pandas series 'astype' conversion different from numpy.array 'astype' conversion

[jacksheng]

I am confused by the dtype conversion function "astype" for pandas series. According to documentation, it should be the same as numpy.ndarray dtype conversion, but seems there is stil some difference. For example, if I try to convert columns of datetime type to object dtype, I got completely different results using astype function on pandas series versus on numpy array. My question is what sort of additional hanlding exists inside pandas Series dtype conversion? Example code as follows, note that pandas Series dtype conversion converts numpy.datetime64 to datetime.datetime, while numpy.array dtype conversion converts numpy.datetime64 to long. By the way, my numpy is version 1.8.0, and my pandas is version 0.12.0, and running on python 2.7.

import datetime
import numpy
import pandas
dd = pandas.DataFrame({"val": [datetime.datetime(2014, 1, 1, 9)]})
dd["val"].astype("object")
Out[1]:
0 2014-01-01 09:00:00
Name: val, dtype: object
numpy.array(dd["val"]).astype("object")
Out[1]: array([1388566800000000000L], dtype=object)

[jreback]

This is as expected.

Pandas is quite restrictive when astyping datetime64 and timedelta64 types. These are kept internally as datetime64[ns] and timedelta64[ns] so pretty much only object dtype is allowed (timedeltat64 can be cast slightly different; its in the docs: http://pandas.pydata.org/pandas-docs/dev/timeseries.html#time-deltas-conversions)

The int64 you are getting back in essentially (for datetime64) is the nanoseconds since 1970, and is used for all sorts of things internally (numpy has the same reference).

What are you doing that you actually need to astype these? (their is a limited set of operations where you would actually want to astype).

Alll that said; this could be doc'd better. Care to do a PR?

[jreback]

closing as this is expected