In [109]: s1
Out[109]: 
0 1.118076
1 -0.352051
2 -1.242883
3 -1.277155
4 -0.641184
dtype: float64
In [110]: s1.idxmin(), s1.idxmax()
Out[110]: (3, 0)
In [111]: df1 = pd.DataFrame(np.random.randn(5, 3), columns=["A", "B", "C"])
In [112]: df1
Out[112]: 
 A B C
0 -0.327863 -0.946180 -0.137570
1 -0.186235 -0.257213 -0.486567
2 -0.507027 -0.871259 -0.111110
3 2.000339 -2.430505 0.089759
4 -0.321434 -0.033695 0.096271
In [113]: df1.idxmin(axis=0)
Out[113]: 
A 2
B 3
C 1
dtype: int64
In [114]: df1.idxmax(axis=1)
Out[114]: 
0 C
1 A
2 C
3 A
4 C
dtype: object

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

In [115]: df3 = pd.DataFrame([2, 1, 1, 3, np.nan], columns=["A"], index=list("edcba"))
In [116]: df3
Out[116]: 
 A
e 2.0
d 1.0
c 1.0
b 3.0
a NaN
In [117]: df3["A"].idxmin()
Out[117]: 'd'

Note

idxmin and idxmax are called argmin and argmax in NumPy.

Value counts (histogramming) / mode#

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

In [118]: data = np.random.randint(0, 7, size=50)
In [119]: data
Out[119]: 
array([6, 6, 2, 3, 5, 3, 2, 5, 4, 5, 4, 3, 4, 5, 0, 2, 0, 4, 2, 0, 3, 2,
 2, 5, 6, 5, 3, 4, 6, 4, 3, 5, 6, 4, 3, 6, 2, 6, 6, 2, 3, 4, 2, 1,
 6, 2, 6, 1, 5, 4])
In [120]: s = pd.Series(data)
In [121]: s.value_counts()
Out[121]: 
6 10
2 10
4 9
3 8
5 8
0 3
1 2
Name: count, dtype: int64

The value_counts() method can be used to count combinations across multiple columns. By default all columns are used but a subset can be selected using the subset argument.

In [122]: data = {"a": [1, 2, 3, 4], "b": ["x", "x", "y", "y"]}
In [123]: frame = pd.DataFrame(data)
In [124]: frame.value_counts()
Out[124]: 
a b
1 x 1
2 x 1
3 y 1
4 y 1
Name: count, dtype: int64

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

In [125]: s5 = pd.Series([1, 1, 3, 3, 3, 5, 5, 7, 7, 7])
In [126]: s5.mode()
Out[126]: 
0 3
1 7
dtype: int64
In [127]: df5 = pd.DataFrame(
 .....:  {
 .....:  "A": np.random.randint(0, 7, size=50),
 .....:  "B": np.random.randint(-10, 15, size=50),
 .....:  }
 .....: )
 .....: 
In [128]: df5.mode()
Out[128]: 
 A B
0 1.0 -9
1 NaN 10
2 NaN 13

Discretization and quantiling#

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

In [129]: arr = np.random.randn(20)
In [130]: factor = pd.cut(arr, 4)
In [131]: factor
Out[131]: 
[(-0.251, 0.464], (-0.968, -0.251], (0.464, 1.179], (-0.251, 0.464], (-0.968, -0.251], ..., (-0.251, 0.464], (-0.968, -0.251], (-0.968, -0.251], (-0.968, -0.251], (-0.968, -0.251]]
Length: 20
Categories (4, interval[float64, right]): [(-0.968, -0.251] < (-0.251, 0.464] < (0.464, 1.179] <
 (1.179, 1.893]]
In [132]: factor = pd.cut(arr, [-5, -1, 0, 1, 5])
In [133]: factor
Out[133]: 
[(0, 1], (-1, 0], (0, 1], (0, 1], (-1, 0], ..., (-1, 0], (-1, 0], (-1, 0], (-1, 0], (-1, 0]]
Length: 20
Categories (4, interval[int64, right]): [(-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 [134]: arr = np.random.randn(30)
In [135]: factor = pd.qcut(arr, [0, 0.25, 0.5, 0.75, 1])
In [136]: factor
Out[136]: 
[(0.569, 1.184], (-2.278, -0.301], (-2.278, -0.301], (0.569, 1.184], (0.569, 1.184], ..., (-0.301, 0.569], (1.184, 2.346], (1.184, 2.346], (-0.301, 0.569], (-2.278, -0.301]]
Length: 30
Categories (4, interval[float64, right]): [(-2.278, -0.301] < (-0.301, 0.569] < (0.569, 1.184] <
 (1.184, 2.346]]

We can also pass infinite values to define the bins:

In [137]: arr = np.random.randn(20)
In [138]: factor = pd.cut(arr, [-np.inf, 0, np.inf])
In [139]: factor
Out[139]: 
[(-inf, 0.0], (0.0, inf], (0.0, inf], (-inf, 0.0], (-inf, 0.0], ..., (-inf, 0.0], (-inf, 0.0], (-inf, 0.0], (0.0, inf], (0.0, inf]]
Length: 20
Categories (2, interval[float64, right]): [(-inf, 0.0] < (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. Aggregation API: agg() and transform()

4. Applying Elementwise Functions: map()

Tablewise function application#

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

First some setup:

In [140]: defextract_city_name(df):
 .....: """
 .....:  Chicago, IL -> Chicago for city_name column
 .....:  """
 .....:  df["city_name"] = df["city_and_code"].str.split(",").str.get(0)
 .....:  return df
 .....: 
In [141]: defadd_country_name(df, country_name=None):
 .....: """
 .....:  Chicago -> Chicago-US for city_name column
 .....:  """
 .....:  col = "city_name"
 .....:  df["city_and_country"] = df[col] + country_name
 .....:  return df
 .....: 
In [142]: df_p = pd.DataFrame({"city_and_code": ["Chicago, IL"]})

extract_city_name and add_country_name are functions taking and returning DataFrames.

Now compare the following:

In [143]: add_country_name(extract_city_name(df_p), country_name="US")
Out[143]: 
 city_and_code city_name city_and_country
0 Chicago, IL Chicago ChicagoUS

Is equivalent to:

In [144]: df_p.pipe(extract_city_name).pipe(add_country_name, country_name="US")
Out[144]: 
 city_and_code city_name city_and_country
0 Chicago, IL Chicago ChicagoUS

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 extract_city_name and add_country_name each expected a 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.ols, 'data') to pipe:

In [147]: importstatsmodels.formula.apiassm
In [148]: bb = pd.read_csv("data/baseball.csv", index_col="id")
In [149]: (
 .....:  bb.query("h > 0")
 .....:  .assign(ln_h=lambda df: np.log(df.h))
 .....:  .pipe((sm.ols, "data"), "hr ~ ln_h + year + g + C(lg)")
 .....:  .fit()
 .....:  .summary()
 .....: )
 .....:
Out[149]:
<class 'statsmodels.iolib.summary.Summary'>
"""
 OLS Regression Results
==============================================================================
Dep. Variable: hr R-squared: 0.685
Model: OLS Adj. R-squared: 0.665
Method: Least Squares F-statistic: 34.28
Date: 2022年11月22日 Prob (F-statistic): 3.48e-15
Time: 05:34:17 Log-Likelihood: -205.92
No. Observations: 68 AIC: 421.8
Df Residuals: 63 BIC: 432.9
Df Model: 4
Covariance Type: nonrobust
===============================================================================
 coef std err t P>|t| [0.025 0.975]
-------------------------------------------------------------------------------
Intercept -8484.7720 4664.146 -1.819 0.074 -1.78e+04 835.780
C(lg)[T.NL] -2.2736 1.325 -1.716 0.091 -4.922 0.375
ln_h -1.3542 0.875 -1.547 0.127 -3.103 0.395
year 4.2277 2.324 1.819 0.074 -0.417 8.872
g 0.1841 0.029 6.258 0.000 0.125 0.243
==============================================================================
Omnibus: 10.875 Durbin-Watson: 1.999
Prob(Omnibus): 0.004 Jarque-Bera (JB): 17.298
Skew: 0.537 Prob(JB): 0.000175
Kurtosis: 5.225 Cond. No. 1.49e+07
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 1.49e+07. This might indicate that there are
strong multicollinearity or other numerical problems.
"""

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 of pipe().

Row or column-wise function application#

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

In [145]: df.apply(lambda x: np.mean(x))
Out[145]: 
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [146]: df.apply(lambda x: np.mean(x), axis=1)
Out[146]: 
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64
In [147]: df.apply(lambda x: x.max() - x.min())
Out[147]: 
one 1.051928
two 1.632779
three 1.840607
dtype: float64
In [148]: df.apply(np.cumsum)
Out[148]: 
 one two three
a 1.394981 1.772517 NaN
b 1.738035 3.684640 -0.050390
c 2.433281 5.163008 1.177045
d NaN 5.442353 0.563873
In [149]: df.apply(np.exp)
Out[149]: 
 one two three
a 4.034899 5.885648 NaN
b 1.409244 6.767440 0.950858
c 2.004201 4.385785 3.412466
d NaN 1.322262 0.541630

The apply() method will also dispatch on a string method name.

In [150]: df.apply("mean")
Out[150]: 
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [151]: df.apply("mean", axis=1)
Out[151]: 
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64

The return type of the function passed to apply() affects the type of the final output from DataFrame.apply for the default behaviour:

• If the applied function returns a Series, the final output is a DataFrame. The columns match the index of the Series returned by the applied function.

• If the applied function returns any other type, the final output is a Series.

This default behaviour can be overridden using the result_type, which accepts three options: reduce, broadcast, and expand. These will determine how list-likes return values expand (or not) to a DataFrame.

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 [152]: tsdf = pd.DataFrame(
 .....:  np.random.randn(1000, 3),
 .....:  columns=["A", "B", "C"],
 .....:  index=pd.date_range("1/1/2000", periods=1000),
 .....: )
 .....: 
In [153]: tsdf.apply(lambda x: x.idxmax())
Out[153]: 
A 2000年08月06日
B 2001年01月18日
C 2001年07月18日
dtype: datetime64[ns]

You may also pass additional arguments and keyword arguments to the apply() method.

In [154]: defsubtract_and_divide(x, sub, divide=1):
 .....:  return (x - sub) / divide
 .....: 
In [155]: df_udf = pd.DataFrame(np.ones((2, 2)))
In [156]: df_udf.apply(subtract_and_divide, args=(5,), divide=3)
Out[156]: 
 0 1
0 -1.333333 -1.333333
1 -1.333333 -1.333333

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

In [157]: tsdf = pd.DataFrame(
 .....:  np.random.randn(10, 3),
 .....:  columns=["A", "B", "C"],
 .....:  index=pd.date_range("1/1/2000", periods=10),
 .....: )
 .....: 
In [158]: tsdf.iloc[3:7] = np.nan
In [159]: tsdf
Out[159]: 
 A B C
2000年01月01日 -0.158131 -0.232466 0.321604
2000年01月02日 -1.810340 -3.105758 0.433834
2000年01月03日 -1.209847 -1.156793 -0.136794
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 -0.653602 0.178875 1.008298
2000年01月09日 1.007996 0.462824 0.254472
2000年01月10日 0.307473 0.600337 1.643950
In [160]: tsdf.apply(pd.Series.interpolate)
Out[160]: 
 A B C
2000年01月01日 -0.158131 -0.232466 0.321604
2000年01月02日 -1.810340 -3.105758 0.433834
2000年01月03日 -1.209847 -1.156793 -0.136794
2000年01月04日 -1.098598 -0.889659 0.092225
2000年01月05日 -0.987349 -0.622526 0.321243
2000年01月06日 -0.876100 -0.355392 0.550262
2000年01月07日 -0.764851 -0.088259 0.779280
2000年01月08日 -0.653602 0.178875 1.008298
2000年01月09日 1.007996 0.462824 0.254472
2000年01月10日 0.307473 0.600337 1.643950

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.

Aggregation API#

The aggregation API allows one to express possibly multiple aggregation operations in a single concise way. This API is similar across pandas objects, see groupby API, the window API, and the resample API. The entry point for aggregation is DataFrame.aggregate(), or the alias DataFrame.agg().

We will use a similar starting frame from above:

In [161]: tsdf = pd.DataFrame(
 .....:  np.random.randn(10, 3),
 .....:  columns=["A", "B", "C"],
 .....:  index=pd.date_range("1/1/2000", periods=10),
 .....: )
 .....: 
In [162]: tsdf.iloc[3:7] = np.nan
In [163]: tsdf
Out[163]: 
 A B C
2000年01月01日 1.257606 1.004194 0.167574
2000年01月02日 -0.749892 0.288112 -0.757304
2000年01月03日 -0.207550 -0.298599 0.116018
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 0.814347 -0.257623 0.869226
2000年01月09日 -0.250663 -1.206601 0.896839
2000年01月10日 2.169758 -1.333363 0.283157

Using a single function is equivalent to apply(). You can also pass named methods as strings. These will return a Series of the aggregated output:

In [164]: tsdf.agg(lambda x: np.sum(x))
Out[164]: 
A 3.033606
B -1.803879
C 1.575510
dtype: float64
In [165]: tsdf.agg("sum")
Out[165]: 
A 3.033606
B -1.803879
C 1.575510
dtype: float64
# these are equivalent to a ``.sum()`` because we are aggregating
# on a single function
In [166]: tsdf.sum()
Out[166]: 
A 3.033606
B -1.803879
C 1.575510
dtype: float64

Single aggregations on a Series this will return a scalar value:

In [167]: tsdf["A"].agg("sum")
Out[167]: 3.033606102414146

Aggregating with multiple functions#

You can pass multiple aggregation arguments as a list. The results of each of the passed functions will be a row in the resulting DataFrame. These are naturally named from the aggregation function.

In [168]: tsdf.agg(["sum"])
Out[168]: 
 A B C
sum 3.033606 -1.803879 1.57551

Multiple functions yield multiple rows:

In [169]: tsdf.agg(["sum", "mean"])
Out[169]: 
 A B C
sum 3.033606 -1.803879 1.575510
mean 0.505601 -0.300647 0.262585

On a Series, multiple functions return a Series, indexed by the function names:

In [170]: tsdf["A"].agg(["sum", "mean"])
Out[170]: 
sum 3.033606
mean 0.505601
Name: A, dtype: float64

Passing a lambda function will yield a <lambda> named row:

In [171]: tsdf["A"].agg(["sum", lambda x: x.mean()])
Out[171]: 
sum 3.033606
<lambda> 0.505601
Name: A, dtype: float64

Passing a named function will yield that name for the row:

In [172]: defmymean(x):
 .....:  return x.mean()
 .....: 
In [173]: tsdf["A"].agg(["sum", mymean])
Out[173]: 
sum 3.033606
mymean 0.505601
Name: A, dtype: float64

Aggregating with a dict#

Passing a dictionary of column names to a scalar or a list of scalars, to DataFrame.agg allows you to customize which functions are applied to which columns. Note that the results are not in any particular order, you can use an OrderedDict instead to guarantee ordering.

In [174]: tsdf.agg({"A": "mean", "B": "sum"})
Out[174]: 
A 0.505601
B -1.803879
dtype: float64

Passing a list-like will generate a DataFrame output. You will get a matrix-like output of all of the aggregators. The output will consist of all unique functions. Those that are not noted for a particular column will be NaN:

In [175]: tsdf.agg({"A": ["mean", "min"], "B": "sum"})
Out[175]: 
 A B
mean 0.505601 NaN
min -0.749892 NaN
sum NaN -1.803879

Custom describe#

With .agg() it is possible to easily create a custom describe function, similar to the built in describe function.

In [176]: fromfunctoolsimport partial
In [177]: q_25 = partial(pd.Series.quantile, q=0.25)
In [178]: q_25.__name__ = "25%"
In [179]: q_75 = partial(pd.Series.quantile, q=0.75)
In [180]: q_75.__name__ = "75%"
In [181]: tsdf.agg(["count", "mean", "std", "min", q_25, "median", q_75, "max"])
Out[181]: 
 A B C
count 6.000000 6.000000 6.000000
mean 0.505601 -0.300647 0.262585
std 1.103362 0.887508 0.606860
min -0.749892 -1.333363 -0.757304
25% -0.239885 -0.979600 0.128907
median 0.303398 -0.278111 0.225365
75% 1.146791 0.151678 0.722709
max 2.169758 1.004194 0.896839

Transform API#

The transform() method returns an object that is indexed the same (same size) as the original. This API allows you to provide multiple operations at the same time rather than one-by-one. Its API is quite similar to the .agg API.

We create a frame similar to the one used in the above sections.

In [182]: tsdf = pd.DataFrame(
 .....:  np.random.randn(10, 3),
 .....:  columns=["A", "B", "C"],
 .....:  index=pd.date_range("1/1/2000", periods=10),
 .....: )
 .....: 
In [183]: tsdf.iloc[3:7] = np.nan
In [184]: tsdf
Out[184]: 
 A B C
2000年01月01日 -0.428759 -0.864890 -0.675341
2000年01月02日 -0.168731 1.338144 -1.279321
2000年01月03日 -1.621034 0.438107 0.903794
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 0.254374 -1.240447 -0.201052
2000年01月09日 -0.157795 0.791197 -1.144209
2000年01月10日 -0.030876 0.371900 0.061932

Transform the entire frame. .transform() allows input functions as: a NumPy function, a string function name or a user defined function.

In [185]: tsdf.transform(np.abs)
Out[185]: 
 A B C
2000年01月01日 0.428759 0.864890 0.675341
2000年01月02日 0.168731 1.338144 1.279321
2000年01月03日 1.621034 0.438107 0.903794
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 0.254374 1.240447 0.201052
2000年01月09日 0.157795 0.791197 1.144209
2000年01月10日 0.030876 0.371900 0.061932
In [186]: tsdf.transform("abs")
Out[186]: 
 A B C
2000年01月01日 0.428759 0.864890 0.675341
2000年01月02日 0.168731 1.338144 1.279321
2000年01月03日 1.621034 0.438107 0.903794
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 0.254374 1.240447 0.201052
2000年01月09日 0.157795 0.791197 1.144209
2000年01月10日 0.030876 0.371900 0.061932
In [187]: tsdf.transform(lambda x: x.abs())
Out[187]: 
 A B C
2000年01月01日 0.428759 0.864890 0.675341
2000年01月02日 0.168731 1.338144 1.279321
2000年01月03日 1.621034 0.438107 0.903794
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 0.254374 1.240447 0.201052
2000年01月09日 0.157795 0.791197 1.144209
2000年01月10日 0.030876 0.371900 0.061932

Here transform() received a single function; this is equivalent to a ufunc application.

In [188]: np.abs(tsdf)
Out[188]: 
 A B C
2000年01月01日 0.428759 0.864890 0.675341
2000年01月02日 0.168731 1.338144 1.279321
2000年01月03日 1.621034 0.438107 0.903794
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 0.254374 1.240447 0.201052
2000年01月09日 0.157795 0.791197 1.144209
2000年01月10日 0.030876 0.371900 0.061932

Passing a single function to .transform() with a Series will yield a single Series in return.

In [189]: tsdf["A"].transform(np.abs)
Out[189]: 
2000年01月01日 0.428759
2000年01月02日 0.168731
2000年01月03日 1.621034
2000年01月04日 NaN
2000年01月05日 NaN
2000年01月06日 NaN
2000年01月07日 NaN
2000年01月08日 0.254374
2000年01月09日 0.157795
2000年01月10日 0.030876
Freq: D, Name: A, dtype: float64

Transform with multiple functions#

Passing multiple functions will yield a column MultiIndexed DataFrame. The first level will be the original frame column names; the second level will be the names of the transforming functions.

In [190]: tsdf.transform([np.abs, lambda x: x + 1])
Out[190]: 
 A B C 
 absolute <lambda> absolute <lambda> absolute <lambda>
2000年01月01日 0.428759 0.571241 0.864890 0.135110 0.675341 0.324659
2000年01月02日 0.168731 0.831269 1.338144 2.338144 1.279321 -0.279321
2000年01月03日 1.621034 -0.621034 0.438107 1.438107 0.903794 1.903794
2000年01月04日 NaN NaN NaN NaN NaN NaN
2000年01月05日 NaN NaN NaN NaN NaN NaN
2000年01月06日 NaN NaN NaN NaN NaN NaN
2000年01月07日 NaN NaN NaN NaN NaN NaN
2000年01月08日 0.254374 1.254374 1.240447 -0.240447 0.201052 0.798948
2000年01月09日 0.157795 0.842205 0.791197 1.791197 1.144209 -0.144209
2000年01月10日 0.030876 0.969124 0.371900 1.371900 0.061932 1.061932

Passing multiple functions to a Series will yield a DataFrame. The resulting column names will be the transforming functions.

In [191]: tsdf["A"].transform([np.abs, lambda x: x + 1])
Out[191]: 
 absolute <lambda>
2000年01月01日 0.428759 0.571241
2000年01月02日 0.168731 0.831269
2000年01月03日 1.621034 -0.621034
2000年01月04日 NaN NaN
2000年01月05日 NaN NaN
2000年01月06日 NaN NaN
2000年01月07日 NaN NaN
2000年01月08日 0.254374 1.254374
2000年01月09日 0.157795 0.842205
2000年01月10日 0.030876 0.969124

Transforming with a dict#

Passing a dict of functions will allow selective transforming per column.

In [192]: tsdf.transform({"A": np.abs, "B": lambda x: x + 1})
Out[192]: 
 A B
2000年01月01日 0.428759 0.135110
2000年01月02日 0.168731 2.338144
2000年01月03日 1.621034 1.438107
2000年01月04日 NaN NaN
2000年01月05日 NaN NaN
2000年01月06日 NaN NaN
2000年01月07日 NaN NaN
2000年01月08日 0.254374 -0.240447
2000年01月09日 0.157795 1.791197
2000年01月10日 0.030876 1.371900

Passing a dict of lists will generate a MultiIndexed DataFrame with these selective transforms.

In [193]: tsdf.transform({"A": np.abs, "B": [lambda x: x + 1, "sqrt"]})
Out[193]: 
 A B 
 absolute <lambda> sqrt
2000年01月01日 0.428759 0.135110 NaN
2000年01月02日 0.168731 2.338144 1.156782
2000年01月03日 1.621034 1.438107 0.661897
2000年01月04日 NaN NaN NaN
2000年01月05日 NaN NaN NaN
2000年01月06日 NaN NaN NaN
2000年01月07日 NaN NaN NaN
2000年01月08日 0.254374 -0.240447 NaN
2000年01月09日 0.157795 1.791197 0.889493
2000年01月10日 0.030876 1.371900 0.609836

Applying elementwise functions#

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

In [194]: df4 = df.copy()
In [195]: df4
Out[195]: 
 one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [196]: deff(x):
 .....:  return len(str(x))
 .....: 
In [197]: df4["one"].map(f)
Out[197]: 
a 18
b 19
c 18
d 3
Name: one, dtype: int64
In [198]: df4.map(f)
Out[198]: 
 one two three
a 18 17 3
b 19 18 20
c 18 18 16
d 3 19 19

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

In [199]: s = pd.Series(
 .....:  ["six", "seven", "six", "seven", "six"], index=["a", "b", "c", "d", "e"]
 .....: )
 .....: 
In [200]: t = pd.Series({"six": 6.0, "seven": 7.0})
In [201]: s
Out[201]: 
a six
b seven
c six
d seven
e six
dtype: object
In [202]: s.map(t)
Out[202]: 
a 6.0
b 7.0
c 6.0
d 7.0
e 6.0
dtype: float64

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 [203]: s = pd.Series(np.random.randn(5), index=["a", "b", "c", "d", "e"])
In [204]: s
Out[204]: 
a 1.695148
b 1.328614
c 1.234686
d -0.385845
e -1.326508
dtype: float64
In [205]: s.reindex(["e", "b", "f", "d"])
Out[205]: 
e -1.326508
b 1.328614
f NaN
d -0.385845
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 [206]: df
Out[206]: 
 one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [207]: df.reindex(index=["c", "f", "b"], columns=["three", "two", "one"])
Out[207]: 
 three two one
c 1.227435 1.478369 0.695246
f NaN NaN NaN
b -0.050390 1.912123 0.343054

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 [208]: rs = s.reindex(df.index)
In [209]: rs
Out[209]: 
a 1.695148
b 1.328614
c 1.234686
d -0.385845
dtype: float64
In [210]: rs.index is df.index
Out[210]: True

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

DataFrame.reindex() also supports an "axis-style" calling convention, where you specify a single labels argument and the axis it applies to.

In [211]: df.reindex(["c", "f", "b"], axis="index")
Out[211]: 
 one two three
c 0.695246 1.478369 1.227435
f NaN NaN NaN
b 0.343054 1.912123 -0.050390
In [212]: df.reindex(["three", "two", "one"], axis="columns")
Out[212]: 
 three two one
a NaN 1.772517 1.394981
b -0.050390 1.912123 0.343054
c 1.227435 1.478369 0.695246
d -0.613172 0.279344 NaN

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 [213]: df2 = df.reindex(["a", "b", "c"], columns=["one", "two"])
In [214]: df3 = df2 - df2.mean()
In [215]: df2
Out[215]: 
 one two
a 1.394981 1.772517
b 0.343054 1.912123
c 0.695246 1.478369
In [216]: df3
Out[216]: 
 one two
a 0.583888 0.051514
b -0.468040 0.191120
c -0.115848 -0.242634
In [217]: df.reindex_like(df2)
Out[217]: 
 one two
a 1.394981 1.772517
b 0.343054 1.912123
c 0.695246 1.478369

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 [218]: s = pd.Series(np.random.randn(5), index=["a", "b", "c", "d", "e"])
In [219]: s1 = s[:4]
In [220]: s2 = s[1:]
In [221]: s1.align(s2)
Out[221]: 
(a -0.186646
 b -1.692424
 c -0.303893
 d -1.425662
 e NaN
 dtype: float64,
 a NaN
 b -1.692424
 c -0.303893
 d -1.425662
 e 1.114285
 dtype: float64)
In [222]: s1.align(s2, join="inner")
Out[222]: 
(b -1.692424
 c -0.303893
 d -1.425662
 dtype: float64,
 b -1.692424
 c -0.303893
 d -1.425662
 dtype: float64)
In [223]: s1.align(s2, join="left")
Out[223]: 
(a -0.186646
 b -1.692424
 c -0.303893
 d -1.425662
 dtype: float64,
 a NaN
 b -1.692424
 c -0.303893
 d -1.425662
 dtype: float64)

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

In [224]: df.align(df2, join="inner")
Out[224]: 
( one two
 a 1.394981 1.772517
 b 0.343054 1.912123
 c 0.695246 1.478369,
 one two
 a 1.394981 1.772517
 b 0.343054 1.912123
 c 0.695246 1.478369)

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

In [225]: df.align(df2, join="inner", axis=0)
Out[225]: 
( one two three
 a 1.394981 1.772517 NaN
 b 0.343054 1.912123 -0.050390
 c 0.695246 1.478369 1.227435,
 one two
 a 1.394981 1.772517
 b 0.343054 1.912123
 c 0.695246 1.478369)

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 [226]: df.align(df2.iloc[0], axis=1)
Out[226]: 
( one three two
 a 1.394981 NaN 1.772517
 b 0.343054 -0.050390 1.912123
 c 0.695246 1.227435 1.478369
 d NaN -0.613172 0.279344,
 one 1.394981
 three NaN
 two 1.772517
 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 [227]: rng = pd.date_range("1/3/2000", periods=8)
In [228]: ts = pd.Series(np.random.randn(8), index=rng)
In [229]: ts2 = ts.iloc[[0, 3, 6]]
In [230]: ts
Out[230]: 
2000年01月03日 0.183051
2000年01月04日 0.400528
2000年01月05日 -0.015083
2000年01月06日 2.395489
2000年01月07日 1.414806
2000年01月08日 0.118428
2000年01月09日 0.733639
2000年01月10日 -0.936077
Freq: D, dtype: float64
In [231]: ts2
Out[231]: 
2000年01月03日 0.183051
2000年01月06日 2.395489
2000年01月09日 0.733639
Freq: 3D, dtype: float64
In [232]: ts2.reindex(ts.index)
Out[232]: 
2000年01月03日 0.183051
2000年01月04日 NaN
2000年01月05日 NaN
2000年01月06日 2.395489
2000年01月07日 NaN
2000年01月08日 NaN
2000年01月09日 0.733639
2000年01月10日 NaN
Freq: D, dtype: float64
In [233]: ts2.reindex(ts.index, method="ffill")
Out[233]: 
2000年01月03日 0.183051
2000年01月04日 0.183051
2000年01月05日 0.183051
2000年01月06日 2.395489
2000年01月07日 2.395489
2000年01月08日 2.395489
2000年01月09日 0.733639
2000年01月10日 0.733639
Freq: D, dtype: float64
In [234]: ts2.reindex(ts.index, method="bfill")
Out[234]: 
2000年01月03日 0.183051
2000年01月04日 2.395489
2000年01月05日 2.395489
2000年01月06日 2.395489
2000年01月07日 0.733639
2000年01月08日 0.733639
2000年01月09日 0.733639
2000年01月10日 NaN
Freq: D, dtype: float64
In [235]: ts2.reindex(ts.index, method="nearest")
Out[235]: 
2000年01月03日 0.183051
2000年01月04日 0.183051
2000年01月05日 2.395489
2000年01月06日 2.395489
2000年01月07日 2.395489
2000年01月08日 0.733639
2000年01月09日 0.733639
2000年01月10日 0.733639
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 ffill (except for method='nearest') or interpolate:

In [236]: ts2.reindex(ts.index).ffill()
Out[236]: 
2000年01月03日 0.183051
2000年01月04日 0.183051
2000年01月05日 0.183051
2000年01月06日 2.395489
2000年01月07日 2.395489
2000年01月08日 2.395489
2000年01月09日 0.733639
2000年01月10日 0.733639
Freq: D, dtype: float64

reindex() will raise a ValueError if the index is not monotonically increasing or decreasing. fillna() and interpolate() will not perform 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 [237]: ts2.reindex(ts.index, method="ffill", limit=1)
Out[237]: 
2000年01月03日 0.183051
2000年01月04日 0.183051
2000年01月05日 NaN
2000年01月06日 2.395489
2000年01月07日 2.395489
2000年01月08日 NaN
2000年01月09日 0.733639
2000年01月10日 0.733639
Freq: D, dtype: float64

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

In [238]: ts2.reindex(ts.index, method="ffill", tolerance="1 day")
Out[238]: 
2000年01月03日 0.183051
2000年01月04日 0.183051
2000年01月05日 NaN
2000年01月06日 2.395489
2000年01月07日 2.395489
2000年01月08日 NaN
2000年01月09日 0.733639
2000年01月10日 0.733639
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 [239]: df
Out[239]: 
 one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [240]: df.drop(["a", "d"], axis=0)
Out[240]: 
 one two three
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
In [241]: df.drop(["one"], axis=1)
Out[241]: 
 two three
a 1.772517 NaN
b 1.912123 -0.050390
c 1.478369 1.227435
d 0.279344 -0.613172

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

In [242]: df.reindex(df.index.difference(["a", "d"]))
Out[242]: 
 one two three
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435

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 [243]: s
Out[243]: 
a -0.186646
b -1.692424
c -0.303893
d -1.425662
e 1.114285
dtype: float64
In [244]: s.rename(str.upper)
Out[244]: 
A -0.186646
B -1.692424
C -0.303893
D -1.425662
E 1.114285
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). A dict or Series can also be used:

In [245]: df.rename(
 .....:  columns={"one": "foo", "two": "bar"},
 .....:  index={"a": "apple", "b": "banana", "d": "durian"},
 .....: )
 .....: 
Out[245]: 
 foo bar three
apple 1.394981 1.772517 NaN
banana 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
durian NaN 0.279344 -0.613172

If the mapping doesn’t include a column/index label, it isn’t renamed. Note that extra labels in the mapping don’t throw an error.

DataFrame.rename() also supports an "axis-style" calling convention, where you specify a single mapper and the axis to apply that mapping to.

In [246]: df.rename({"one": "foo", "two": "bar"}, axis="columns")
Out[246]: 
 foo bar three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [247]: df.rename({"a": "apple", "b": "banana", "d": "durian"}, axis="index")
Out[247]: 
 one two three
apple 1.394981 1.772517 NaN
banana 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
durian NaN 0.279344 -0.613172

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

In [248]: s.rename("scalar-name")
Out[248]: 
a -0.186646
b -1.692424
c -0.303893
d -1.425662
e 1.114285
Name: scalar-name, dtype: float64

The methods DataFrame.rename_axis() and Series.rename_axis() allow specific names of a MultiIndex to be changed (as opposed to the labels).

In [249]: df = pd.DataFrame(
 .....:  {"x": [1, 2, 3, 4, 5, 6], "y": [10, 20, 30, 40, 50, 60]},
 .....:  index=pd.MultiIndex.from_product(
 .....:  [["a", "b", "c"], [1, 2]], names=["let", "num"]
 .....:  ),
 .....: )
 .....: 
In [250]: df
Out[250]: 
 x y
let num 
a 1 1 10
 2 2 20
b 1 3 30
 2 4 40
c 1 5 50
 2 6 60
In [251]: df.rename_axis(index={"let": "abc"})
Out[251]: 
 x y
abc num 
a 1 1 10
 2 2 20
b 1 3 30
 2 4 40
c 1 5 50
 2 6 60
In [252]: df.rename_axis(index=str.upper)
Out[252]: 
 x y
LET NUM 
a 1 1 10
 2 2 20
b 1 3 30
 2 4 40
c 1 5 50
 2 6 60

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. DataFrames 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

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

In [253]: df = pd.DataFrame(
 .....:  {"col1": np.random.randn(3), "col2": np.random.randn(3)}, index=["a", "b", "c"]
 .....: )
 .....: 
In [254]: for col in df:
 .....:  print(col)
 .....: 
col1
col2

pandas objects also have the dict-like items() 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 with 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 [255]: df = pd.DataFrame({"a": [1, 2, 3], "b": ["a", "b", "c"]})
In [256]: for index, row in df.iterrows():
 .....:  row["a"] = 10
 .....: 
In [257]: df
Out[257]: 
 a b
0 1 a
1 2 b
2 3 c

items#

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

• Series: (index, scalar value) pairs

• DataFrame: (column, Series) pairs

For example:

In [258]: for label, ser in df.items():
 .....:  print(label)
 .....:  print(ser)
 .....: 
a
0 1
1 2
2 3
Name: a, dtype: int64
b
0 a
1 b
2 c
Name: b, dtype: object

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 [259]: for row_index, row in df.iterrows():
 .....:  print(row_index, row, sep="\n")
 .....: 
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 [260]: df_orig = pd.DataFrame([[1, 1.5]], columns=["int", "float"])
In [261]: df_orig.dtypes
Out[261]: 
int int64
float float64
dtype: object
In [262]: row = next(df_orig.iterrows())[1]
In [263]: row
Out[263]: 
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 [264]: row["int"].dtype
Out[264]: dtype('float64')
In [265]: df_orig["int"].dtype
Out[265]: 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 than iterrows().

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

In [266]: df2 = pd.DataFrame({"x": [1, 2, 3], "y": [4, 5, 6]})
In [267]: print(df2)
 x y
0 1 4
1 2 5
2 3 6
In [268]: print(df2.T)
 0 1 2
x 1 2 3
y 4 5 6
In [269]: df2_t = pd.DataFrame({idx: values for idx, values in df2.iterrows()})
In [270]: 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 [271]: 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; it merely 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 [272]: s = pd.Series(pd.date_range("20130101 09:10:12", periods=4))
In [273]: s
Out[273]: 
0 2013年01月01日 09:10:12
1 2013年01月02日 09:10:12
2 2013年01月03日 09:10:12
3 2013年01月04日 09:10:12
dtype: datetime64[ns]
In [274]: s.dt.hour
Out[274]: 
0 9
1 9
2 9
3 9
dtype: int32
In [275]: s.dt.second
Out[275]: 
0 12
1 12
2 12
3 12
dtype: int32
In [276]: s.dt.day
Out[276]: 
0 1
1 2
2 3
3 4
dtype: int32

This enables nice expressions like this:

In [277]: s[s.dt.day == 2]
Out[277]: 
1 2013年01月02日 09:10:12
dtype: datetime64[ns]

You can easily produces tz aware transformations:

In [278]: stz = s.dt.tz_localize("US/Eastern")
In [279]: stz
Out[279]: 
0 2013年01月01日 09:10:12-05:00
1 2013年01月02日 09:10:12-05:00
2 2013年01月03日 09:10:12-05:00
3 2013年01月04日 09:10:12-05:00
dtype: datetime64[ns, US/Eastern]
In [280]: stz.dt.tz
Out[280]: <DstTzInfo 'US/Eastern' LMT-1 day, 19:04:00 STD>

You can also chain these types of operations:

In [281]: s.dt.tz_localize("UTC").dt.tz_convert("US/Eastern")
Out[281]: 
0 2013年01月01日 04:10:12-05:00
1 2013年01月02日 04:10:12-05:00
2 2013年01月03日 04:10:12-05:00
3 2013年01月04日 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 [282]: s = pd.Series(pd.date_range("20130101", periods=4))
In [283]: s
Out[283]: 
0 2013年01月01日
1 2013年01月02日
2 2013年01月03日
3 2013年01月04日
dtype: datetime64[ns]
In [284]: s.dt.strftime("%Y/%m/%d")
Out[284]: 
0 2013年01月01日
1 2013年01月02日
2 2013年01月03日
3 2013年01月04日
dtype: object

# PeriodIndex
In [285]: s = pd.Series(pd.period_range("20130101", periods=4))
In [286]: s
Out[286]: 
0 2013年01月01日
1 2013年01月02日
2 2013年01月03日
3 2013年01月04日
dtype: period[D]
In [287]: s.dt.strftime("%Y/%m/%d")
Out[287]: 
0 2013年01月01日
1 2013年01月02日
2 2013年01月03日
3 2013年01月04日
dtype: object

The .dt accessor works for period and timedelta dtypes.

# period
In [288]: s = pd.Series(pd.period_range("20130101", periods=4, freq="D"))
In [289]: s
Out[289]: 
0 2013年01月01日
1 2013年01月02日
2 2013年01月03日
3 2013年01月04日
dtype: period[D]
In [290]: s.dt.year
Out[290]: 
0 2013
1 2013
2 2013
3 2013
dtype: int64
In [291]: s.dt.day
Out[291]: 
0 1
1 2
2 3
3 4
dtype: int64

# timedelta
In [292]: s = pd.Series(pd.timedelta_range("1 day 00:00:05", periods=4, freq="s"))
In [293]: s
Out[293]: 
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 [294]: s.dt.days
Out[294]: 
0 1
1 1
2 1
3 1
dtype: int64
In [295]: s.dt.seconds
Out[295]: 
0 5
1 6
2 7
3 8
dtype: int32
In [296]: s.dt.components
Out[296]: 
 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-datetime-like 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 [297]: s = pd.Series(
 .....:  ["A", "B", "C", "Aaba", "Baca", np.nan, "CABA", "dog", "cat"], dtype="string"
 .....: )
 .....: 
In [298]: s.str.lower()
Out[298]: 
0 a
1 b
2 c
3 aaba
4 baca
5 <NA>
6 caba
7 dog
8 cat
dtype: string

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).

Note

Prior to pandas 1.0, string methods were only available on object -dtype Series. pandas 1.0 added the StringDtype which is dedicated to strings. See Text data types for more.

Please see Vectorized String Methods for a complete description.

Sorting#

pandas supports three kinds of sorting: sorting by index labels, sorting by column values, and sorting by a combination of both.

By index#

The Series.sort_index() and DataFrame.sort_index() methods are used to sort a pandas object by its index levels.

In [299]: 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 [300]: unsorted_df = df.reindex(
 .....:  index=["a", "d", "c", "b"], columns=["three", "two", "one"]
 .....: )
 .....: 
In [301]: unsorted_df
Out[301]: 
 three two one
a NaN -1.152244 0.562973
d -0.252916 -0.109597 NaN
c 1.273388 -0.167123 0.640382
b -0.098217 0.009797 -1.299504
# DataFrame
In [302]: unsorted_df.sort_index()
Out[302]: 
 three two one
a NaN -1.152244 0.562973
b -0.098217 0.009797 -1.299504
c 1.273388 -0.167123 0.640382
d -0.252916 -0.109597 NaN
In [303]: unsorted_df.sort_index(ascending=False)
Out[303]: 
 three two one
d -0.252916 -0.109597 NaN
c 1.273388 -0.167123 0.640382
b -0.098217 0.009797 -1.299504
a NaN -1.152244 0.562973
In [304]: unsorted_df.sort_index(axis=1)
Out[304]: 
 one three two
a 0.562973 NaN -1.152244
d NaN -0.252916 -0.109597
c 0.640382 1.273388 -0.167123
b -1.299504 -0.098217 0.009797
# Series
In [305]: unsorted_df["three"].sort_index()
Out[305]: 
a NaN
b -0.098217
c 1.273388
d -0.252916
Name: three, dtype: float64

Sorting by index also supports a key parameter that takes a callable function to apply to the index being sorted. For MultiIndex objects, the key is applied per-level to the levels specified by level.

In [306]: s1 = pd.DataFrame({"a": ["B", "a", "C"], "b": [1, 2, 3], "c": [2, 3, 4]}).set_index(
 .....:  list("ab")
 .....: )
 .....: 
In [307]: s1
Out[307]: 
 c
a b 
B 1 2
a 2 3
C 3 4

In [308]: s1.sort_index(level="a")
Out[308]: 
 c
a b 
B 1 2
C 3 4
a 2 3
In [309]: s1.sort_index(level="a", key=lambda idx: idx.str.lower())
Out[309]: 
 c
a b 
a 2 3
B 1 2
C 3 4

For information on key sorting by value, see value sorting.

By values#

The Series.sort_values() method is used to sort a Series by its values. The DataFrame.sort_values() method is used to sort a DataFrame by its column or row values. The optional by parameter to DataFrame.sort_values() may used to specify one or more columns to use to determine the sorted order.

In [310]: df1 = pd.DataFrame(
 .....:  {"one": [2, 1, 1, 1], "two": [1, 3, 2, 4], "three": [5, 4, 3, 2]}
 .....: )
 .....: 
In [311]: df1.sort_values(by="two")
Out[311]: 
 one two three
0 2 1 5
2 1 2 3
1 1 3 4
3 1 4 2

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

In [312]: df1[["one", "two", "three"]].sort_values(by=["one", "two"])
Out[312]: 
 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 [313]: s[2] = np.nan
In [314]: s.sort_values()
Out[314]: 
0 A
3 Aaba
1 B
4 Baca
6 CABA
8 cat
7 dog
2 <NA>
5 <NA>
dtype: string
In [315]: s.sort_values(na_position="first")
Out[315]: 
2 <NA>
5 <NA>
0 A
3 Aaba
1 B
4 Baca
6 CABA
8 cat
7 dog
dtype: string

Sorting also supports a key parameter that takes a callable function to apply to the values being sorted.

In [316]: s1 = pd.Series(["B", "a", "C"])

In [317]: s1.sort_values()
Out[317]: 
0 B
2 C
1 a
dtype: object
In [318]: s1.sort_values(key=lambda x: x.str.lower())
Out[318]: 
1 a
0 B
2 C
dtype: object

key will be given the Series of values and should return a Series or array of the same shape with the transformed values. For DataFrame objects, the key is applied per column, so the key should still expect a Series and return a Series, e.g.

In [319]: df = pd.DataFrame({"a": ["B", "a", "C"], "b": [1, 2, 3]})

In [320]: df.sort_values(by="a")
Out[320]: 
 a b
0 B 1
2 C 3
1 a 2
In [321]: df.sort_values(by="a", key=lambda col: col.str.lower())
Out[321]: 
 a b
1 a 2
0 B 1
2 C 3

The name or type of each column can be used to apply different functions to different columns.

By indexes and values#

Strings passed as the by parameter to DataFrame.sort_values() may refer to either columns or index level names.

# Build MultiIndex
In [322]: idx = pd.MultiIndex.from_tuples(
 .....:  [("a", 1), ("a", 2), ("a", 2), ("b", 2), ("b", 1), ("b", 1)]
 .....: )
 .....: 
In [323]: idx.names = ["first", "second"]
# Build DataFrame
In [324]: df_multi = pd.DataFrame({"A": np.arange(6, 0, -1)}, index=idx)
In [325]: df_multi
Out[325]: 
 A
first second 
a 1 6
 2 5
 2 4
b 2 3
 1 2
 1 1

Sort by ‘second’ (index) and ‘A’ (column)

In [326]: df_multi.sort_values(by=["second", "A"])
Out[326]: 
 A
first second 
b 1 1
 1 2
a 1 6
b 2 3
a 2 4
 2 5

Note

If a string matches both a column name and an index level name then a warning is issued and the column takes precedence. This will result in an ambiguity error in a future version.

searchsorted#

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

In [327]: ser = pd.Series([1, 2, 3])
In [328]: ser.searchsorted([0, 3])
Out[328]: array([0, 2])
In [329]: ser.searchsorted([0, 4])
Out[329]: array([0, 3])
In [330]: ser.searchsorted([1, 3], side="right")
Out[330]: array([1, 3])
In [331]: ser.searchsorted([1, 3], side="left")
Out[331]: array([0, 2])
In [332]: ser = pd.Series([3, 1, 2])
In [333]: ser.searchsorted([0, 3], sorter=np.argsort(ser))
Out[333]: array([0, 2])

smallest / largest values#

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 [334]: s = pd.Series(np.random.permutation(10))
In [335]: s
Out[335]: 
0 2
1 0
2 3
3 7
4 1
5 5
6 9
7 6
8 8
9 4
dtype: int64
In [336]: s.sort_values()
Out[336]: 
1 0
4 1
0 2
2 3
9 4
5 5
7 6
3 7
8 8
6 9
dtype: int64
In [337]: s.nsmallest(3)
Out[337]: 
1 0
4 1
0 2
dtype: int64
In [338]: s.nlargest(3)
Out[338]: 
6 9
8 8
3 7
dtype: int64

DataFrame also has the nlargest and nsmallest methods.

In [339]: 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 [340]: df.nlargest(3, "a")
Out[340]: 
 a b c
5 11 f 3.0
3 10 c 3.2
4 8 e NaN
In [341]: df.nlargest(5, ["a", "c"])
Out[341]: 
 a b c
5 11 f 3.0
3 10 c 3.2
4 8 e NaN
2 1 d 4.0
6 -1 f 4.0
In [342]: df.nsmallest(3, "a")
Out[342]: 
 a b c
0 -2 a 1.0
1 -1 b 2.0
6 -1 f 4.0
In [343]: df.nsmallest(5, ["a", "c"])
Out[343]: 
 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 MultiIndex column#

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

In [344]: df1.columns = pd.MultiIndex.from_tuples(
 .....:  [("a", "one"), ("a", "two"), ("b", "three")]
 .....: )
 .....: 
In [345]: df1.sort_values(by=("a", "two"))
Out[345]: 
 a b
 one two three
0 2 1 5
2 1 2 3
1 1 3 4
3 1 4 2

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 method has the side effect of modifying your data; almost every method returns a new object, leaving the original object untouched. If the data is modified, it is because you did so explicitly.

dtypes#

For the most part, pandas uses NumPy arrays and dtypes for Series or individual columns of a DataFrame. NumPy provides support for float, int, bool, timedelta64[ns] and datetime64[ns] (note that NumPy does not support timezone-aware datetimes).

pandas and third-party libraries extend NumPy’s type system in a few places. This section describes the extensions pandas has made internally. See Extension types for how to write your own extension that works with pandas. See the ecosystem page for a list of third-party libraries that have implemented an extension.

The following table lists all of pandas extension types. For methods requiring dtype arguments, strings can be specified as indicated. See the respective documentation sections for more on each type.

Kind of Data	Data Type		Scalar		Array	String Aliases
tz-aware datetime	DatetimeTZDtype		Timestamp		arrays.DatetimeArray	'datetime64[ns, <tz>]'
Categorical	CategoricalDtype		(none)		Categorical	'category'
period (time spans)	PeriodDtype		Period		arrays.PeriodArray 'Period[<freq>]'	'period[<freq>]',
sparse	SparseDtype		(none)		arrays.SparseArray	'Sparse', 'Sparse[int]', 'Sparse[float]'
intervals	IntervalDtype		Interval		arrays.IntervalArray	'interval', 'Interval', 'Interval[<numpy_dtype>]', 'Interval[datetime64[ns, <tz>]]', 'Interval[timedelta64[<freq>]]'
nullable integer	Int64Dtype, ...		(none)		arrays.IntegerArray	'Int8', 'Int16', 'Int32', 'Int64', 'UInt8', 'UInt16', 'UInt32', 'UInt64'
nullable float	Float64Dtype, ...		(none)		arrays.FloatingArray	'Float32', 'Float64'
Strings	StringDtype		str		arrays.StringArray	'string'
Boolean (with NA)	BooleanDtype		bool		arrays.BooleanArray	'boolean'

pandas has two ways to store strings.

1. object dtype, which can hold any Python object, including strings.

2. StringDtype, which is dedicated to strings.

Generally, we recommend using StringDtype. See Text data types for more.

Finally, arbitrary objects may be stored using the object dtype, but should be avoided to the extent possible (for performance and interoperability with other libraries and methods. See object conversion).

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

In [346]: dft = pd.DataFrame(
 .....:  {
 .....:  "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 [347]: dft
Out[347]: 
 A B C D E F G
0 0.035962 1 foo 2001年01月02日 1.0 False 1
1 0.701379 1 foo 2001年01月02日 1.0 False 1
2 0.281885 1 foo 2001年01月02日 1.0 False 1
In [348]: dft.dtypes
Out[348]: 
A float64
B int64
C object
D datetime64[s]
E float32
F bool
G int8
dtype: object

On a Series object, use the dtype attribute.

In [349]: dft["A"].dtype
Out[349]: dtype('float64')

If a pandas object contains data with 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 [350]: pd.Series([1, 2, 3, 4, 5, 6.0])
Out[350]: 
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 [351]: pd.Series([1, 2, 3, 6.0, "foo"])
Out[351]: 
0 1
1 2
2 3
3 6.0
4 foo
dtype: object

The number of columns of each type in a DataFrame can be found by calling DataFrame.dtypes.value_counts().

In [352]: dft.dtypes.value_counts()
Out[352]: 
float64 1
int64 1
object 1
datetime64[s] 1
float32 1
bool 1
int8 1
Name: count, dtype: int64

Numeric dtypes will propagate and can coexist in DataFrames. 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 [353]: df1 = pd.DataFrame(np.random.randn(8, 1), columns=["A"], dtype="float32")
In [354]: df1
Out[354]: 
 A
0 0.224364
1 1.890546
2 0.182879
3 0.787847
4 -0.188449
5 0.667715
6 -0.011736
7 -0.399073
In [355]: df1.dtypes
Out[355]: 
A float32
dtype: object
In [356]: df2 = pd.DataFrame(
 .....:  {
 .....:  "A": pd.Series(np.random.randn(8), dtype="float16"),
 .....:  "B": pd.Series(np.random.randn(8)),
 .....:  "C": pd.Series(np.random.randint(0, 255, size=8), dtype="uint8"), # [0,255] (range of uint8)
 .....:  }
 .....: )
 .....: 
In [357]: df2
Out[357]: 
 A B C
0 0.823242 0.256090 26
1 1.607422 1.426469 86
2 -0.333740 -0.416203 46
3 -0.063477 1.139976 212
4 -1.014648 -1.193477 26
5 0.678711 0.096706 7
6 -0.040863 -1.956850 184
7 -0.357422 -0.714337 206
In [358]: df2.dtypes
Out[358]: 
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 [359]: pd.DataFrame([1, 2], columns=["a"]).dtypes
Out[359]: 
a int64
dtype: object
In [360]: pd.DataFrame({"a": [1, 2]}).dtypes
Out[360]: 
a int64
dtype: object
In [361]: pd.DataFrame({"a": 1}, index=list(range(2))).dtypes
Out[361]: 
a int64
dtype: object

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

In [362]: 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 (e.g. int to float).

In [363]: df3 = df1.reindex_like(df2).fillna(value=0.0) + df2
In [364]: df3
Out[364]: 
 A B C
0 1.047606 0.256090 26.0
1 3.497968 1.426469 86.0
2 -0.150862 -0.416203 46.0
3 0.724370 1.139976 212.0
4 -1.203098 -1.193477 26.0
5 1.346426 0.096706 7.0
6 -0.052599 -1.956850 184.0
7 -0.756495 -0.714337 206.0
In [365]: df3.dtypes
Out[365]: 
A float32
B float64
C float64
dtype: object

DataFrame.to_numpy() will 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 [366]: df3.to_numpy().dtype
Out[366]: 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 [367]: df3
Out[367]: 
 A B C
0 1.047606 0.256090 26.0
1 3.497968 1.426469 86.0
2 -0.150862 -0.416203 46.0
3 0.724370 1.139976 212.0
4 -1.203098 -1.193477 26.0
5 1.346426 0.096706 7.0
6 -0.052599 -1.956850 184.0
7 -0.756495 -0.714337 206.0
In [368]: df3.dtypes
Out[368]: 
A float32
B float64
C float64
dtype: object
# conversion of dtypes
In [369]: df3.astype("float32").dtypes
Out[369]: 
A float32
B float32
C float32
dtype: object

Convert a subset of columns to a specified type using astype().

In [370]: dft = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6], "c": [7, 8, 9]})
In [371]: dft[["a", "b"]] = dft[["a", "b"]].astype(np.uint8)
In [372]: dft
Out[372]: 
 a b c
0 1 4 7
1 2 5 8
2 3 6 9
In [373]: dft.dtypes
Out[373]: 
a uint8
b uint8
c int64
dtype: object

Convert certain columns to a specific dtype by passing a dict to astype().

In [374]: dft1 = pd.DataFrame({"a": [1, 0, 1], "b": [4, 5, 6], "c": [7, 8, 9]})
In [375]: dft1 = dft1.astype({"a": np.bool_, "c": np.float64})
In [376]: dft1
Out[376]: 
 a b c
0 True 4 7.0
1 False 5 8.0
2 True 6 9.0
In [377]: dft1.dtypes
Out[377]: 
a bool
b int64
c float64
dtype: object

Note

When trying to convert a subset of columns to a specified type using astype() and loc(), upcasting occurs.

loc() tries to fit in what we are assigning to the current dtypes, while [] will overwrite them taking the dtype from the right hand side. Therefore the following piece of code produces the unintended result.

In [378]: dft = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6], "c": [7, 8, 9]})
In [379]: dft.loc[:, ["a", "b"]].astype(np.uint8).dtypes
Out[379]: 
a uint8
b uint8
dtype: object
In [380]: dft.loc[:, ["a", "b"]] = dft.loc[:, ["a", "b"]].astype(np.uint8)
In [381]: dft.dtypes
Out[381]: 
a int64
b int64
c int64
dtype: object

object conversion#

pandas offers various functions to try to force conversion of types from the object dtype to other types. In cases where the data is already of the correct type, but stored in an object array, the DataFrame.infer_objects() and Series.infer_objects() methods can be used to soft convert to the correct type.

In [382]: importdatetime
In [383]: df = pd.DataFrame(
 .....:  [
 .....:  [1, 2],
 .....:  ["a", "b"],
 .....:  [datetime.datetime(2016, 3, 2), datetime.datetime(2016, 3, 2)],
 .....:  ]
 .....: )
 .....: 
In [384]: df = df.T
In [385]: df
Out[385]: 
 0 1 2
0 1 a 2016年03月02日 00:00:00
1 2 b 2016年03月02日 00:00:00
In [386]: df.dtypes
Out[386]: 
0 object
1 object
2 object
dtype: object

Because the data was transposed the original inference stored all columns as object, which infer_objects will correct.

In [387]: df.infer_objects().dtypes
Out[387]: 
0 int64
1 object
2 datetime64[ns]
dtype: object

The following functions are available for one dimensional object arrays or scalars to perform hard conversion of objects to a specified type:

• to_numeric() (conversion to numeric dtypes)

  In [388]: m = ["1.1", 2, 3]
  In [389]: pd.to_numeric(m)
  Out[389]: array([1.1, 2. , 3. ])
  

• to_datetime() (conversion to datetime objects)

  In [390]: importdatetime
  In [391]: m = ["2016年07月09日", datetime.datetime(2016, 3, 2)]
  In [392]: pd.to_datetime(m)
  Out[392]: DatetimeIndex(['2016年07月09日', '2016年03月02日'], dtype='datetime64[ns]', freq=None)
  

• to_timedelta() (conversion to timedelta objects)

  In [393]: m = ["5us", pd.Timedelta("1day")]
  In [394]: pd.to_timedelta(m)
  Out[394]: TimedeltaIndex(['0 days 00:00:00.000005', '1 days 00:00:00'], dtype='timedelta64[ns]', freq=None)
  

To force a conversion, we can pass in an errors argument, which specifies how pandas should deal with elements that cannot be converted to desired dtype or object. By default, errors='raise', meaning that any errors encountered will be raised during the conversion process. However, if errors='coerce', these errors will be ignored and pandas will convert problematic elements to pd.NaT (for datetime and timedelta) or np.nan (for numeric). This might be useful if you are reading in data which is mostly of the desired dtype (e.g. numeric, datetime), but occasionally has non-conforming elements intermixed that you want to represent as missing:

In [395]: importdatetime
In [396]: m = ["apple", datetime.datetime(2016, 3, 2)]
In [397]: pd.to_datetime(m, errors="coerce")
Out[397]: DatetimeIndex(['NaT', '2016年03月02日'], dtype='datetime64[ns]', freq=None)
In [398]: m = ["apple", 2, 3]
In [399]: pd.to_numeric(m, errors="coerce")
Out[399]: array([nan, 2., 3.])
In [400]: m = ["apple", pd.Timedelta("1day")]
In [401]: pd.to_timedelta(m, errors="coerce")
Out[401]: TimedeltaIndex([NaT, '1 days'], dtype='timedelta64[ns]', freq=None)

In addition to object conversion, to_numeric() provides another argument downcast, which gives the option of downcasting the newly (or already) numeric data to a smaller dtype, which can conserve memory:

In [402]: m = ["1", 2, 3]
In [403]: pd.to_numeric(m, downcast="integer") # smallest signed int dtype
Out[403]: array([1, 2, 3], dtype=int8)
In [404]: pd.to_numeric(m, downcast="signed") # same as 'integer'
Out[404]: array([1, 2, 3], dtype=int8)
In [405]: pd.to_numeric(m, downcast="unsigned") # smallest unsigned int dtype
Out[405]: array([1, 2, 3], dtype=uint8)
In [406]: pd.to_numeric(m, downcast="float") # smallest float dtype
Out[406]: array([1., 2., 3.], dtype=float32)

As these methods apply only to one-dimensional arrays, lists or scalars; they cannot be used directly on multi-dimensional objects such as DataFrames. However, with apply(), we can "apply" the function over each column efficiently:

In [407]: importdatetime
In [408]: df = pd.DataFrame([["2016年07月09日", datetime.datetime(2016, 3, 2)]] * 2, dtype="O")
In [409]: df
Out[409]: 
 0 1
0 2016年07月09日 2016年03月02日 00:00:00
1 2016年07月09日 2016年03月02日 00:00:00
In [410]: df.apply(pd.to_datetime)
Out[410]: 
 0 1
0 2016年07月09日 2016年03月02日
1 2016年07月09日 2016年03月02日
In [411]: df = pd.DataFrame([["1.1", 2, 3]] * 2, dtype="O")
In [412]: df
Out[412]: 
 0 1 2
0 1.1 2 3
1 1.1 2 3
In [413]: df.apply(pd.to_numeric)
Out[413]: 
 0 1 2
0 1.1 2 3
1 1.1 2 3
In [414]: df = pd.DataFrame([["5us", pd.Timedelta("1day")]] * 2, dtype="O")
In [415]: df
Out[415]: 
 0 1
0 5us 1 days 00:00:00
1 5us 1 days 00:00:00
In [416]: df.apply(pd.to_timedelta)
Out[416]: 
 0 1
0 0 days 00:00:00.000005 1 days
1 0 days 00:00:00.000005 1 days

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. See also Support for integer NA.

In [417]: dfi = df3.astype("int32")
In [418]: dfi["E"] = 1
In [419]: dfi
Out[419]: 
 A B C E
0 1 0 26 1
1 3 1 86 1
2 0 0 46 1
3 0 1 212 1
4 -1 -1 26 1
5 1 0 7 1
6 0 -1 184 1
7 0 0 206 1
In [420]: dfi.dtypes
Out[420]: 
A int32
B int32
C int32
E int64
dtype: object
In [421]: casted = dfi[dfi > 0]
In [422]: casted
Out[422]: 
 A B C E
0 1.0 NaN 26 1
1 3.0 1.0 86 1
2 NaN NaN 46 1
3 NaN 1.0 212 1
4 NaN NaN 26 1
5 1.0 NaN 7 1
6 NaN NaN 184 1
7 NaN NaN 206 1
In [423]: casted.dtypes
Out[423]: 
A float64
B float64
C int32
E int64
dtype: object

While float dtypes are unchanged.

In [424]: dfa = df3.copy()
In [425]: dfa["A"] = dfa["A"].astype("float32")
In [426]: dfa.dtypes
Out[426]: 
A float32
B float64
C float64
dtype: object
In [427]: casted = dfa[df2 > 0]
In [428]: casted
Out[428]: 
 A B C
0 1.047606 0.256090 26.0
1 3.497968 1.426469 86.0
2 NaN NaN 46.0
3 NaN 1.139976 212.0
4 NaN NaN 26.0
5 1.346426 0.096706 7.0
6 NaN NaN 184.0
7 NaN NaN 206.0
In [429]: casted.dtypes
Out[429]: 
A float32
B float64
C float64
dtype: object

Selecting columns based on dtype#

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 [430]: 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),
 .....:  "category": pd.Series(list("ABC")).astype("category"),
 .....:  }
 .....: )
 .....: 
In [431]: df["tdeltas"] = df.dates.diff()
In [432]: df["uint64"] = np.arange(3, 6).astype("u8")
In [433]: df["other_dates"] = pd.date_range("20130101", periods=3)
In [434]: df["tz_aware_dates"] = pd.date_range("20130101", periods=3, tz="US/Eastern")
In [435]: df
Out[435]: 
 string int64 uint8 ... uint64 other_dates tz_aware_dates
0 a 1 3 ... 3 2013年01月01日 2013年01月01日 00:00:00-05:00
1 b 2 4 ... 4 2013年01月02日 2013年01月02日 00:00:00-05:00
2 c 3 5 ... 5 2013年01月03日 2013年01月03日 00:00:00-05:00
[3 rows x 12 columns]

And the dtypes:

In [436]: df.dtypes
Out[436]: 
string object
int64 int64
uint8 uint8
float64 float64
bool1 bool
bool2 bool
dates datetime64[ns]
category category
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 [437]: df.select_dtypes(include=[bool])
Out[437]: 
 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 [438]: df.select_dtypes(include=["bool"])
Out[438]: 
 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 [439]: df.select_dtypes(include=["number", "bool"], exclude=["unsignedinteger"])
Out[439]: 
 int64 float64 bool1 bool2 tdeltas
0 1 4.0 True False NaT
1 2 5.0 False True 1 days
2 3 6.0 True False 1 days

To select string columns you must use the object dtype:

In [440]: df.select_dtypes(include=["object"])
Out[440]: 
 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 [441]: defsubdtypes(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 [442]: subdtypes(np.generic)
Out[442]: 
[numpy.generic,
 [[numpy.number,
 [[numpy.integer,
 [[numpy.signedinteger,
 [numpy.int8,
 numpy.int16,
 numpy.int32,
 numpy.int64,
 numpy.longlong,
 numpy.timedelta64]],
 [numpy.unsignedinteger,
 [numpy.uint8,
 numpy.uint16,
 numpy.uint32,
 numpy.uint64,
 numpy.ulonglong]]]],
 [numpy.inexact,
 [[numpy.floating,
 [numpy.float16, numpy.float32, numpy.float64, numpy.longdouble]],
 [numpy.complexfloating,
 [numpy.complex64, numpy.complex128, numpy.clongdouble]]]]]],
 [numpy.flexible,
 [[numpy.character, [numpy.bytes_, numpy.str_]],
 [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 won’t show up with the above function.