问题
I have a column in my dataset which represents a date in ms and sometimes its values is nan (actually my columns is of type str and sometimes its valus is 'nan'). I want to compute the epoch in days of this column. The problem is that when doing the difference of two dates:
(pd.to_datetime('now') - pd.to_datetime(np.nan)).days
if one is nan it is converted to NaT and the difference is of type NaTType which hasn't the attribute days.
In my case I would like to have nan as a result.
Other approach I have tried: np.datetime64 cannot be used, since it cannot take as argument nan. My data cannot be converted to int since int doesn't have nan.
回答1:
It will just work even if you filter first:
In [201]:
df = pd.DataFrame({'date':[dt.datetime.now(), pd.NaT, dt.datetime(2015,1,1)]})
df
Out[201]:
date
0 2015-08-28 12:12:12.851729
1 NaT
2 2015-01-01 00:00:00.000000
In [203]:
df.loc[df['date'].notnull(), 'days'] = (pd.to_datetime('now') - df['date']).dt.days
df
Out[203]:
date days
0 2015-08-28 12:12:12.851729 -1
1 NaT NaN
2 2015-01-01 00:00:00.000000 239
回答2:
For me upgrading to pandas 0.20.3 from pandas 0.19.2 helped resolve this error.
pip install --upgrade pandas
来源：https://stackoverflow.com/questions/32269676/nattype-object-has-no-attribute-days