不管是为 pandas 对象应用自定义函数,还是应用其它第三方函数,都离不开以下三种方法。用哪种方法取决于操作的对象是 DataFrame
或 Series
,是行或列,还是元素。
apply()
虽然可以把 DataFrame
与 Series
传递给函数。不过,通过链式调用函数时,最好使用 pipe()
方法。对比以下两种方式:
# f, g, and h are functions taking and returning ``DataFrames``
>>> f(g(h(df), arg1=1), arg2=2, arg3=3)
下列代码与上述代码等效
>>> (df.pipe(h)
... .pipe(g, arg1=1)
... .pipe(f, arg2=2, arg3=3))
pandas 鼓励使用第二种方式,即链式方法。在链式方法中调用自定义函数或第三方支持库函数时,用 pipe
更容易,与用 pandas 自身方法一样。
上例中,f
、g
与 h
这几个函数都把 DataFrame
当作首位参数。要是想把数据作为第二个参数,该怎么办?本例中,pipe
为元组 (callable,data_keyword
)形式。.pipe
把 DataFrame
作为元组里指定的参数。
下例用 statsmodels 拟合回归。该 API 先接收一个公式,DataFrame
是第二个参数,data
。要传递函数,则要用pipe
接收关键词对 (sm.ols,'data'
)。
In [138]: import statsmodels.formula.api as sm
In [139]: bb = pd.read_csv('data/baseball.csv', index_col='id')
In [140]: (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[140]:
<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: Thu, 22 Aug 2019 Prob (F-statistic): 3.48e-15
Time: 15:48:59 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
==============================================================================
Warnings:
[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.
unix 的 pipe
与后来出现的 dplyr 及 magrittr 启发了pipe
方法,在此,引入了 R 语言里用于读取 pipe 的操作符 (%>%
)。pipe
的实现思路非常清晰,仿佛 Python 源生的一样。强烈建议大家阅读 pipe()
的源代码。
apply()
方法可以沿着 DataFrame 的轴应用任何函数,比如,描述性统计方法,该方法支持 axis
参数。
In [141]: df.apply(np.mean)
Out[141]:
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [142]: df.apply(np.mean, axis=1)
Out[142]:
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64
In [143]: df.apply(lambda x: x.max() - x.min())
Out[143]:
one 1.051928
two 1.632779
three 1.840607
dtype: float64
In [144]: df.apply(np.cumsum)
Out[144]:
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 [145]: df.apply(np.exp)
Out[145]:
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
apply()
方法还支持通过函数名字符串调用函数。
In [146]: df.apply('mean')
Out[146]:
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [147]: df.apply('mean', axis=1)
Out[147]:
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64
默认情况下,apply()
调用的函数返回的类型会影响 DataFrame.apply
输出结果的类型。
Series
时,最终输出的结果是 DataFrame
。输出的列与函数返回的 Series
索引相匹配。Series
。result_type
会覆盖默认行为,该参数有三个选项:reduce
、broadcast
、expand
。这些选项决定了列表型返回值是否扩展为 DataFrame
。
用好 apply()
可以了解数据集的很多信息。比如可以提取每列的最大值对应的日期:
In [148]: tsdf = pd.DataFrame(np.random.randn(1000, 3), columns=['A', 'B', 'C'],
.....: index=pd.date_range('1/1/2000', periods=1000))
.....:
In [149]: tsdf.apply(lambda x: x.idxmax())
Out[149]:
A 2000-08-06
B 2001-01-18
C 2001-07-18
dtype: datetime64[ns]
还可以向 apply()
方法传递额外的参数与关键字参数。比如下例中要应用的这个函数:
def subtract_and_divide(x, sub, divide=1):
return (x - sub) / divide
可以用下列方式应用该函数:
df.apply(subtract_and_divide, args=(5,), divide=3)
为每行或每列执行 Series
方法的功能也很实用:
In [150]: tsdf
Out[150]:
A B C
2000-01-01 -0.158131 -0.232466 0.321604
2000-01-02 -1.810340 -3.105758 0.433834
2000-01-03 -1.209847 -1.156793 -0.136794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 -0.653602 0.178875 1.008298
2000-01-09 1.007996 0.462824 0.254472
2000-01-10 0.307473 0.600337 1.643950
In [151]: tsdf.apply(pd.Series.interpolate)
Out[151]:
A B C
2000-01-01 -0.158131 -0.232466 0.321604
2000-01-02 -1.810340 -3.105758 0.433834
2000-01-03 -1.209847 -1.156793 -0.136794
2000-01-04 -1.098598 -0.889659 0.092225
2000-01-05 -0.987349 -0.622526 0.321243
2000-01-06 -0.876100 -0.355392 0.550262
2000-01-07 -0.764851 -0.088259 0.779280
2000-01-08 -0.653602 0.178875 1.008298
2000-01-09 1.007996 0.462824 0.254472
2000-01-10 0.307473 0.600337 1.643950
apply()
有一个参数 raw
,默认值为 False
,在应用函数前,使用该参数可以将每行或列转换为 Series
。该参数为 True
时,传递的函数接收 ndarray 对象,若不需要索引功能,这种操作能显著提高性能。
0.20.0 版新增。
聚合 API 可以快速、简洁地执行多个聚合操作。Pandas 对象支持多个类似的 API,如 groupby API、window functions API、resample API。聚合函数为DataFrame.aggregate()
,它的别名是 DataFrame.agg()
。
这里使用与前例类似的 DataFrame
:
In [152]: tsdf = pd.DataFrame(np.random.randn(10, 3), columns=['A', 'B', 'C'],
.....: index=pd.date_range('1/1/2000', periods=10))
.....:
In [153]: tsdf.iloc[3:7] = np.nan
In [154]: tsdf
Out[154]:
A B C
2000-01-01 1.257606 1.004194 0.167574
2000-01-02 -0.749892 0.288112 -0.757304
2000-01-03 -0.207550 -0.298599 0.116018
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.814347 -0.257623 0.869226
2000-01-09 -0.250663 -1.206601 0.896839
2000-01-10 2.169758 -1.333363 0.283157
应用单个函数时,该操作与 apply()
等效,这里也可以用字符串表示聚合函数名。下面的聚合函数输出的结果为 Series
:
In [155]: tsdf.agg(np.sum)
Out[155]:
A 3.033606
B -1.803879
C 1.575510
dtype: float64
In [156]: tsdf.agg('sum')
Out[156]:
A 3.033606
B -1.803879
C 1.575510
dtype: float64
# 因为应用的是单个函数,该操作与`.sum()` 是等效的
In [157]: tsdf.sum()
Out[157]:
A 3.033606
B -1.803879
C 1.575510
dtype: float64
对 Series
进行单个聚合操作,返回的是标量值:
In [158]: tsdf.A.agg('sum')
Out[158]: 3.033606102414146
还可以用列表形式传递多个聚合函数。每个函数在输出结果 DataFrame
里以行的形式显示,行名是每个聚合函数的函数名。
In [159]: tsdf.agg(['sum'])
Out[159]:
A B C
sum 3.033606 -1.803879 1.57551
多个函数输出多行:
In [160]: tsdf.agg(['sum', 'mean'])
Out[160]:
A B C
sum 3.033606 -1.803879 1.575510
mean 0.505601 -0.300647 0.262585
对于 Series
,多个函数返回的结果也是 Series
,其索引为函数名:
In [161]: tsdf.A.agg(['sum', 'mean'])
Out[161]:
sum 3.033606
mean 0.505601
Name: A, dtype: float64
传递 lambda
函数时,输出名为 <lambda>
的行:
In [162]: tsdf.A.agg(['sum', lambda x: x.mean()])
Out[162]:
sum 3.033606
<lambda> 0.505601
Name: A, dtype: float64
应用自定义函数时,则该函数名为输出结果的行名:
In [163]: def mymean(x):
.....: return x.mean()
.....:
In [164]: tsdf.A.agg(['sum', mymean])
Out[164]:
sum 3.033606
mymean 0.505601
Name: A, dtype: float64
指定为哪些列应用哪些聚合函数时,需要把包含列名与标量(或标量列表)的字典传递给 DataFrame.agg
。
注意:这里输出结果的顺序不是固定的,要想让输出顺序与输入顺序一致,请使用 OrderedDict
。
In [165]: tsdf.agg({'A': 'mean', 'B': 'sum'})
Out[165]:
A 0.505601
B -1.803879
dtype: float64
输入的参数是列表时,输出结果为 DataFrame
,并以矩阵形式显示所有聚合函数的计算结果,且输出结果由所有唯一函数组成。未执行聚合操作的列输出结果为 NaN
值:
In [166]: tsdf.agg({'A': ['mean', 'min'], 'B': 'sum'})
Out[166]:
A B
mean 0.505601 NaN
min -0.749892 NaN
sum NaN -1.803879
DataFrame
里包含不能执行聚合操作的多种 Dtype 时,.agg
只计算可以执行聚合的列。这与 groupby
的 .agg
操作类似:
In [167]: mdf = pd.DataFrame({'A': [1, 2, 3],
.....: 'B': [1., 2., 3.],
.....: 'C': ['foo', 'bar', 'baz'],
.....: 'D': pd.date_range('20130101', periods=3)})
.....:
In [168]: mdf.dtypes
Out[168]:
A int64
B float64
C object
D datetime64[ns]
dtype: object
In [169]: mdf.agg(['min', 'sum'])
Out[169]:
A B C D
min 1 1.0 bar 2013-01-01
sum 6 6.0 foobarbaz NaT
用 .agg()
可以轻松地创建与内置 describe 函数类似的自定义 describe 函数。
In [170]: from functools import partial
In [171]: q_25 = partial(pd.Series.quantile, q=0.25)
In [172]: q_25.__name__ = '25%'
In [173]: q_75 = partial(pd.Series.quantile, q=0.75)
In [174]: q_75.__name__ = '75%'
In [175]: tsdf.agg(['count', 'mean', 'std', 'min', q_25, 'median', q_75, 'max'])
Out[175]:
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
0.20.0 版新增。
transform()
方法返回的结果与原始数据具有同样索引,且大小相同。这个 API 支持同时处理多种操作,不用一个一个操作,且该 API 与 .agg
API 类似。
下面先创建一个 DataFrame:
In [176]: tsdf = pd.DataFrame(np.random.randn(10, 3), columns=['A', 'B', 'C'],
.....: index=pd.date_range('1/1/2000', periods=10))
.....:
In [177]: tsdf.iloc[3:7] = np.nan
In [178]: tsdf
Out[178]:
A B C
2000-01-01 -0.428759 -0.864890 -0.675341
2000-01-02 -0.168731 1.338144 -1.279321
2000-01-03 -1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 -1.240447 -0.201052
2000-01-09 -0.157795 0.791197 -1.144209
2000-01-10 -0.030876 0.371900 0.061932
这里转换的是整个 DataFrame。.transform()
支持 Numpy 函数、字符串函数及自定义函数。
In [179]: tsdf.transform(np.abs)
Out[179]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
In [180]: tsdf.transform('abs')
Out[180]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
In [181]: tsdf.transform(lambda x: x.abs())
Out[181]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
这里的 transform()
接受单个函数;与 ufunc 等效。
In [182]: np.abs(tsdf)
Out[182]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
.transform()
向 Series
传递单个函数时,返回的结果也是单个 Series
。
In [183]: tsdf.A.transform(np.abs)
Out[183]:
2000-01-01 0.428759
2000-01-02 0.168731
2000-01-03 1.621034
2000-01-04 NaN
2000-01-05 NaN
2000-01-06 NaN
2000-01-07 NaN
2000-01-08 0.254374
2000-01-09 0.157795
2000-01-10 0.030876
Freq: D, Name: A, dtype: float64
transform()
调用多个函数时,将生成多重索引 DataFrame。第一层是原始数据集的列名;第二层是 transform()
调用的函数名。
In [184]: tsdf.transform([np.abs, lambda x: x + 1])
Out[184]:
A B C
absolute <lambda> absolute <lambda> absolute <lambda>
2000-01-01 0.428759 0.571241 0.864890 0.135110 0.675341 0.324659
2000-01-02 0.168731 0.831269 1.338144 2.338144 1.279321 -0.279321
2000-01-03 1.621034 -0.621034 0.438107 1.438107 0.903794 1.903794
2000-01-04 NaN NaN NaN NaN NaN NaN
2000-01-05 NaN NaN NaN NaN NaN NaN
2000-01-06 NaN NaN NaN NaN NaN NaN
2000-01-07 NaN NaN NaN NaN NaN NaN
2000-01-08 0.254374 1.254374 1.240447 -0.240447 0.201052 0.798948
2000-01-09 0.157795 0.842205 0.791197 1.791197 1.144209 -0.144209
2000-01-10 0.030876 0.969124 0.371900 1.371900 0.061932 1.061932
为 Series 应用多个函数时,输出结果是 DataFrame,列名是 transform()
调用的函数名。
In [185]: tsdf.A.transform([np.abs, lambda x: x + 1])
Out[185]:
absolute <lambda>
2000-01-01 0.428759 0.571241
2000-01-02 0.168731 0.831269
2000-01-03 1.621034 -0.621034
2000-01-04 NaN NaN
2000-01-05 NaN NaN
2000-01-06 NaN NaN
2000-01-07 NaN NaN
2000-01-08 0.254374 1.254374
2000-01-09 0.157795 0.842205
2000-01-10 0.030876 0.969124
函数字典可以为每列执行指定 transform()
操作。
In [186]: tsdf.transform({'A': np.abs, 'B': lambda x: x + 1})
Out[186]:
A B
2000-01-01 0.428759 0.135110
2000-01-02 0.168731 2.338144
2000-01-03 1.621034 1.438107
2000-01-04 NaN NaN
2000-01-05 NaN NaN
2000-01-06 NaN NaN
2000-01-07 NaN NaN
2000-01-08 0.254374 -0.240447
2000-01-09 0.157795 1.791197
2000-01-10 0.030876 1.371900
transform()
的参数是列表字典时,生成的是以 transform()
调用的函数为名的多重索引 DataFrame。
In [187]: tsdf.transform({'A': np.abs, 'B': [lambda x: x + 1, 'sqrt']})
Out[187]:
A B
absolute <lambda> sqrt
2000-01-01 0.428759 0.135110 NaN
2000-01-02 0.168731 2.338144 1.156782
2000-01-03 1.621034 1.438107 0.661897
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 -0.240447 NaN
2000-01-09 0.157795 1.791197 0.889493
2000-01-10 0.030876 1.371900 0.609836
并非所有函数都能矢量化,即接受 Numpy 数组,返回另一个数组或值,DataFrame 的 applymap()
及 Series 的 map()
,支持任何接收单个值并返回单个值的 Python 函数。
示例如下:
In [188]: df4
Out[188]:
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 [189]: def f(x):
.....: return len(str(x))
.....:
In [190]: df4['one'].map(f)
Out[190]:
a 18
b 19
c 18
d 3
Name: one, dtype: int64
In [191]: df4.applymap(f)
Out[191]:
one two three
a 18 17 3
b 19 18 20
c 18 18 16
d 3 19 19
Series.map()
还有个功能,可以“连接”或“映射”第二个 Series 定义的值。这与 merging/joining 功能联系非常紧密:
In [192]: s = pd.Series(['six', 'seven', 'six', 'seven', 'six'],
.....: index=['a', 'b', 'c', 'd', 'e'])
.....:
In [193]: t = pd.Series({'six': 6., 'seven': 7.})
In [194]: s
Out[194]:
a six
b seven
c six
d seven
e six
dtype: object
In [195]: s.map(t)
Out[195]:
a 6.0
b 7.0
c 6.0
d 7.0
e 6.0
dtype: float64