pandas convert string columns to datetime, allowing missing but not invalid

问题

I have a pandas data frame with multiple columns of strings representing dates, with empty strings representing missing dates. For example

import numpy as np
import pandas as pd

# expected date format is 'm/%d/%Y'

custId = np.array(list(range(1,6)))
eventDate = np.array(["06/10/1992","08/24/2012","04/24/2015","","10/14/2009"])
registerDate = np.array(["06/08/2002","08/20/2012","04/20/2015","","10/10/2009"])

# both date columns of dfGood should convert to datetime without error
dfGood = pd.DataFrame({'custId':custId, 'eventDate':eventDate, 'registerDate':registerDate}) 

I am trying to:

• Efficiently convert columns where all strings are valid dates or empty into columns of type datetime64 (with NaT for the empty)
• Raise ValueError when any non-empty string does not conform to the expected format,

Example of where ValueError should be raised:

# 2nd string invalid
registerDate = np.array(["06/08/2002","20/08/2012","04/20/2015","","10/10/2009"]) 
# eventDate column should convert, registerDate column should raise ValueError
dfBad = pd.DataFrame({'custId':custId, 'eventDate':eventDate, 'registerDate':registerDate})

This function does what I want at the element level:

from datetime import datetime

def parseStrToDt(s, format = '%m/%d/%Y'):
    """Parse a string to datetime with the supplied format."""
    return pd.NaT if s=='' else datetime.strptime(s, format)

print(parseStrToDt("")) # correctly returns NaT
print(parseStrToDt("12/31/2011")) # correctly returns 2011-12-31 00:00:00
print(parseStrToDt("12/31/11")) # correctly raises ValueError

However, I have read that string operations shouldn't be np.vectorize-d. I thought this could be done efficiently using pandas.DataFrame.apply, as in:

dfGood[['eventDate','registerDate']].applymap(lambda s: parseStrToDt(s)) # raises TypeError

dfGood.loc[:,'eventDate'].apply(lambda s: parseStrToDt(s)) # raises same TypeError

I'm guessing that the TypeError has something to do with my function returning a different dtype, but I do want to take advantage of dynamic typing and replace the string with a datetime (unless ValueError is raise)... so how can I do this?

回答1:

pandas doesn't have an option that exactly replicates what you want, here's one way to do it, which should be relatively efficient.

In [4]: dfBad
Out[4]: 
   custId   eventDate registerDate
0       1  06/10/1992   06/08/2002
1       2  08/24/2012   20/08/2012
2       3  04/24/2015   04/20/2015
3       4                         
4       5  10/14/2009   10/10/2009

In [7]: cols
Out[7]: ['eventDate', 'registerDate']

In [9]: dts = dfBad[cols].apply(lambda x: pd.to_datetime(x, errors='coerce', format='%m/%d/%Y'))

In [10]: dts
Out[10]: 
   eventDate registerDate
0 1992-06-10   2002-06-08
1 2012-08-24          NaT
2 2015-04-24   2015-04-20
3        NaT          NaT
4 2009-10-14   2009-10-10

In [11]: mask = pd.isnull(dts) & (dfBad[cols] != '')

In [12]: mask
Out[12]: 
  eventDate registerDate
0     False        False
1     False         True
2     False        False
3     False        False
4     False        False


In [13]: mask.any()
Out[13]: 
eventDate       False
registerDate     True
dtype: bool

In [14]: is_bad = mask.any()

In [23]: if is_bad.any():
    ...:     raise ValueError("bad dates in col(s) {0}".format(is_bad[is_bad].index.tolist()))
    ...: else:
    ...:     df[cols] = dts
    ...:     
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-23-579c06ce3c77> in <module>()
      1 if is_bad.any():
----> 2     raise ValueError("bad dates in col(s) {0}".format(is_bad[is_bad].index.tolist()))
      3 else:
      4     df[cols] = dts
      5 

ValueError: bad dates in col(s) ['registerDate']

回答2:

Just to take the accepted answer a little further, I replaced the columns of all valid or missing strings with their parsed datetimes, and then raised an error for the remaining unparsed columns:

dtCols = ['eventDate', 'registerDate']
dts = dfBad[dtCols].apply(lambda x: pd.to_datetime(x, errors='coerce', format='%m/%d/%Y'))

mask = pd.isnull(dts) & (dfBad[dtCols] != '')
colHasError = mask.any()

invalidCols = colHasError[colHasError].index.tolist() 
validCols = list(set(dtCols) - set(invalidCols))

dfBad[validCols] = dts[validCols] # replace the completely valid/empty string cols with dates
if colHasError.any():
    raise ValueError("bad dates in col(s) {0}".format(invalidCols))
# raises:  ValueError: bad dates in col(s) ['registerDate']

print(dfBad) # eventDate got converted, registerDate didn't

The accepted answer contains the main insight, though, which is to go ahead and coerce errors to NaT and then distinguish the non-empty but invalid strings from the empty ones with the mask.

来源：https://stackoverflow.com/questions/38594246/pandas-convert-string-columns-to-datetime-allowing-missing-but-not-invalid