python - 使用Python的Pandas包将hdf5文件中的列从int64转换为日期时间

Question

我对 Pandas 和一般编程都很陌生，因此我们将不胜感激。

我在将从 hdf5 文件加载的 Pandas 数据框中的一列数据转换为日期时间对象时遇到困难。数据太大，无法使用文本文件，因此我使用以下代码将其转换为 hdf5 文件:

# get text file from zip file and unzip
file = urllib.request.urlretrieve(file, dir)           
z = zipfile.ZipFile(dir)             
data = z.open(z.namelist()[0])

# column names from text file
colnames = ['Patent#','App#','Small','Filing Date','Issue Date', 'Event Date', 'Event Code'] 

# load the data in chunks and concat into single DataFrame        
mfees = pd.read_table(data, index_col=0, sep='\s+', header = None, names = colnames, chunksize=1000, iterator=True)
df = pd.concat([chunk for chunk in mfees], ignore_index=False)

# close files        
z.close()
data.close()

# convert to hdf5 file
data = data.to_hdf('mfees.h5','raw_data',format='table')

此后我的数据采用以下格式:

data['Filing Date']

输出:

Patent#
4287053    19801222
4287053    19801222
4289713    19810105
4289713    19810105
4289713    19810105
4289713    19810105
4289713    19810105
4289713    19810105
Name: Filing Date, Length: 11887679, dtype: int64

但是，当我使用 to_datetime 函数时，我得到以下结果:

data['Filing Date'] = pd.to_datetime(data['Filing Date'])
data['Filing Date']

输出:

Patent#
4287053   1970-01-01 00:00:00.019801222
4287053   1970-01-01 00:00:00.019801222
4289713   1970-01-01 00:00:00.019810105
4289713   1970-01-01 00:00:00.019810105
4289713   1970-01-01 00:00:00.019810105
4289713   1970-01-01 00:00:00.019810105
4289713   1970-01-01 00:00:00.019810105
4289713   1970-01-01 00:00:00.019810105
4289713   1970-01-01 00:00:00.019810105
4291808   1970-01-01 00:00:00.019801212
4291808   1970-01-01 00:00:00.019801212
4292069   1970-01-01 00:00:00.019810123
4292069   1970-01-01 00:00:00.019810123
4292069   1970-01-01 00:00:00.019810123
4292069   1970-01-01 00:00:00.019810123
Name: Filing Date, Length: 11887679, dtype: datetime64[ns]

我不确定为什么我会得到上述日期时间对象的输出。我可以做些什么来纠正这个问题并将日期正确转换为日期时间对象吗？谢谢!

Answer 1

最简单的方法是在您阅读时进行转换(请注意，我复制粘贴了您的数据，因此您只需添加 parse_dates=[1] 选项

In [31]: df = read_csv(StringIO(data),sep='\s+',header=None,parse_dates=[1],names=['num','date']).set_index('num')

In [32]: df
Out[32]: 
                       date
num                        
4287053 1980-12-22 00:00:00
4287053 1980-12-22 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00

In [33]: df.dtypes
Out[33]: 
date    datetime64[ns]
dtype: object

然后 hdf 将处理该列

In [46]: df.to_hdf('test.h5','df',mode='w',format='table')

In [47]: pd.read_hdf('test.h5','df')
Out[47]: 
                       date
num                        
4287053 1980-12-22 00:00:00
4287053 1980-12-22 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00
4289713 1981-01-05 00:00:00

In [48]: pd.read_hdf('test.h5','df').dtypes
Out[48]: 
date    datetime64[ns]
dtype: object

这是一个类似 int 的日期转换器，应该很快

In [18]: s = Series([19801222,19801222] + [19810105]*5)

In [19]: s
Out[19]: 
0    19801222
1    19801222
2    19810105
3    19810105
4    19810105
5    19810105
6    19810105
dtype: int64

In [20]: s = s.values.astype(object)

In [21]: Series(pd.lib.try_parse_year_month_day(s/10000,s/100 % 100, s % 100))
Out[21]: 
0   1980-12-22 00:00:00
1   1980-12-22 00:00:00
2   1981-01-05 00:00:00
3   1981-01-05 00:00:00
4   1981-01-05 00:00:00
5   1981-01-05 00:00:00
6   1981-01-05 00:00:00
dtype: datetime64[ns]