设置parse_date = ['列名']时,pd.read_csv无法正确解析日期/月份字段

Question

我试图通过parse_dates解析几个日期时遇到了这个bug pandas.read_csv().在下面的代码片段中,我试图解析具有dd/mm/yy导致我转换不正确的格式的日期.在某些情况下,日期字段被视为月份,反之亦然.

为了简单起见,有些情况下dd/mm/yy转换为yyyy-dd-mm而不是yyyy-mm-dd.

情况1:

  04/10/96 is parsed as 1996-04-10, which is wrong.

案例2:

  15/07/97 is parsed as 1997-07-15, which is correct.

案例3:

  10/12/97 is parsed as 1997-10-12, which is wrong.

代码示例

import pandas as pd

df = pd.read_csv('date_time.csv') 
print 'Data in csv:'
print df
print df['start_date'].dtypes

print '----------------------------------------------'

df = pd.read_csv('date_time.csv', parse_dates = ['start_date'])
print 'Data after parsing:'
print df
print df['start_date'].dtypes

电流输出

----------------------
Data in csv:
----------------------
  start_date
0   04/10/96
1   15/07/97
2   10/12/97
3   06/03/99
4     //1994
5   /02/1967
object
----------------------
Data after parsing:
----------------------
   start_date
0 1996-04-10
1 1997-07-15
2 1997-10-12
3 1999-06-03
4 1994-01-01
5 1967-02-01
datetime64[ns]

预期产出

----------------------
Data in csv:
----------------------
   start_date
0   04/10/96
1   15/07/97
2   10/12/97
3   06/03/99
4     //1994
5   /02/1967
object
----------------------
Data after parsing:
----------------------
  start_date

0 1996-10-04
1 1997-07-15
2 1997-12-10
3 1999-03-06
4 1994-01-01
5 1967-02-01
datetime64[ns]

更多评论:

我可以使用date_parser或pandas.to_datetime()指定适当的日期格式.但在我的情况下,我有几个日期字段['//1997', '/02/1967'],我需要转换['01/01/1997','01/02/1967'].这parse_dates有助于我将这些类型的日期字段转换为预期的格式,而不会让我编写额外的代码行.

这有什么解决方案吗？

Bug Link @GitHub:https://github.com/pydata/pandas/issues/13063

Answer 1

在熊猫版本中，0.18.0您可以添加参数dayfirst=True，然后它可以工作：

import pandas as pd
import io

temp=u"""start_date
04/10/96
15/07/97
10/12/97
06/03/99
//1994
/02/1967
"""
#after testing replace io.StringIO(temp) to filename
df = pd.read_csv(io.StringIO(temp),  parse_dates = ['start_date'], dayfirst=True)
  start_date
0 1996-10-04
1 1997-07-15
2 1997-12-10
3 1999-03-06
4 1994-01-01
5 1967-02-01

另一个解决方案：

您可以解析to_datetime使用不同的参数format，并errors='coerce'再combine_first：

date1 = pd.to_datetime(df['start_date'], format='%d/%m/%y', errors='coerce')
print date1
0   1996-10-04
1   1997-07-15
2   1997-12-10
3   1999-03-06
4          NaT
5          NaT
Name: start_date, dtype: datetime64[ns]

date2 = pd.to_datetime(df['start_date'], format='/%m/%Y', errors='coerce')
print date2
0          NaT
1          NaT
2          NaT
3          NaT
4          NaT
5   1967-02-01
Name: start_date, dtype: datetime64[ns]

date3 = pd.to_datetime(df['start_date'], format='//%Y', errors='coerce')
print date3
0          NaT
1          NaT
2          NaT
3          NaT
4   1994-01-01
5          NaT
Name: start_date, dtype: datetime64[ns]

print date1.combine_first(date2).combine_first(date3)
0   1996-10-04
1   1997-07-15
2   1997-12-10
3   1999-03-06
4   1994-01-01
5   1967-02-01
Name: start_date, dtype: datetime64[ns]