MultiIndex / 高级索引#

本节涵盖了使用 MultiIndex 进行索引其他高级索引功能

有关通用索引文档,请参阅索引和选择数据

有关一些高级策略,请参阅手册

分层索引(MultiIndex)#

分层/多级索引非常令人兴奋,它为更复杂的数据分析和操作打开了大门,尤其是在处理更高维度数据时。本质上,它使您能够将具有任意数量维数的数据存储和操作到较低维数的数据结构中,如Series(1d)和DataFrame(2d)。

在本节中,我们将展示“分层”索引的确切含义,以及它如何与前面几节和本节描述的所有 pandas 索引功能集成。稍后,在讨论分组依据数据重塑和透视时,我们将展示非平凡的应用,以说明它如何有助于构建用于分析的数据。

有关一些高级策略,请参阅手册

创建 MultiIndex(分层索引)对象#

MultiIndex 对象是标准 Index 对象的层级对应物,后者通常存储 pandas 对象中的轴标签。您可以将 MultiIndex 视为一个元组数组,其中每个元组都是唯一的。可以通过数组列表(使用 MultiIndex.from_arrays())、元组数组(使用 MultiIndex.from_tuples())、可迭代对象的交叉集(使用 MultiIndex.from_product())或 DataFrame(使用 MultiIndex.from_frame())来创建 MultiIndex。当 Index 构造函数接收到元组列表时,它将尝试返回一个 MultiIndex。以下示例演示了初始化 MultiIndex 的不同方法。

In [1]: arrays = [
   ...:     ["bar", "bar", "baz", "baz", "foo", "foo", "qux", "qux"],
   ...:     ["one", "two", "one", "two", "one", "two", "one", "two"],
   ...: ]
   ...: 

In [2]: tuples = list(zip(*arrays))

In [3]: tuples
Out[3]: 
[('bar', 'one'),
 ('bar', 'two'),
 ('baz', 'one'),
 ('baz', 'two'),
 ('foo', 'one'),
 ('foo', 'two'),
 ('qux', 'one'),
 ('qux', 'two')]

In [4]: index = pd.MultiIndex.from_tuples(tuples, names=["first", "second"])

In [5]: index
Out[5]: 
MultiIndex([('bar', 'one'),
            ('bar', 'two'),
            ('baz', 'one'),
            ('baz', 'two'),
            ('foo', 'one'),
            ('foo', 'two'),
            ('qux', 'one'),
            ('qux', 'two')],
           names=['first', 'second'])

In [6]: s = pd.Series(np.random.randn(8), index=index)

In [7]: s
Out[7]: 
first  second
bar    one       0.469112
       two      -0.282863
baz    one      -1.509059
       two      -1.135632
foo    one       1.212112
       two      -0.173215
qux    one       0.119209
       two      -1.044236
dtype: float64

当您需要两个可迭代对象的所有元素配对时,使用 MultiIndex.from_product() 方法可能会更简单。

In [8]: iterables = [["bar", "baz", "foo", "qux"], ["one", "two"]]

In [9]: pd.MultiIndex.from_product(iterables, names=["first", "second"])
Out[9]: 
MultiIndex([('bar', 'one'),
            ('bar', 'two'),
            ('baz', 'one'),
            ('baz', 'two'),
            ('foo', 'one'),
            ('foo', 'two'),
            ('qux', 'one'),
            ('qux', 'two')],
           names=['first', 'second'])

您还可以直接使用 MultiIndex.from_frame() 方法从 DataFrame 构建 MultiIndex。这是 MultiIndex.to_frame() 方法的补充。

In [10]: df = pd.DataFrame(
   ....:     [["bar", "one"], ["bar", "two"], ["foo", "one"], ["foo", "two"]],
   ....:     columns=["first", "second"],
   ....: )
   ....: 

In [11]: pd.MultiIndex.from_frame(df)
Out[11]: 
MultiIndex([('bar', 'one'),
            ('bar', 'two'),
            ('foo', 'one'),
            ('foo', 'two')],
           names=['first', 'second'])

为了方便起见,您可以直接将数组列表传递给 SeriesDataFrame 以自动构建 MultiIndex

In [12]: arrays = [
   ....:     np.array(["bar", "bar", "baz", "baz", "foo", "foo", "qux", "qux"]),
   ....:     np.array(["one", "two", "one", "two", "one", "two", "one", "two"]),
   ....: ]
   ....: 

In [13]: s = pd.Series(np.random.randn(8), index=arrays)

In [14]: s
Out[14]: 
bar  one   -0.861849
     two   -2.104569
baz  one   -0.494929
     two    1.071804
foo  one    0.721555
     two   -0.706771
qux  one   -1.039575
     two    0.271860
dtype: float64

In [15]: df = pd.DataFrame(np.random.randn(8, 4), index=arrays)

In [16]: df
Out[16]: 
                0         1         2         3
bar one -0.424972  0.567020  0.276232 -1.087401
    two -0.673690  0.113648 -1.478427  0.524988
baz one  0.404705  0.577046 -1.715002 -1.039268
    two -0.370647 -1.157892 -1.344312  0.844885
foo one  1.075770 -0.109050  1.643563 -1.469388
    two  0.357021 -0.674600 -1.776904 -0.968914
qux one -1.294524  0.413738  0.276662 -0.472035
    two -0.013960 -0.362543 -0.006154 -0.923061

所有 MultiIndex 构造函数都接受一个 names 参数,该参数存储了层级的字符串名称。如果未提供名称,则会分配 None

In [17]: df.index.names
Out[17]: FrozenList([None, None])

此索引可以支持 pandas 对象的任何轴,并且索引的**级别**数量由您决定。

In [18]: df = pd.DataFrame(np.random.randn(3, 8), index=["A", "B", "C"], columns=index)

In [19]: df
Out[19]: 
first        bar                 baz  ...       foo       qux          
second       one       two       one  ...       two       one       two
A       0.895717  0.805244 -1.206412  ...  1.340309 -1.170299 -0.226169
B       0.410835  0.813850  0.132003  ... -1.187678  1.130127 -1.436737
C      -1.413681  1.607920  1.024180  ... -2.211372  0.974466 -2.006747

[3 rows x 8 columns]

In [20]: pd.DataFrame(np.random.randn(6, 6), index=index[:6], columns=index[:6])
Out[20]: 
first              bar                 baz                 foo          
second             one       two       one       two       one       two
first second                                                            
bar   one    -0.410001 -0.078638  0.545952 -1.219217 -1.226825  0.769804
      two    -1.281247 -0.727707 -0.121306 -0.097883  0.695775  0.341734
baz   one     0.959726 -1.110336 -0.619976  0.149748 -0.732339  0.687738
      two     0.176444  0.403310 -0.154951  0.301624 -2.179861 -1.369849
foo   one    -0.954208  1.462696 -1.743161 -0.826591 -0.345352  1.314232
      two     0.690579  0.995761  2.396780  0.014871  3.357427 -0.317441

我们已经“稀疏化”了索引的较高层级,以便在控制台中显示更易读。请注意,索引的显示方式可以通过 pandas.set_options() 中的 multi_sparse 选项进行控制。

In [21]: with pd.option_context("display.multi_sparse", False):
   ....:     df
   ....: 

值得注意的是,没有任何东西可以阻止您将元组用作轴上的原子标签。

In [22]: pd.Series(np.random.randn(8), index=tuples)
Out[22]: 
(bar, one)   -1.236269
(bar, two)    0.896171
(baz, one)   -0.487602
(baz, two)   -0.082240
(foo, one)   -2.182937
(foo, two)    0.380396
(qux, one)    0.084844
(qux, two)    0.432390
dtype: float64

MultiIndex 之所以重要,是因为它允许我们进行分组、选择和重塑操作,正如我们将在下文和后续文档部分中所述。正如您将在后续部分中看到的,您可能会发现自己在处理分层索引数据,而无需自己显式创建 MultiIndex。但是,在从文件加载数据时,您可能希望在准备数据集时生成自己的 MultiIndex

重建层级标签#

get_level_values() 方法将返回特定层级上每个位置的标签向量。

In [23]: index.get_level_values(0)
Out[23]: Index(['bar', 'bar', 'baz', 'baz', 'foo', 'foo', 'qux', 'qux'], dtype='str', name='first')

In [24]: index.get_level_values("second")
Out[24]: Index(['one', 'two', 'one', 'two', 'one', 'two', 'one', 'two'], dtype='str', name='second')

使用 MultiIndex 在轴上进行基本索引#

分层索引的一个重要特性是,您可以通过一个“部分”标签来选择数据,该标签标识数据中的一个子组。**部分**选择以与选择常规 DataFrame 中的列完全相同的方式,“丢弃”分层索引的层级。

In [25]: df["bar"]
Out[25]: 
second       one       two
A       0.895717  0.805244
B       0.410835  0.813850
C      -1.413681  1.607920

In [26]: df["bar", "one"]
Out[26]: 
A    0.895717
B    0.410835
C   -1.413681
Name: (bar, one), dtype: float64

In [27]: df["bar"]["one"]
Out[27]: 
A    0.895717
B    0.410835
C   -1.413681
Name: one, dtype: float64

In [28]: s["qux"]
Out[28]: 
one   -1.039575
two    0.271860
dtype: float64

有关如何在更深的层级上进行选择,请参阅分层索引的横截面

已定义的层级#

MultiIndex 保留了索引的所有已定义层级,即使它们实际上并未被使用。切片索引时,您可能会注意到这一点。例如:

In [29]: df.columns.levels  # original MultiIndex
Out[29]: FrozenList([['bar', 'baz', 'foo', 'qux'], ['one', 'two']])

In [30]: df[["foo","qux"]].columns.levels  # sliced
Out[30]: FrozenList([['bar', 'baz', 'foo', 'qux'], ['one', 'two']])

这样做是为了避免重新计算层级,从而使切片具有很高的性能。如果您只想查看使用的层级,可以使用 get_level_values() 方法。

In [31]: df[["foo", "qux"]].columns.to_numpy()
Out[31]: 
array([('foo', 'one'), ('foo', 'two'), ('qux', 'one'), ('qux', 'two')],
      dtype=object)

# for a specific level
In [32]: df[["foo", "qux"]].columns.get_level_values(0)
Out[32]: Index(['foo', 'foo', 'qux', 'qux'], dtype='str', name='first')

要使用仅包含已用层级的 MultiIndex 进行重建,可以使用 remove_unused_levels() 方法。

In [33]: new_mi = df[["foo", "qux"]].columns.remove_unused_levels()

In [34]: new_mi.levels
Out[34]: FrozenList([['foo', 'qux'], ['one', 'two']])

数据对齐和使用 reindex#

具有 MultiIndex 作为轴的、索引不同的对象之间的操作将按预期工作;数据对齐将与元组索引的工作方式相同。

In [35]: s + s[:-2]
Out[35]: 
bar  one   -1.723698
     two   -4.209138
baz  one   -0.989859
     two    2.143608
foo  one    1.443110
     two   -1.413542
qux  one         NaN
     two         NaN
dtype: float64

In [36]: s + s[::2]
Out[36]: 
bar  one   -1.723698
     two         NaN
baz  one   -0.989859
     two         NaN
foo  one    1.443110
     two         NaN
qux  one   -2.079150
     two         NaN
dtype: float64

Series/DataFramesreindex() 方法可以与另一个 MultiIndex,甚至是一个元组列表或数组一起调用。

In [37]: s.reindex(index[:3])
Out[37]: 
first  second
bar    one      -0.861849
       two      -2.104569
baz    one      -0.494929
dtype: float64

In [38]: s.reindex([("foo", "two"), ("bar", "one"), ("qux", "one"), ("baz", "one")])
Out[38]: 
foo  two   -0.706771
bar  one   -0.861849
qux  one   -1.039575
baz  one   -0.494929
dtype: float64

使用分层索引进行高级索引#

在语法上将 MultiIndex 集成到使用 .loc 的高级索引中有点挑战性,但我们已尽一切努力做到这一点。通常,MultiIndex 键采用元组的形式。例如,以下内容按预期工作:

In [39]: df = df.T

In [40]: df
Out[40]: 
                     A         B         C
first second                              
bar   one     0.895717  0.410835 -1.413681
      two     0.805244  0.813850  1.607920
baz   one    -1.206412  0.132003  1.024180
      two     2.565646 -0.827317  0.569605
foo   one     1.431256 -0.076467  0.875906
      two     1.340309 -1.187678 -2.211372
qux   one    -1.170299  1.130127  0.974466
      two    -0.226169 -1.436737 -2.006747

In [41]: df.loc[("bar", "two")]
Out[41]: 
A    0.805244
B    0.813850
C    1.607920
Name: (bar, two), dtype: float64

请注意,在此示例中 df.loc['bar', 'two'] 也可以正常工作,但这种简写符号在一般情况下可能导致歧义。

如果您还想使用 .loc 索引特定列,则必须使用如下元组:

In [42]: df.loc[("bar", "two"), "A"]
Out[42]: np.float64(0.8052440253863785)

您不必通过传递元组的第一个元素来指定 MultiIndex 的所有层级。例如,您可以使用“部分”索引来获取第一个层级中所有包含 bar 的元素,如下所示:

In [43]: df.loc["bar"]
Out[43]: 
               A         B         C
second                              
one     0.895717  0.410835 -1.413681
two     0.805244  0.813850  1.607920

这是更冗长的表示法 df.loc[('bar',),](在此示例中等同于 df.loc['bar',])的快捷方式。

“部分”切片也能很好地工作。

In [44]: df.loc["baz":"foo"]
Out[44]: 
                     A         B         C
first second                              
baz   one    -1.206412  0.132003  1.024180
      two     2.565646 -0.827317  0.569605
foo   one     1.431256 -0.076467  0.875906
      two     1.340309 -1.187678 -2.211372

您可以通过提供元组的切片来使用“范围”值进行切片。

In [45]: df.loc[("baz", "two"):("qux", "one")]
Out[45]: 
                     A         B         C
first second                              
baz   two     2.565646 -0.827317  0.569605
foo   one     1.431256 -0.076467  0.875906
      two     1.340309 -1.187678 -2.211372
qux   one    -1.170299  1.130127  0.974466

In [46]: df.loc[("baz", "two"):"foo"]
Out[46]: 
                     A         B         C
first second                              
baz   two     2.565646 -0.827317  0.569605
foo   one     1.431256 -0.076467  0.875906
      two     1.340309 -1.187678 -2.211372

传递标签列表或元组列表,其工作方式类似于重新索引。

In [47]: df.loc[[("bar", "two"), ("qux", "one")]]
Out[47]: 
                     A         B         C
first second                              
bar   two     0.805244  0.813850  1.607920
qux   one    -1.170299  1.130127  0.974466

注意

值得注意的是,在 pandas 中,元组和列表在索引方面并非被同等对待。元组被解释为一个多级键,而列表用于指定多个键。换句话说,元组是水平移动(遍历层级),列表是垂直移动(扫描层级)。

重要的是,元组列表索引了几个完整的 MultiIndex 键,而列表元组则引用了层级中的多个值。

In [48]: s = pd.Series(
   ....:     [1, 2, 3, 4, 5, 6],
   ....:     index=pd.MultiIndex.from_product([["A", "B"], ["c", "d", "e"]]),
   ....: )
   ....: 

In [49]: s.loc[[("A", "c"), ("B", "d")]]  # list of tuples
Out[49]: 
A  c    1
B  d    5
dtype: int64

In [50]: s.loc[(["A", "B"], ["c", "d"])]  # tuple of lists
Out[50]: 
A  c    1
   d    2
B  c    4
   d    5
dtype: int64

使用切片器#

您可以通过提供多个索引器来切片 MultiIndex

您可以提供任何索引器,就像按标签进行索引一样,请参阅按标签选择,包括切片、标签列表、标签和布尔索引器。

您可以使用 slice(None) 来选择该层级的所有内容。您无需指定所有**更深层级**的,它们将默认为 slice(None)

一如既往,切片器的**两侧**都被包含在内,因为这是标签索引。

警告

您应该在 .loc 指定符中指定所有轴,即**索引**和**列**的索引器。存在一些模棱两可的情况,此时传递的索引器可能会被误解为索引**两个**轴,而不是索引行的 MultiIndex

您应该这样做:

df.loc[(slice("A1", "A3"), ...), :]  # noqa: E999

您**不**应该这样做:

df.loc[(slice("A1", "A3"), ...)]  # noqa: E999
In [51]: def mklbl(prefix, n):
   ....:     return ["%s%s" % (prefix, i) for i in range(n)]
   ....: 

In [52]: miindex = pd.MultiIndex.from_product(
   ....:     [mklbl("A", 4), mklbl("B", 2), mklbl("C", 4), mklbl("D", 2)]
   ....: )
   ....: 

In [53]: micolumns = pd.MultiIndex.from_tuples(
   ....:     [("a", "foo"), ("a", "bar"), ("b", "foo"), ("b", "bah")], names=["lvl0", "lvl1"]
   ....: )
   ....: 

In [54]: dfmi = (
   ....:     pd.DataFrame(
   ....:         np.arange(len(miindex) * len(micolumns)).reshape(
   ....:             (len(miindex), len(micolumns))
   ....:         ),
   ....:         index=miindex,
   ....:         columns=micolumns,
   ....:     )
   ....:     .sort_index()
   ....:     .sort_index(axis=1)
   ....: )
   ....: 

In [55]: dfmi
Out[55]: 
lvl0           a         b     
lvl1         bar  foo  bah  foo
A0 B0 C0 D0    1    0    3    2
         D1    5    4    7    6
      C1 D0    9    8   11   10
         D1   13   12   15   14
      C2 D0   17   16   19   18
...          ...  ...  ...  ...
A3 B1 C1 D1  237  236  239  238
      C2 D0  241  240  243  242
         D1  245  244  247  246
      C3 D0  249  248  251  250
         D1  253  252  255  254

[64 rows x 4 columns]

使用切片、列表和标签的基本 MultiIndex 切片。

In [56]: dfmi.loc[(slice("A1", "A3"), slice(None), ["C1", "C3"]), :]
Out[56]: 
lvl0           a         b     
lvl1         bar  foo  bah  foo
A1 B0 C1 D0   73   72   75   74
         D1   77   76   79   78
      C3 D0   89   88   91   90
         D1   93   92   95   94
   B1 C1 D0  105  104  107  106
...          ...  ...  ...  ...
A3 B0 C3 D1  221  220  223  222
   B1 C1 D0  233  232  235  234
         D1  237  236  239  238
      C3 D0  249  248  251  250
         D1  253  252  255  254

[24 rows x 4 columns]

您可以使用 pandas.IndexSlice 来实现使用 : 的更自然的语法,而不是使用 slice(None)

In [57]: idx = pd.IndexSlice

In [58]: dfmi.loc[idx[:, :, ["C1", "C3"]], idx[:, "foo"]]
Out[58]: 
lvl0           a    b
lvl1         foo  foo
A0 B0 C1 D0    8   10
         D1   12   14
      C3 D0   24   26
         D1   28   30
   B1 C1 D0   40   42
...          ...  ...
A3 B0 C3 D1  220  222
   B1 C1 D0  232  234
         D1  236  238
      C3 D0  248  250
         D1  252  254

[32 rows x 2 columns]

使用此方法可以在多个轴上同时执行相当复杂的选择。

In [59]: dfmi.loc["A1", (slice(None), "foo")]
Out[59]: 
lvl0        a    b
lvl1      foo  foo
B0 C0 D0   64   66
      D1   68   70
   C1 D0   72   74
      D1   76   78
   C2 D0   80   82
...       ...  ...
B1 C1 D1  108  110
   C2 D0  112  114
      D1  116  118
   C3 D0  120  122
      D1  124  126

[16 rows x 2 columns]

In [60]: dfmi.loc[idx[:, :, ["C1", "C3"]], idx[:, "foo"]]
Out[60]: 
lvl0           a    b
lvl1         foo  foo
A0 B0 C1 D0    8   10
         D1   12   14
      C3 D0   24   26
         D1   28   30
   B1 C1 D0   40   42
...          ...  ...
A3 B0 C3 D1  220  222
   B1 C1 D0  232  234
         D1  236  238
      C3 D0  248  250
         D1  252  254

[32 rows x 2 columns]

使用布尔索引器,您可以提供与**值**相关的选择。

In [61]: mask = dfmi[("a", "foo")] > 200

In [62]: dfmi.loc[idx[mask, :, ["C1", "C3"]], idx[:, "foo"]]
Out[62]: 
lvl0           a    b
lvl1         foo  foo
A3 B0 C1 D1  204  206
      C3 D0  216  218
         D1  220  222
   B1 C1 D0  232  234
         D1  236  238
      C3 D0  248  250
         D1  252  254

您还可以指定 .locaxis 参数来解释在单个轴上传递的切片。

In [63]: dfmi.loc(axis=0)[:, :, ["C1", "C3"]]
Out[63]: 
lvl0           a         b     
lvl1         bar  foo  bah  foo
A0 B0 C1 D0    9    8   11   10
         D1   13   12   15   14
      C3 D0   25   24   27   26
         D1   29   28   31   30
   B1 C1 D0   41   40   43   42
...          ...  ...  ...  ...
A3 B0 C3 D1  221  220  223  222
   B1 C1 D0  233  232  235  234
         D1  237  236  239  238
      C3 D0  249  248  251  250
         D1  253  252  255  254

[32 rows x 4 columns]

此外,您还可以使用以下方法**设置**值。

In [64]: df2 = dfmi.copy()

In [65]: df2.loc(axis=0)[:, :, ["C1", "C3"]] = -10

In [66]: df2
Out[66]: 
lvl0           a         b     
lvl1         bar  foo  bah  foo
A0 B0 C0 D0    1    0    3    2
         D1    5    4    7    6
      C1 D0  -10  -10  -10  -10
         D1  -10  -10  -10  -10
      C2 D0   17   16   19   18
...          ...  ...  ...  ...
A3 B1 C1 D1  -10  -10  -10  -10
      C2 D0  241  240  243  242
         D1  245  244  247  246
      C3 D0  -10  -10  -10  -10
         D1  -10  -10  -10  -10

[64 rows x 4 columns]

您也可以使用可对齐对象的右侧。

In [67]: df2 = dfmi.copy()

In [68]: df2.loc[idx[:, :, ["C1", "C3"]], :] = df2 * 1000

In [69]: df2
Out[69]: 
lvl0              a               b        
lvl1            bar     foo     bah     foo
A0 B0 C0 D0       1       0       3       2
         D1       5       4       7       6
      C1 D0    9000    8000   11000   10000
         D1   13000   12000   15000   14000
      C2 D0      17      16      19      18
...             ...     ...     ...     ...
A3 B1 C1 D1  237000  236000  239000  238000
      C2 D0     241     240     243     242
         D1     245     244     247     246
      C3 D0  249000  248000  251000  250000
         D1  253000  252000  255000  254000

[64 rows x 4 columns]

横截面#

DataFramexs() 方法还接受一个 level 参数,以便于在 MultiIndex 的特定层级上选择数据。

In [70]: df
Out[70]: 
                     A         B         C
first second                              
bar   one     0.895717  0.410835 -1.413681
      two     0.805244  0.813850  1.607920
baz   one    -1.206412  0.132003  1.024180
      two     2.565646 -0.827317  0.569605
foo   one     1.431256 -0.076467  0.875906
      two     1.340309 -1.187678 -2.211372
qux   one    -1.170299  1.130127  0.974466
      two    -0.226169 -1.436737 -2.006747

In [71]: df.xs("one", level="second")
Out[71]: 
              A         B         C
first                              
bar    0.895717  0.410835 -1.413681
baz   -1.206412  0.132003  1.024180
foo    1.431256 -0.076467  0.875906
qux   -1.170299  1.130127  0.974466
# using the slicers
In [72]: df.loc[(slice(None), "one"), :]
Out[72]: 
                     A         B         C
first second                              
bar   one     0.895717  0.410835 -1.413681
baz   one    -1.206412  0.132003  1.024180
foo   one     1.431256 -0.076467  0.875906
qux   one    -1.170299  1.130127  0.974466

您还可以通过提供 axis 参数,使用 xs 选择列。

In [73]: df = df.T

In [74]: df.xs("one", level="second", axis=1)
Out[74]: 
first       bar       baz       foo       qux
A      0.895717 -1.206412  1.431256 -1.170299
B      0.410835  0.132003 -0.076467  1.130127
C     -1.413681  1.024180  0.875906  0.974466
# using the slicers
In [75]: df.loc[:, (slice(None), "one")]
Out[75]: 
first        bar       baz       foo       qux
second       one       one       one       one
A       0.895717 -1.206412  1.431256 -1.170299
B       0.410835  0.132003 -0.076467  1.130127
C      -1.413681  1.024180  0.875906  0.974466

xs 也允许使用多个键进行选择。

In [76]: df.xs(("one", "bar"), level=("second", "first"), axis=1)
Out[76]: 
first        bar
second       one
A       0.895717
B       0.410835
C      -1.413681
# using the slicers
In [77]: df.loc[:, ("bar", "one")]
Out[77]: 
A    0.895717
B    0.410835
C   -1.413681
Name: (bar, one), dtype: float64

您可以将 drop_level=False 传递给 xs,以保留被选中的层级。

In [78]: df.xs("one", level="second", axis=1, drop_level=False)
Out[78]: 
first        bar       baz       foo       qux
second       one       one       one       one
A       0.895717 -1.206412  1.431256 -1.170299
B       0.410835  0.132003 -0.076467  1.130127
C      -1.413681  1.024180  0.875906  0.974466

将以上结果与使用 drop_level=True(默认值)的结果进行比较。

In [79]: df.xs("one", level="second", axis=1, drop_level=True)
Out[79]: 
first       bar       baz       foo       qux
A      0.895717 -1.206412  1.431256 -1.170299
B      0.410835  0.132003 -0.076467  1.130127
C     -1.413681  1.024180  0.875906  0.974466

高级重新索引和对齐#

在 pandas 对象的 reindex()align() 方法中使用 level 参数,有助于在层级上广播值。例如:

In [80]: midx = pd.MultiIndex(
   ....:     levels=[["zero", "one"], ["x", "y"]], codes=[[1, 1, 0, 0], [1, 0, 1, 0]]
   ....: )
   ....: 

In [81]: df = pd.DataFrame(np.random.randn(4, 2), index=midx)

In [82]: df
Out[82]: 
               0         1
one  y  1.519970 -0.493662
     x  0.600178  0.274230
zero y  0.132885 -0.023688
     x  2.410179  1.450520

In [83]: df2 = df.groupby(level=0).mean()

In [84]: df2
Out[84]: 
             0         1
one   1.060074 -0.109716
zero  1.271532  0.713416

In [85]: df2.reindex(df.index, level=0)
Out[85]: 
               0         1
one  y  1.060074 -0.109716
     x  1.060074 -0.109716
zero y  1.271532  0.713416
     x  1.271532  0.713416

# aligning
In [86]: df_aligned, df2_aligned = df.align(df2, level=0)

In [87]: df_aligned
Out[87]: 
               0         1
one  y  1.519970 -0.493662
     x  0.600178  0.274230
zero y  0.132885 -0.023688
     x  2.410179  1.450520

In [88]: df2_aligned
Out[88]: 
               0         1
one  y  1.060074 -0.109716
     x  1.060074 -0.109716
zero y  1.271532  0.713416
     x  1.271532  0.713416

使用 swaplevel 交换层级#

swaplevel() 方法可以改变两个层级的顺序。

In [89]: df[:5]
Out[89]: 
               0         1
one  y  1.519970 -0.493662
     x  0.600178  0.274230
zero y  0.132885 -0.023688
     x  2.410179  1.450520

In [90]: df[:5].swaplevel(0, 1, axis=0)
Out[90]: 
               0         1
y one   1.519970 -0.493662
x one   0.600178  0.274230
y zero  0.132885 -0.023688
x zero  2.410179  1.450520

使用 reorder_levels 重排层级#

reorder_levels() 方法泛化了 swaplevel 方法,允许您一步排列分层索引层级。

In [91]: df[:5].reorder_levels([1, 0], axis=0)
Out[91]: 
               0         1
y one   1.519970 -0.493662
x one   0.600178  0.274230
y zero  0.132885 -0.023688
x zero  2.410179  1.450520

重命名 IndexMultiIndex 的名称#

rename() 方法用于重命名 MultiIndex 的标签,通常用于重命名 DataFrame 的列。 renamecolumns 参数允许指定一个字典,其中只包含您希望重命名的列。

In [92]: df.rename(columns={0: "col0", 1: "col1"})
Out[92]: 
            col0      col1
one  y  1.519970 -0.493662
     x  0.600178  0.274230
zero y  0.132885 -0.023688
     x  2.410179  1.450520

此方法也可用于重命名 DataFrame 主索引的特定标签。

In [93]: df.rename(index={"one": "two", "y": "z"})
Out[93]: 
               0         1
two  z  1.519970 -0.493662
     x  0.600178  0.274230
zero z  0.132885 -0.023688
     x  2.410179  1.450520

rename_axis() 方法用于重命名 IndexMultiIndex 的名称。特别是,可以指定 MultiIndex 层级的名称,这在稍后使用 reset_index()MultiIndex 的值移到列中时很有用。

In [94]: df.rename_axis(index=["abc", "def"])
Out[94]: 
                 0         1
abc  def                    
one  y    1.519970 -0.493662
     x    0.600178  0.274230
zero y    0.132885 -0.023688
     x    2.410179  1.450520

请注意,DataFrame 的列是一个索引,因此使用带有 columns 参数的 rename_axis 将更改该索引的名称。

In [95]: df.rename_axis(columns="Cols").columns
Out[95]: RangeIndex(start=0, stop=2, step=1, name='Cols')

renamerename_axis 都支持指定字典、Series 或映射函数来将标签/名称映射到新值。

当直接使用 Index 对象而不是通过 DataFrame 工作时,可以使用 Index.set_names() 来更改名称。

In [96]: mi = pd.MultiIndex.from_product([[1, 2], ["a", "b"]], names=["x", "y"])

In [97]: mi.names
Out[97]: FrozenList(['x', 'y'])

In [98]: mi2 = mi.rename("new name", level=0)

In [99]: mi2
Out[99]: 
MultiIndex([(1, 'a'),
            (1, 'b'),
            (2, 'a'),
            (2, 'b')],
           names=['new name', 'y'])

您无法通过层级设置 MultiIndex 的名称。

In [100]: mi.levels[0].name = "name via level"
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[100], line 1
----> 1 mi.levels[0].name = "name via level"

File ~/work/pandas/pandas/pandas/core/indexes/base.py:1809, in Index.name(self, value)
   1805 @name.setter
   1806 def name(self, value: Hashable) -> None:
   1807     if self._no_setting_name:
   1808         # Used in MultiIndex.levels to avoid silently ignoring name updates.
-> 1809         raise RuntimeError(
   1810             "Cannot set name on a level of a MultiIndex. Use "
   1811             "'MultiIndex.set_names' instead."
   1812         )
   1813     maybe_extract_name(value, None, type(self))
   1814     self._name = value

RuntimeError: Cannot set name on a level of a MultiIndex. Use 'MultiIndex.set_names' instead.

请改用 Index.set_names()

排序 MultiIndex#

为了有效地索引和切片 MultiIndex 对象,需要对其进行排序。与任何索引一样,您可以使用 sort_index()

In [101]: import random

In [102]: random.shuffle(tuples)

In [103]: s = pd.Series(np.random.randn(8), index=pd.MultiIndex.from_tuples(tuples))

In [104]: s
Out[104]: 
bar  one    0.206053
baz  one   -0.251905
qux  one   -2.213588
foo  one    1.063327
     two    1.266143
bar  two    0.299368
baz  two   -0.863838
qux  two    0.408204
dtype: float64

In [105]: s.sort_index()
Out[105]: 
bar  one    0.206053
     two    0.299368
baz  one   -0.251905
     two   -0.863838
foo  one    1.063327
     two    1.266143
qux  one   -2.213588
     two    0.408204
dtype: float64

In [106]: s.sort_index(level=0)
Out[106]: 
bar  one    0.206053
     two    0.299368
baz  one   -0.251905
     two   -0.863838
foo  one    1.063327
     two    1.266143
qux  one   -2.213588
     two    0.408204
dtype: float64

In [107]: s.sort_index(level=1)
Out[107]: 
bar  one    0.206053
baz  one   -0.251905
foo  one    1.063327
qux  one   -2.213588
bar  two    0.299368
baz  two   -0.863838
foo  two    1.266143
qux  two    0.408204
dtype: float64

如果 MultiIndex 层级已命名,您也可以将层级名称传递给 sort_index

In [108]: s.index = s.index.set_names(["L1", "L2"])

In [109]: s.sort_index(level="L1")
Out[109]: 
L1   L2 
bar  one    0.206053
     two    0.299368
baz  one   -0.251905
     two   -0.863838
foo  one    1.063327
     two    1.266143
qux  one   -2.213588
     two    0.408204
dtype: float64

In [110]: s.sort_index(level="L2")
Out[110]: 
L1   L2 
bar  one    0.206053
baz  one   -0.251905
foo  one    1.063327
qux  one   -2.213588
bar  two    0.299368
baz  two   -0.863838
foo  two    1.266143
qux  two    0.408204
dtype: float64

在更高维度的对象上,您可以按层级对任何其他轴进行排序,如果它们具有 MultiIndex

In [111]: df.T.sort_index(level=1, axis=1)
Out[111]: 
        one      zero       one      zero
          x         x         y         y
0  0.600178  2.410179  1.519970  0.132885
1  0.274230  1.450520 -0.493662 -0.023688

即使数据未排序,索引也能工作,但效率会相当低(并显示 PerformanceWarning)。它还将返回数据的副本而不是视图。

In [112]: dfm = pd.DataFrame(
   .....:     {"jim": [0, 0, 1, 1], "joe": ["x", "x", "z", "y"], "jolie": np.random.rand(4)}
   .....: )
   .....: 

In [113]: dfm = dfm.set_index(["jim", "joe"])

In [114]: dfm
Out[114]: 
            jolie
jim joe          
0   x    0.490671
    x    0.120248
1   z    0.537020
    y    0.110968

In [115]: dfm.loc[(1, 'z')]
Out[115]: 
           jolie
jim joe         
1   z    0.53702

此外,如果您尝试索引一个未完全按字典序排序的内容,这可能会引发错误。

In [116]: dfm.loc[(0, 'y'):(1, 'z')]
---------------------------------------------------------------------------
UnsortedIndexError                        Traceback (most recent call last)
Cell In[116], line 1
----> 1 dfm.loc[(0, 'y'):(1, 'z')]

File ~/work/pandas/pandas/pandas/core/indexing.py:1207, in _LocationIndexer.__getitem__(self, key)
   1205 maybe_callable = com.apply_if_callable(key, self.obj)
   1206 maybe_callable = self._raise_callable_usage(key, maybe_callable)
-> 1207 return self._getitem_axis(maybe_callable, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexing.py:1429, in _LocIndexer._getitem_axis(self, key, axis)
   1427 if isinstance(key, slice):
   1428     self._validate_key(key, axis)
-> 1429     return self._get_slice_axis(key, axis=axis)
   1430 elif com.is_bool_indexer(key):
   1431     return self._getbool_axis(key, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexing.py:1461, in _LocIndexer._get_slice_axis(self, slice_obj, axis)
   1458     return obj.copy(deep=False)
   1460 labels = obj._get_axis(axis)
-> 1461 indexer = labels.slice_indexer(slice_obj.start, slice_obj.stop, slice_obj.step)
   1463 if isinstance(indexer, slice):
   1464     return self.obj._slice(indexer, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexes/base.py:6804, in Index.slice_indexer(self, start, end, step)
   6753 def slice_indexer(
   6754     self,
   6755     start: Hashable | None = None,
   6756     end: Hashable | None = None,
   6757     step: int | None = None,
   6758 ) -> slice:
   6759     """
   6760     Compute the slice indexer for input labels and step.
   6761 
   (...)   6802     slice(1, 3, None)
   6803     """
-> 6804     start_slice, end_slice = self.slice_locs(start, end, step=step)
   6806     # return a slice
   6807     if not is_scalar(start_slice):

File ~/work/pandas/pandas/pandas/core/indexes/multi.py:3382, in MultiIndex.slice_locs(self, start, end, step)
   3329 """
   3330 For an ordered MultiIndex, compute the slice locations for input
   3331 labels.
   (...)   3378                       sequence of such.
   3379 """
   3380 # This function adds nothing to its parent implementation (the magic
   3381 # happens in get_slice_bound method), but it adds meaningful doc.
-> 3382 return super().slice_locs(start, end, step)

File ~/work/pandas/pandas/pandas/core/indexes/base.py:7062, in Index.slice_locs(self, start, end, step)
   7060 start_slice = None
   7061 if start is not None:
-> 7062     start_slice = self.get_slice_bound(start, "left")
   7063 if start_slice is None:
   7064     start_slice = 0

File ~/work/pandas/pandas/pandas/core/indexes/multi.py:3324, in MultiIndex.get_slice_bound(self, label, side)
   3322 if not isinstance(label, tuple):
   3323     label = (label,)
-> 3324 result = self._partial_tup_index(label, side=side)
   3325 result = maybe_unbox_numpy_scalar(result)
   3326 return result

File ~/work/pandas/pandas/pandas/core/indexes/multi.py:3386, in MultiIndex._partial_tup_index(self, tup, side)
   3384 def _partial_tup_index(self, tup: tuple, side: Literal["left", "right"] = "left"):
   3385     if len(tup) > self._lexsort_depth:
-> 3386         raise UnsortedIndexError(
   3387             f"Key length ({len(tup)}) was greater than MultiIndex lexsort depth "
   3388             f"({self._lexsort_depth})"
   3389         )
   3391     n = len(tup)
   3392     start, end = 0, len(self)

UnsortedIndexError: 'Key length (2) was greater than MultiIndex lexsort depth (1)'

is_monotonic_increasing() 方法在 MultiIndex 上显示索引是否已排序。

In [117]: dfm.index.is_monotonic_increasing
Out[117]: False
In [118]: dfm = dfm.sort_index()

In [119]: dfm
Out[119]: 
            jolie
jim joe          
0   x    0.490671
    x    0.120248
1   y    0.110968
    z    0.537020

In [120]: dfm.index.is_monotonic_increasing
Out[120]: True

现在选择按预期工作。

In [121]: dfm.loc[(0, "y"):(1, "z")]
Out[121]: 
            jolie
jim joe          
1   y    0.110968
    z    0.537020

take 方法#

与 NumPy ndarrays 类似,pandas 的 IndexSeriesDataFrame 也提供了 take() 方法,该方法在给定索引处沿给定轴检索元素。给定的索引必须是整数索引位置的列表或 ndarray。take 也接受负整数作为相对于对象末尾的位置。

In [122]: index = pd.Index(np.random.randint(0, 1000, 10))

In [123]: index
Out[123]: Index([214, 502, 712, 567, 786, 175, 993, 133, 758, 329], dtype='int64')

In [124]: positions = [0, 9, 3]

In [125]: index[positions]
Out[125]: Index([214, 329, 567], dtype='int64')

In [126]: index.take(positions)
Out[126]: Index([214, 329, 567], dtype='int64')

In [127]: ser = pd.Series(np.random.randn(10))

In [128]: ser.iloc[positions]
Out[128]: 
0   -0.179666
9    1.824375
3    0.392149
dtype: float64

In [129]: ser.take(positions)
Out[129]: 
0   -0.179666
9    1.824375
3    0.392149
dtype: float64

对于 DataFrame,给定的索引应该是 1d 列表或 ndarray,用于指定行或列位置。

In [130]: frm = pd.DataFrame(np.random.randn(5, 3))

In [131]: frm.take([1, 4, 3])
Out[131]: 
          0         1         2
1 -1.237881  0.106854 -1.276829
4  0.629675 -1.425966  1.857704
3  0.979542 -1.633678  0.615855

In [132]: frm.take([0, 2], axis=1)
Out[132]: 
          0         2
0  0.595974  0.601544
1 -1.237881 -1.276829
2 -0.767101  1.499591
3  0.979542  0.615855
4  0.629675  1.857704

重要的是要注意,pandas 对象上的 take 方法不适用于布尔索引,并且可能返回意外结果。

In [133]: arr = np.random.randn(10)

In [134]: arr.take([False, False, True, True])
Out[134]: array([-1.1935, -1.1935,  0.6775,  0.6775])

In [135]: arr[[0, 1]]
Out[135]: array([-1.1935,  0.6775])

In [136]: ser = pd.Series(np.random.randn(10))

In [137]: ser.take([False, False, True, True])
Out[137]: 
0    0.233141
0    0.233141
1   -0.223540
1   -0.223540
dtype: float64

In [138]: ser.iloc[[0, 1]]
Out[138]: 
0    0.233141
1   -0.223540
dtype: float64

最后,作为一个关于性能的小提示,由于 take 方法处理的输入范围更窄,因此它可能比复杂的索引提供更好的性能。

In [139]: arr = np.random.randn(10000, 5)

In [140]: indexer = np.arange(10000)

In [141]: random.shuffle(indexer)

In [142]: %timeit arr[indexer]
   .....: %timeit arr.take(indexer, axis=0)
   .....: 
255 us +- 6.25 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)
56.2 us +- 1.18 us per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
In [143]: ser = pd.Series(arr[:, 0])

In [144]: %timeit ser.iloc[indexer]
   .....: %timeit ser.take(indexer)
   .....: 
141 us +- 12.7 us per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
126 us +- 22 us per loop (mean +- std. dev. of 7 runs, 10,000 loops each)

索引类型#

我们在前面的章节中已经详细讨论了 MultiIndex。有关 DatetimeIndexPeriodIndex 的文档在此处显示,有关 TimedeltaIndex 的文档在此处找到

在接下来的子节中,我们将重点介绍一些其他索引类型。

CategoricalIndex#

CategoricalIndex 是一种索引类型,它对于支持带有重复项的索引很有用。这是 Categorical 的一个容器,允许高效地索引和存储具有大量重复元素的索引。

In [145]: from pandas.api.types import CategoricalDtype

In [146]: df = pd.DataFrame({"A": np.arange(6), "B": list("aabbca")})

In [147]: df["B"] = df["B"].astype(CategoricalDtype(list("cab")))

In [148]: df
Out[148]: 
   A  B
0  0  a
1  1  a
2  2  b
3  3  b
4  4  c
5  5  a

In [149]: df.dtypes
Out[149]: 
A       int64
B    category
dtype: object

In [150]: df["B"].cat.categories
Out[150]: Index(['c', 'a', 'b'], dtype='str')

设置索引将创建一个 CategoricalIndex

In [151]: df2 = df.set_index("B")

In [152]: df2.index
Out[152]: CategoricalIndex(['a', 'a', 'b', 'b', 'c', 'a'], categories=['c', 'a', 'b'], ordered=False, dtype='category', name='B')

使用 __getitem__/.iloc/.loc 进行索引,其工作方式类似于带有重复项的 Index。索引器**必须**在类别中,否则操作将引发 KeyError

In [153]: df2.loc["a"]
Out[153]: 
   A
B   
a  0
a  1
a  5

CategoricalIndex 在索引后**保持不变**。

In [154]: df2.loc["a"].index
Out[154]: CategoricalIndex(['a', 'a', 'a'], categories=['c', 'a', 'b'], ordered=False, dtype='category', name='B')

对索引进行排序将按类别顺序排序(回想一下,我们使用 CategoricalDtype(list('cab')) 创建了索引,所以排序后的顺序是 cab)。

In [155]: df2.sort_index()
Out[155]: 
   A
B   
c  4
a  0
a  1
a  5
b  2
b  3

在索引上进行分组操作将保留索引的性质。

In [156]: df2.groupby(level=0, observed=True).sum()
Out[156]: 
   A
B   
c  4
a  6
b  5

In [157]: df2.groupby(level=0, observed=True).sum().index
Out[157]: CategoricalIndex(['c', 'a', 'b'], categories=['c', 'a', 'b'], ordered=False, dtype='category', name='B')

重新索引操作将根据传递的索引器的类型返回一个结果索引。传递列表将返回一个普通 Index;使用 Categorical 进行索引将返回一个 CategoricalIndex,该索引根据**传递的** Categorical dtype 的类别进行索引。这允许我们任意索引这些,即使是**不在**类别中的值,这与您如何重新索引**任何** pandas 索引类似。

In [158]: df3 = pd.DataFrame(
   .....:     {"A": np.arange(3), "B": pd.Series(list("abc")).astype("category")}
   .....: )
   .....: 

In [159]: df3 = df3.set_index("B")

In [160]: df3
Out[160]: 
   A
B   
a  0
b  1
c  2
In [161]: df3.reindex(["a", "e"])
Out[161]: 
     A
B     
a  0.0
e  NaN

In [162]: df3.reindex(["a", "e"]).index
Out[162]: Index(['a', 'e'], dtype='str', name='B')

In [163]: df3.reindex(pd.Categorical(["a", "e"], categories=list("abe")))
Out[163]: 
     A
B     
a  0.0
e  NaN

In [164]: df3.reindex(pd.Categorical(["a", "e"], categories=list("abe"))).index
Out[164]: CategoricalIndex(['a', 'e'], categories=['a', 'b', 'e'], ordered=False, dtype='category', name='B')

警告

CategoricalIndex 上的重塑和比较操作必须具有相同的类别,否则将引发 TypeError

In [165]: df4 = pd.DataFrame({"A": np.arange(2), "B": list("ba")})

In [166]: df4["B"] = df4["B"].astype(CategoricalDtype(list("ab")))

In [167]: df4 = df4.set_index("B")

In [168]: df4.index
Out[168]: CategoricalIndex(['b', 'a'], categories=['a', 'b'], ordered=False, dtype='category', name='B')

In [169]: df5 = pd.DataFrame({"A": np.arange(2), "B": list("bc")})

In [170]: df5["B"] = df5["B"].astype(CategoricalDtype(list("bc")))

In [171]: df5 = df5.set_index("B")

In [172]: df5.index
Out[172]: CategoricalIndex(['b', 'c'], categories=['b', 'c'], ordered=False, dtype='category', name='B')
In [173]: pd.concat([df4, df5])
Out[173]: 
   A
B   
b  0
a  1
b  0
c  1

RangeIndex#

RangeIndexIndex 的一个子类,它为所有 DataFrameSeries 对象提供了默认索引。RangeIndexIndex 的优化版本,可以表示单调有序的集合。这些类似于 Python 的 range 类型RangeIndex 始终具有 int64 的 dtype。

In [174]: idx = pd.RangeIndex(5)

In [175]: idx
Out[175]: RangeIndex(start=0, stop=5, step=1)

RangeIndex 是所有 DataFrameSeries 对象的默认索引。

In [176]: ser = pd.Series([1, 2, 3])

In [177]: ser.index
Out[177]: RangeIndex(start=0, stop=3, step=1)

In [178]: df = pd.DataFrame([[1, 2], [3, 4]])

In [179]: df.index
Out[179]: RangeIndex(start=0, stop=2, step=1)

In [180]: df.columns
Out[180]: RangeIndex(start=0, stop=2, step=1)

RangeIndex 的行为类似于具有 int64 dtype 的 Index,并且 RangeIndex 上的操作,如果结果无法用 RangeIndex 表示,但应具有整数 dtype,则将转换为具有 int64Index。例如:

In [181]: idx[[0, 2]]
Out[181]: RangeIndex(start=0, stop=4, step=2)

IntervalIndex#

IntervalIndex 及其自身 dtype IntervalDtype 以及 Interval 标量类型,为 pandas 中的区间表示法提供了头等支持。

IntervalIndex 允许一些独特的索引,并且还用作 cut()qcut() 的类别返回类型。

使用 IntervalIndex 进行索引#

IntervalIndex 可用作 SeriesDataFrame 中的索引。

In [182]: df = pd.DataFrame(
   .....:     {"A": [1, 2, 3, 4]}, index=pd.IntervalIndex.from_breaks([0, 1, 2, 3, 4])
   .....: )
   .....: 

In [183]: df
Out[183]: 
        A
(0, 1]  1
(1, 2]  2
(2, 3]  3
(3, 4]  4

通过 .loc 对间隔边界进行基于标签的索引,其行为符合预期,会选择该特定间隔。

In [184]: df.loc[2]
Out[184]: 
A    2
Name: (1, 2], dtype: int64

In [185]: df.loc[[2, 3]]
Out[185]: 
        A
(1, 2]  2
(2, 3]  3

如果您选择的标签包含在某个间隔内,该间隔也会被选中。

In [186]: df.loc[2.5]
Out[186]: 
A    3
Name: (2, 3], dtype: int64

In [187]: df.loc[[2.5, 3.5]]
Out[187]: 
        A
(2, 3]  3
(3, 4]  4

使用 Interval 进行选择只会返回精确匹配项。

In [188]: df.loc[pd.Interval(1, 2)]
Out[188]: 
A    2
Name: (1, 2], dtype: int64

尝试选择一个未完全包含在 IntervalIndex 中的 Interval 将会引发 KeyError

In [189]: df.loc[pd.Interval(0.5, 2.5)]
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
Cell In[189], line 1
----> 1 df.loc[pd.Interval(0.5, 2.5)]

File ~/work/pandas/pandas/pandas/core/indexing.py:1207, in _LocationIndexer.__getitem__(self, key)
   1205 maybe_callable = com.apply_if_callable(key, self.obj)
   1206 maybe_callable = self._raise_callable_usage(key, maybe_callable)
-> 1207 return self._getitem_axis(maybe_callable, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexing.py:1449, in _LocIndexer._getitem_axis(self, key, axis)
   1447 # fall thru to straight lookup
   1448 self._validate_key(key, axis)
-> 1449 return self._get_label(key, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexing.py:1399, in _LocIndexer._get_label(self, label, axis)
   1397 def _get_label(self, label, axis: AxisInt):
   1398     # GH#5567 this will fail if the label is not present in the axis.
-> 1399     return self.obj.xs(label, axis=axis)

File ~/work/pandas/pandas/pandas/core/generic.py:4253, in NDFrame.xs(self, key, axis, level, drop_level)
   4251             new_index = index[loc]
   4252 else:
-> 4253     loc = index.get_loc(key)
   4255     if isinstance(loc, np.ndarray):
   4256         if loc.dtype == np.bool_:

File ~/work/pandas/pandas/pandas/core/indexes/interval.py:814, in IntervalIndex.get_loc(self, key)
    812 matches = mask.sum()
    813 if matches == 0:
--> 814     raise KeyError(key)
    815 if matches == 1:
    816     return maybe_unbox_numpy_scalar(mask.argmax())

KeyError: Interval(0.5, 2.5, closed='right')

可以使用 overlaps() 方法来选择与给定 Interval 重叠的所有 Interval,以创建布尔索引器。

In [190]: idxr = df.index.overlaps(pd.Interval(0.5, 2.5))

In [191]: idxr
Out[191]: array([ True,  True,  True, False])

In [192]: df[idxr]
Out[192]: 
        A
(0, 1]  1
(1, 2]  2
(2, 3]  3

使用 cutqcut 对数据进行分箱#

cut()qcut() 都返回一个 Categorical 对象,并且它们创建的分箱存储在其 .categories 属性中,类型为 IntervalIndex

In [193]: c = pd.cut(range(4), bins=2)

In [194]: c
Out[194]: 
[(-0.003, 1.5], (-0.003, 1.5], (1.5, 3.0], (1.5, 3.0]]
Categories (2, interval[float64, right]): [(-0.003, 1.5] < (1.5, 3.0]]

In [195]: c.categories
Out[195]: IntervalIndex([(-0.003, 1.5], (1.5, 3.0]], dtype='interval[float64, right]')

cut() 还接受 IntervalIndex 作为其 bins 参数,这启用了一种有用的 pandas 惯用法。首先,我们使用一些数据调用 cut(),并将 bins 设置为固定数量,以生成分箱。然后,我们将 .categories 的值作为后续调用 cut()bins 参数,提供新数据将被分箱到相同的分箱中。

In [196]: pd.cut([0, 3, 5, 1], bins=c.categories)
Out[196]: 
[(-0.003, 1.5], (1.5, 3.0], NaN, (-0.003, 1.5]]
Categories (2, interval[float64, right]): [(-0.003, 1.5] < (1.5, 3.0]]

落在所有分箱之外的任何值都将被分配一个 NaN 值。

生成间隔范围#

如果我们需要在规则频率上生成间隔,我们可以使用 interval_range() 函数,通过 startendperiods 的各种组合来创建 IntervalIndexinterval_range 的默认频率对于数值间隔是 1,对于类似日期时间的间隔是日历天。

In [197]: pd.interval_range(start=0, end=5)
Out[197]: IntervalIndex([(0, 1], (1, 2], (2, 3], (3, 4], (4, 5]], dtype='interval[int64, right]')

In [198]: pd.interval_range(start=pd.Timestamp("2017-01-01"), periods=4)
Out[198]: 
IntervalIndex([(2017-01-01 00:00:00, 2017-01-02 00:00:00],
               (2017-01-02 00:00:00, 2017-01-03 00:00:00],
               (2017-01-03 00:00:00, 2017-01-04 00:00:00],
               (2017-01-04 00:00:00, 2017-01-05 00:00:00]],
              dtype='interval[datetime64[us], right]')

In [199]: pd.interval_range(end=pd.Timedelta("3 days"), periods=3)
Out[199]: 
IntervalIndex([(0 days 00:00:00, 1 days 00:00:00],
               (1 days 00:00:00, 2 days 00:00:00],
               (2 days 00:00:00, 3 days 00:00:00]],
              dtype='interval[timedelta64[us], right]')

freq 参数可用于指定非默认频率,并可利用各种 频率别名 来处理类似日期时间的间隔。

In [200]: pd.interval_range(start=0, periods=5, freq=1.5)
Out[200]: IntervalIndex([(0.0, 1.5], (1.5, 3.0], (3.0, 4.5], (4.5, 6.0], (6.0, 7.5]], dtype='interval[float64, right]')

In [201]: pd.interval_range(start=pd.Timestamp("2017-01-01"), periods=4, freq="W")
Out[201]: 
IntervalIndex([(2017-01-01 00:00:00, 2017-01-08 00:00:00],
               (2017-01-08 00:00:00, 2017-01-15 00:00:00],
               (2017-01-15 00:00:00, 2017-01-22 00:00:00],
               (2017-01-22 00:00:00, 2017-01-29 00:00:00]],
              dtype='interval[datetime64[us], right]')

In [202]: pd.interval_range(start=pd.Timedelta("0 days"), periods=3, freq="9h")
Out[202]: 
IntervalIndex([(0 days 00:00:00, 0 days 09:00:00],
               (0 days 09:00:00, 0 days 18:00:00],
               (0 days 18:00:00, 1 days 03:00:00]],
              dtype='interval[timedelta64[us], right]')

此外,closed 参数可用于指定间隔是封闭在哪个(哪些)侧。默认情况下,间隔是右侧封闭的。

In [203]: pd.interval_range(start=0, end=4, closed="both")
Out[203]: IntervalIndex([[0, 1], [1, 2], [2, 3], [3, 4]], dtype='interval[int64, both]')

In [204]: pd.interval_range(start=0, end=4, closed="neither")
Out[204]: IntervalIndex([(0, 1), (1, 2), (2, 3), (3, 4)], dtype='interval[int64, neither]')

指定 startendperiods 将从 startend(含)生成一系列等距间隔,在生成的 IntervalIndex 中包含 periods 个元素。

In [205]: pd.interval_range(start=0, end=6, periods=4)
Out[205]: IntervalIndex([(0.0, 1.5], (1.5, 3.0], (3.0, 4.5], (4.5, 6.0]], dtype='interval[float64, right]')

In [206]: pd.interval_range(pd.Timestamp("2018-01-01"), pd.Timestamp("2018-02-28"), periods=3)
Out[206]: 
IntervalIndex([(2018-01-01 00:00:00, 2018-01-20 08:00:00],
               (2018-01-20 08:00:00, 2018-02-08 16:00:00],
               (2018-02-08 16:00:00, 2018-02-28 00:00:00]],
              dtype='interval[datetime64[us], right]')

杂项索引常见问题解答#

整数索引#

具有整数轴标签的基于标签的索引是一个棘手的问题。它已经在邮件列表和科学 Python 社区的各个成员之间进行了广泛讨论。在 pandas 中,我们的普遍观点是标签比整数位置更重要。因此,对于只有整数轴索引的情况,使用 .loc 等标准工具只能进行基于标签的索引。以下代码将引发异常:

In [207]: s = pd.Series(range(5))

In [208]: s[-1]
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
File ~/work/pandas/pandas/pandas/core/indexes/range.py:521, in RangeIndex.get_loc(self, key)
    520 try:
--> 521     return self._range.index(new_key)
    522 except ValueError as err:

ValueError: -1 is not in range

The above exception was the direct cause of the following exception:

KeyError                                  Traceback (most recent call last)
Cell In[208], line 1
----> 1 s[-1]

File ~/work/pandas/pandas/pandas/core/series.py:959, in Series.__getitem__(self, key)
    954     key = unpack_1tuple(key)
    956 elif key_is_scalar:
    957     # Note: GH#50617 in 3.0 we changed int key to always be treated as
    958     #  a label, matching DataFrame behavior.
--> 959     return self._get_value(key)
    961 # Convert generator to list before going through hashable part
    962 # (We will iterate through the generator there to check for slices)
    963 if is_iterator(key):

File ~/work/pandas/pandas/pandas/core/series.py:1046, in Series._get_value(self, label, takeable)
   1043     return self._values[label]
   1045 # Similar to Index.get_value, but we do not fall back to positional
-> 1046 loc = self.index.get_loc(label)
   1048 if is_integer(loc):
   1049     return self._values[loc]

File ~/work/pandas/pandas/pandas/core/indexes/range.py:523, in RangeIndex.get_loc(self, key)
    521         return self._range.index(new_key)
    522     except ValueError as err:
--> 523         raise KeyError(key) from err
    524 if isinstance(key, Hashable):
    525     raise KeyError(key)

KeyError: -1

In [209]: df = pd.DataFrame(np.random.randn(5, 4))

In [210]: df
Out[210]: 
          0         1         2         3
0 -0.435772 -1.188928 -0.808286 -0.284634
1 -1.815703  1.347213 -0.243487  0.514704
2  1.162969 -0.287725 -0.179734  0.993962
3 -0.212673  0.909872 -0.733333 -0.349893
4  0.456434 -0.306735  0.553396  0.166221

In [211]: df.loc[-2:]
Out[211]: 
          0         1         2         3
0 -0.435772 -1.188928 -0.808286 -0.284634
1 -1.815703  1.347213 -0.243487  0.514704
2  1.162969 -0.287725 -0.179734  0.993962
3 -0.212673  0.909872 -0.733333 -0.349893
4  0.456434 -0.306735  0.553396  0.166221

这一刻意的决定是为了防止歧义和细微的错误(许多用户在 API 更改为停止“回退”到基于位置的索引时报告了发现错误)。

非单调索引需要精确匹配#

如果 SeriesDataFrame 的索引是单调递增或递减的,那么基于标签的切片边界可以超出索引的范围,这与切片标准的 Python list 类似。索引的单调性可以使用 is_monotonic_increasing()is_monotonic_decreasing() 属性进行测试。

In [212]: df = pd.DataFrame(index=[2, 3, 3, 4, 5], columns=["data"], data=list(range(5)))

In [213]: df.index.is_monotonic_increasing
Out[213]: True

# no rows 0 or 1, but still returns rows 2, 3 (both of them), and 4:
In [214]: df.loc[0:4, :]
Out[214]: 
   data
2     0
3     1
3     2
4     3

# slice is are outside the index, so empty DataFrame is returned
In [215]: df.loc[13:15, :]
Out[215]: 
Empty DataFrame
Columns: [data]
Index: []

另一方面,如果索引不是单调的,则切片边界必须是索引的唯一成员。

In [216]: df = pd.DataFrame(index=[2, 3, 1, 4, 3, 5], columns=["data"], data=list(range(6)))

In [217]: df.index.is_monotonic_increasing
Out[217]: False

# OK because 2 and 4 are in the index
In [218]: df.loc[2:4, :]
Out[218]: 
   data
2     0
3     1
1     2
4     3
 # 0 is not in the index
In [219]: df.loc[0:4, :]
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
Cell In[219], line 1
----> 1 df.loc[0:4, :]

File ~/work/pandas/pandas/pandas/core/indexing.py:1200, in _LocationIndexer.__getitem__(self, key)
   1198     if self._is_scalar_access(key):
   1199         return self.obj._get_value(*key, takeable=self._takeable)
-> 1200     return self._getitem_tuple(key)
   1201 else:
   1202     # we by definition only have the 0th axis
   1203     axis = self.axis or 0

File ~/work/pandas/pandas/pandas/core/indexing.py:1395, in _LocIndexer._getitem_tuple(self, tup)
   1392 if self._multi_take_opportunity(tup):
   1393     return self._multi_take(tup)
-> 1395 return self._getitem_tuple_same_dim(tup)

File ~/work/pandas/pandas/pandas/core/indexing.py:1046, in _LocationIndexer._getitem_tuple_same_dim(self, tup)
   1043 if com.is_null_slice(key):
   1044     continue
-> 1046 retval = getattr(retval, self.name)._getitem_axis(key, axis=i)
   1047 # We should never have retval.ndim < self.ndim, as that should
   1048 #  be handled by the _getitem_lowerdim call above.
   1049 assert retval.ndim == self.ndim

File ~/work/pandas/pandas/pandas/core/indexing.py:1429, in _LocIndexer._getitem_axis(self, key, axis)
   1427 if isinstance(key, slice):
   1428     self._validate_key(key, axis)
-> 1429     return self._get_slice_axis(key, axis=axis)
   1430 elif com.is_bool_indexer(key):
   1431     return self._getbool_axis(key, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexing.py:1461, in _LocIndexer._get_slice_axis(self, slice_obj, axis)
   1458     return obj.copy(deep=False)
   1460 labels = obj._get_axis(axis)
-> 1461 indexer = labels.slice_indexer(slice_obj.start, slice_obj.stop, slice_obj.step)
   1463 if isinstance(indexer, slice):
   1464     return self.obj._slice(indexer, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexes/base.py:6804, in Index.slice_indexer(self, start, end, step)
   6753 def slice_indexer(
   6754     self,
   6755     start: Hashable | None = None,
   6756     end: Hashable | None = None,
   6757     step: int | None = None,
   6758 ) -> slice:
   6759     """
   6760     Compute the slice indexer for input labels and step.
   6761 
   (...)   6802     slice(1, 3, None)
   6803     """
-> 6804     start_slice, end_slice = self.slice_locs(start, end, step=step)
   6806     # return a slice
   6807     if not is_scalar(start_slice):

File ~/work/pandas/pandas/pandas/core/indexes/base.py:7062, in Index.slice_locs(self, start, end, step)
   7060 start_slice = None
   7061 if start is not None:
-> 7062     start_slice = self.get_slice_bound(start, "left")
   7063 if start_slice is None:
   7064     start_slice = 0

File ~/work/pandas/pandas/pandas/core/indexes/base.py:6973, in Index.get_slice_bound(self, label, side)
   6971         return self._searchsorted_monotonic(label, side)
   6972     except ValueError:
-> 6973         raise KeyError(
   6974             f"Cannot get {side} slice bound for non-monotonic index "
   6975             f"with a missing label {original_label!r}. "
   6976             "Either sort the index or specify an existing label."
   6977         ) from None
   6979 if isinstance(slc, np.ndarray):
   6980     # get_loc may return a boolean array, which
   6981     # is OK as long as they are representable by a slice.
   6982     assert is_bool_dtype(slc.dtype)

KeyError: 'Cannot get left slice bound for non-monotonic index with a missing label 0. Either sort the index or specify an existing label.'

 # 3 is not a unique label
In [220]: df.loc[2:3, :]
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
Cell In[220], line 1
----> 1 df.loc[2:3, :]

File ~/work/pandas/pandas/pandas/core/indexing.py:1200, in _LocationIndexer.__getitem__(self, key)
   1198     if self._is_scalar_access(key):
   1199         return self.obj._get_value(*key, takeable=self._takeable)
-> 1200     return self._getitem_tuple(key)
   1201 else:
   1202     # we by definition only have the 0th axis
   1203     axis = self.axis or 0

File ~/work/pandas/pandas/pandas/core/indexing.py:1395, in _LocIndexer._getitem_tuple(self, tup)
   1392 if self._multi_take_opportunity(tup):
   1393     return self._multi_take(tup)
-> 1395 return self._getitem_tuple_same_dim(tup)

File ~/work/pandas/pandas/pandas/core/indexing.py:1046, in _LocationIndexer._getitem_tuple_same_dim(self, tup)
   1043 if com.is_null_slice(key):
   1044     continue
-> 1046 retval = getattr(retval, self.name)._getitem_axis(key, axis=i)
   1047 # We should never have retval.ndim < self.ndim, as that should
   1048 #  be handled by the _getitem_lowerdim call above.
   1049 assert retval.ndim == self.ndim

File ~/work/pandas/pandas/pandas/core/indexing.py:1429, in _LocIndexer._getitem_axis(self, key, axis)
   1427 if isinstance(key, slice):
   1428     self._validate_key(key, axis)
-> 1429     return self._get_slice_axis(key, axis=axis)
   1430 elif com.is_bool_indexer(key):
   1431     return self._getbool_axis(key, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexing.py:1461, in _LocIndexer._get_slice_axis(self, slice_obj, axis)
   1458     return obj.copy(deep=False)
   1460 labels = obj._get_axis(axis)
-> 1461 indexer = labels.slice_indexer(slice_obj.start, slice_obj.stop, slice_obj.step)
   1463 if isinstance(indexer, slice):
   1464     return self.obj._slice(indexer, axis=axis)

File ~/work/pandas/pandas/pandas/core/indexes/base.py:6804, in Index.slice_indexer(self, start, end, step)
   6753 def slice_indexer(
   6754     self,
   6755     start: Hashable | None = None,
   6756     end: Hashable | None = None,
   6757     step: int | None = None,
   6758 ) -> slice:
   6759     """
   6760     Compute the slice indexer for input labels and step.
   6761 
   (...)   6802     slice(1, 3, None)
   6803     """
-> 6804     start_slice, end_slice = self.slice_locs(start, end, step=step)
   6806     # return a slice
   6807     if not is_scalar(start_slice):

File ~/work/pandas/pandas/pandas/core/indexes/base.py:7068, in Index.slice_locs(self, start, end, step)
   7066 end_slice = None
   7067 if end is not None:
-> 7068     end_slice = self.get_slice_bound(end, "right")
   7069 if end_slice is None:
   7070     end_slice = len(self)

File ~/work/pandas/pandas/pandas/core/indexes/base.py:6985, in Index.get_slice_bound(self, label, side)
   6983     slc = lib.maybe_booleans_to_slice(slc.view("u1"))
   6984     if isinstance(slc, np.ndarray):
-> 6985         raise KeyError(
   6986             f"Cannot get {side} slice bound for non-unique "
   6987             f"label: {original_label!r}"
   6988         )
   6990 if isinstance(slc, slice):
   6991     if side == "left":

KeyError: 'Cannot get right slice bound for non-unique label: 3'

Index.is_monotonic_increasingIndex.is_monotonic_decreasing 仅检查索引是否是弱单调的。要检查严格单调性,您可以将其中一个与 is_unique() 属性结合使用。

In [221]: weakly_monotonic = pd.Index(["a", "b", "c", "c"])

In [222]: weakly_monotonic
Out[222]: Index(['a', 'b', 'c', 'c'], dtype='str')

In [223]: weakly_monotonic.is_monotonic_increasing
Out[223]: True

In [224]: weakly_monotonic.is_monotonic_increasing & weakly_monotonic.is_unique
Out[224]: False

端点包含在内#

与切片端点不包含在内的标准 Python 序列切片相比,pandas 中的基于标签的切片是包含端点的。这样做的主要原因是,通常无法轻松确定索引中特定标签的“后继”或下一个元素。例如,考虑以下 Series

In [225]: s = pd.Series(np.random.randn(6), index=list("abcdef"))

In [226]: s
Out[226]: 
a   -0.101684
b   -0.734907
c   -0.130121
d   -0.476046
e    0.759104
f    0.213379
dtype: float64

假设我们希望将切片从 ce,使用整数可以这样完成:

In [227]: s[2:5]
Out[227]: 
c   -0.130121
d   -0.476046
e    0.759104
dtype: float64

但是,如果您只有 ce,确定索引中的下一个元素可能会有些复杂。例如,以下代码无效:

In [228]: s.loc['c':'e' + 1]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[228], line 1
----> 1 s.loc['c':'e' + 1]

TypeError: can only concatenate str (not "int") to str

一个非常常见的用例是将时间序列限制在两个特定日期之间开始和结束。为了实现这一点,我们做出了一个设计选择,即使基于标签的切片包含两个端点。

In [229]: s.loc["c":"e"]
Out[229]: 
c   -0.130121
d   -0.476046
e    0.759104
dtype: float64

这绝对是“实用性胜过纯粹性”的原则,但如果您期望基于标签的切片行为与标准 Python 整数切片完全一致,那么这一点值得注意。

索引可能会改变底层 Series 的 dtype#

不同的索引操作可能会改变 Series 的 dtype。

In [230]: series1 = pd.Series([1, 2, 3])

In [231]: series1.dtype
Out[231]: dtype('int64')

In [232]: res = series1.reindex([0, 4])

In [233]: res.dtype
Out[233]: dtype('float64')

In [234]: res
Out[234]: 
0    1.0
4    NaN
dtype: float64
In [235]: series2 = pd.Series([True])

In [236]: series2.dtype
Out[236]: dtype('bool')

In [237]: res = series2.reindex_like(series1)

In [238]: res.dtype
Out[238]: dtype('O')

In [239]: res
Out[239]: 
0    True
1     NaN
2     NaN
dtype: object

这是因为上述(重新)索引操作会默默地插入 NaN,并且 dtype 会相应地更改。这在使用 numpy ufuncs(如 numpy.logical_and)时可能会导致一些问题。

有关更详细的讨论,请参阅 GH 2388