CODEWITHSUNDEEP
pythonpandascsvdatetime
Magellan88
Magellan88

Reputation: 2573

Pandas parse dates from three fields with milliseconds

I've got a csv file with multiple columns one of which is the date, the next one is the time and then there is one with the miliseconds like so:

    DATE        TIME        MSEC 
0   13/01/2017  08:49:37    805102    
1   13/01/2017  08:49:09    675839    
2   13/01/2017  08:39:03    50614    
3   13/01/2017  08:38:30    383081

Now I'm really impressed that I can use pandas to combine the date and the time when reading the csv file like

pd.read_csv(r"~/file.csv", parse_dates=[["DATE","TIME"]])

but I can't seem to squeeze in the milliseconds as well. As the file is quite large I'd be greatful if this could be done in read_csv as well rather then in a consecutive step.

Upvotes: 1

Views: 1484

Answers (1)

MaxU - stand with Ukraine
MaxU - stand with Ukraine

Reputation: 210882

UPDATE: parsing date from different columns on the fly when reading CSV:

In [181]: pd.read_csv(fn, 
                      date_parser=lambda d,t,ms: d + ' ' + t + '.' + ms, 
                      parse_dates={'Timestamp':['DATE','TIME','MSEC']})
Out[181]:
                   Timestamp
0 2017-01-13 08:49:37.805102
1 2017-01-13 08:49:09.675839
2 2017-01-13 08:39:03.506140
3 2017-01-13 08:38:30.383081

OLD answer:

First read your CSV as it is:

df = pd.read_csv(r"~/file.csv")

In [170]: df
Out[170]:
         DATE      TIME    MSEC
0  13/01/2017  08:49:37  805102
1  13/01/2017  08:49:09  675839
2  13/01/2017  08:39:03   50614
3  13/01/2017  08:38:30  383081

In [171]: df.dtypes
Out[171]:
DATE    object
TIME    object
MSEC     int64
dtype: object

now we can convert it:

In [172]: df['TimeStamp'] = pd.to_datetime(df.DATE + ' ' + df.TIME + '.' + df.MSEC.astype(str), format='%d/%m/%Y %H:%M:%S.%f')

In [173]: df
Out[173]:
         DATE      TIME    MSEC                  TimeStamp
0  13/01/2017  08:49:37  805102 2017-01-13 08:49:37.805102
1  13/01/2017  08:49:09  675839 2017-01-13 08:49:09.675839
2  13/01/2017  08:39:03   50614 2017-01-13 08:39:03.506140
3  13/01/2017  08:38:30  383081 2017-01-13 08:38:30.383081


In [174]: df.dtypes
Out[174]:
DATE                 object
TIME                 object
MSEC                  int64
TimeStamp    datetime64[ns]
dtype: object

TIMING:

In [186]: df = pd.concat([df] * 10**3, ignore_index=True)

In [187]: df.shape
Out[187]: (4000, 3)

In [188]: df.to_csv(fn, index=False)

In [189]: pd.options.display.max_rows = 6

In [190]: df
Out[190]:
            DATE      TIME    MSEC
0     13/01/2017  08:49:37  805102
1     13/01/2017  08:49:09  675839
2     13/01/2017  08:39:03   50614
...          ...       ...     ...
3997  13/01/2017  08:49:09  675839
3998  13/01/2017  08:39:03   50614
3999  13/01/2017  08:38:30  383081

[4000 rows x 3 columns]

In [191]: %%timeit
     ...: pd.read_csv(fn,
     ...:             date_parser=lambda d,t,ms: d + ' ' + t + '.' + ms,
     ...:             parse_dates={'Timestamp':['DATE','TIME','MSEC']})
     ...:
1 loop, best of 3: 3.31 s per loop

In [192]: %%timeit
     ...: df = pd.read_csv(fn)
     ...: df['TimeStamp'] = pd.to_datetime(df.pop('DATE') + ' ' +
     ...:                                  df.pop('TIME') + '.' +
     ...:                                  df.pop('MSEC').astype(str),
     ...:                                  format='%d/%m/%Y %H:%M:%S.%f')
     ...:
10 loops, best of 3: 122 ms per loop

Conclusion: reading CSV as it is and parsing date from the DataFrame was 27 times faster for 4.000 rows data set.

Upvotes: 2

Related Questions