熊猫：每60秒箱中仅保留第一行数据

Question

在pandas中只保留每个60秒数据库的第一行的最佳方法是什么？即对于在增加的时间t发生的每一行，我想删除最多t+60秒出现的所有行。

我知道我可能会使用groupby().first()的某种组合，但我见过的代码示例（例如使用pandas.Grouper(freq='60s')）将丢弃原始日期时间，而不是每60秒偏移一次午夜而不是我原来的约会时间。

例如，以下内容：

                            time        value
0  2016-05-11 13:00:10.841015028     0.215978
1  2016-05-11 13:02:05.760595780     0.155666
2  2016-05-11 13:02:05.760903860     0.155666
3  2016-05-11 13:02:18.325613076     0.157788
4  2016-05-11 13:02:18.486519052     0.157788
5  2016-05-11 13:02:20.243748548     0.157788
6  2016-05-11 13:02:20.533101692     0.157788
7  2016-05-11 13:02:20.646061652     0.157788
8  2016-05-11 13:02:21.121409820     0.157788
9  2016-05-11 13:04:24.660609068     0.211649
10 2016-05-11 13:04:24.660845612     0.211649
11 2016-05-11 13:04:24.660957596     0.211649
12 2016-05-11 13:04:24.661378132     0.211649
13 2016-05-11 13:04:24.661450628     0.211649
14 2016-05-11 13:04:24.661607044     0.211649

应该成为这个：

                            time        value
0  2016-05-11 13:00:10.841015028     0.215978
1  2016-05-11 13:02:05.760595780     0.155666
3  2016-05-11 13:04:24.660609068     0.211649

Answer 1

更新：感谢@piRSquared - 他发现我之前的解决方案不正确。这是另一种尝试：

数据：

In [8]: df = pd.DataFrame(dict(time=pd.date_range('2001-01-01', periods=20, freq='9S'), value=np.random.rand(20)))

In [9]: df
Out[9]:
                  time     value
0  2001-01-01 00:00:00  0.440696
1  2001-01-01 00:00:09  0.135540
2  2001-01-01 00:00:18  0.008243
3  2001-01-01 00:00:27  0.389259
4  2001-01-01 00:00:36  0.128253
5  2001-01-01 00:00:45  0.566704
6  2001-01-01 00:00:54  0.386797
7  2001-01-01 00:01:03  0.426411
8  2001-01-01 00:01:12  0.438114
9  2001-01-01 00:01:21  0.918711
10 2001-01-01 00:01:30  0.715565
11 2001-01-01 00:01:39  0.422044
12 2001-01-01 00:01:48  0.199396
13 2001-01-01 00:01:57  0.827872
14 2001-01-01 00:02:06  0.986887
15 2001-01-01 00:02:15  0.305749
16 2001-01-01 00:02:24  0.030092
17 2001-01-01 00:02:33  0.338214
18 2001-01-01 00:02:42  0.773635
19 2001-01-01 00:02:51  0.816478

解决方案：

In [10]: df.groupby((df.time - df.loc[0, 'time']).dt.total_seconds() // 60, as_index=False).first()
Out[10]:
                 time     value
0 2001-01-01 00:00:00  0.440696
1 2001-01-01 00:01:03  0.426411
2 2001-01-01 00:02:06  0.986887

说明：

In [17]: (df.time - df.loc[0, 'time']).dt.total_seconds()
Out[17]:
0       0.0
1       9.0
2      18.0
3      27.0
4      36.0
5      45.0
6      54.0
7      63.0
8      72.0
9      81.0
10     90.0
11     99.0
12    108.0
13    117.0
14    126.0
15    135.0
16    144.0
17    153.0
18    162.0
19    171.0
Name: time, dtype: float64

In [18]: (df.time - df.loc[0, 'time']).dt.total_seconds() // 60
Out[18]:
0    -0.0
1     0.0
2     0.0
3     0.0
4     0.0
5     0.0
6     0.0
7     1.0
8     1.0
9     1.0
10    1.0
11    1.0
12    1.0
13    1.0
14    2.0
15    2.0
16    2.0
17    2.0
18    2.0
19    2.0
Name: time, dtype: float64

OLD错误回答：

In [102]: df[df.time.diff().fillna(pd.Timedelta('60S')) >= pd.Timedelta('60S')]
Out[102]:
                           time     value
0 2016-05-11 13:00:10.841015028  0.215978
1 2016-05-11 13:02:05.760595780  0.155666
9 2016-05-11 13:04:24.660609068  0.211649

说明：

Answer 2

解决方案

def td60(ta):
    d = np.timedelta64(int(6e10))
    tp = ta + d
    j = 0
    yield j
    for i, tx in enumerate(ta):
        if tx > tp[j]:
            yield i
            j = i

def pir(df):
    slc = list(td60(df.time.values))
    return pd.DataFrame(df.values[slc], df.index[slc])

示例用法

pir(df)

设置时间500,000行

pop_n, smp_n = 1000000, 500000
np.random.seed([3,1415])
tidx = pd.date_range('2016-09-08', periods=pop_n, freq='5s')
tidx = np.random.choice(tidx, smp_n, False)
tidx = pd.to_datetime(tidx).sort_values()

df = pd.DataFrame(dict(time=tidx, value=np.random.rand(smp_n)))

时序

<强> Cythonize
在Jupyter

%load_ext Cython

%%cython
import numpy as np
import pandas as pd

def td60(ta):
    d = np.timedelta64(int(6e10))
    tp = ta + d
    j = 0
    yield j
    for i, tx in enumerate(ta):
        if tx > tp[j]:
            yield i
            j = i

def pir(df):
    slc = list(td60(df.time.values))
    return pd.DataFrame(df.values[slc], df.index[slc])

Cythonizing后
差别不大

OP示例的参考设置

from StringIO import StringIO
import pandas as pd

text = """time,value
2016-05-11 13:00:10.841015028,0.215978
2016-05-11 13:02:05.760595780,0.155666
2016-05-11 13:02:05.760903860,0.155666
2016-05-11 13:02:18.325613076,0.157788
2016-05-11 13:02:18.486519052,0.157788
2016-05-11 13:02:20.243748548,0.157788
2016-05-11 13:02:20.533101692,0.157788
2016-05-11 13:02:20.646061652,0.157788
2016-05-11 13:02:21.121409820,0.157788
2016-05-11 13:04:24.660609068,0.211649
2016-05-11 13:04:24.660845612,0.211649
2016-05-11 13:04:24.660957596,0.211649
2016-05-11 13:04:24.661378132,0.211649
2016-05-11 13:04:24.661450628,0.211649
2016-05-11 13:04:24.661607044,0.211649"""

df = pd.read_csv(StringIO(text), parse_dates=[0])