Pandas doesn't accept dtype=np.datetime64

[Wilfred]

In [1]: from pandas import Series

In [3]: import numpy as np  

In [4]: Series([], dtype=np.datetime64)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-4-139e5646d234> in <module>()
----> 1 Series([], dtype=np.datetime64)

/users/is/whughes/pyenvs/1c4f50ee0e346415/lib/python2.7/site-packages/pandas-0.13.1-py2.7-linux-x86_64.egg/pandas/core/series.py in __init__(self, data, index, dtype, name, copy, fastpath)
    218             else:
    219                 data = _sanitize_array(data, index, dtype, copy,
--> 220                                        raise_cast_failure=True)
    221 
    222                 data = SingleBlockManager(data, index, fastpath=True)

/users/is/whughes/pyenvs/1c4f50ee0e346415/lib/python2.7/site-packages/pandas-0.13.1-py2.7-linux-x86_64.egg/pandas/core/series.py in _sanitize_array(data, index, dtype, copy, raise_cast_failure)
   2564 
   2565     else:
-> 2566         subarr = _try_cast(data, False)
   2567 
   2568     # scalar like

/users/is/whughes/pyenvs/1c4f50ee0e346415/lib/python2.7/site-packages/pandas-0.13.1-py2.7-linux-x86_64.egg/pandas/core/series.py in _try_cast(arr, take_fast_path)
   2509 
   2510         try:
-> 2511             arr = _possibly_cast_to_datetime(arr, dtype)
   2512             subarr = pa.array(arr, dtype=dtype, copy=copy)
   2513         except (ValueError, TypeError):

/users/is/whughes/pyenvs/1c4f50ee0e346415/lib/python2.7/site-packages/pandas-0.13.1-py2.7-linux-x86_64.egg/pandas/core/common.py in _possibly_cast_to_datetime(value, dtype, coerce)
   1659                 else:
   1660                     raise TypeError(
-> 1661                         "cannot convert datetimelike to dtype [%s]" % dtype)
   1662             elif is_timedelta64 and dtype != _TD_DTYPE:
   1663                 if dtype.name == 'timedelta64[ns]':

TypeError: cannot convert datetimelike to dtype [datetime64]

The following workaround works:

Series([], dtype=datetime)

but I would expect passing in the numpy type to work too.

[jreback]

all datetimes are internally (and converted from) kept as datetime64[ns] this doesn't make sense.

and conversions are already way complicated.

You are certainly welcome to try to fix this. but is their an actual usecase?

[cpcloud]

these kinds of dtypes are also parameterized whereas things like int64, float64 etc are not, so it doesn't make sense to pass it in unparameterized.

also empty series of datetimes are very edge-casey what's the motivation?

i would even argue that passing in the dtype to series is unnecessary 95% of the time (pandas is pretty good about doing the right thing), if you need to coerce to datetimes use pd.to_datetime

[Wilfred]

I wanted to explicitly create a Series that contained datetimes, and the docstring of Series says:

dtype : numpy.dtype or None
     If None, dtype will be inferred

so I assumed that I could pass in a numpy type. Perhaps it's just a matter of clarifying the docstring?

Whilst I used an empty series as a minimal reproducible example, this does occur for non-empty series too:

Series([datetime.now()], dtype=np.datetime64) # same error
Series([np.datetime64(datetime.now())], dtype=np.datetime64) # same error

This behaviour does actually work with other numpy dtypes, e.g. Series([], dtype=np.int64), which is what set my expectations.

[cpcloud]

I wanted to explicitly create a Series that contained datetime

you can create and then do pd.to_datetime

do you have non-trivial use case?

would help us evaluate whether this is necessary

[jreback]

@Wilfred this is the point of to_datetime. As I said, you could allow this (their are explicity) check for NOT allowing this. But not sure its relevant/usecase. The doc-string is generic. That's why their is an error.

[ischwabacher]

If all you want is a workaround, note that you can do this:

In [1]: import pandas as pd

In [2]: import numpy as np

In [3]: pd.Series(dtype='m8[ns]')
Out[3]: Series([], dtype: timedelta64[ns])

In [4]: pd.Series(dtype=np.timedelta64(0, 'ns').dtype)
Out[4]: Series([], dtype: timedelta64[ns])

The problem is that numpy doesn't make it easy to get from timedelta64 to timedelta64[ns], but pandas demands nanoseconds.

[cpcloud]

numpy doesn't make it easy to get from timedelta64 to timedelta64[ns]

that's because timedelta64 without a unit doesn't make sense

[cpcloud]

even though for some reason numpy still lets you think that it does

In [18]: np.dtype('m8')
Out[18]: dtype('<m8')

[ischwabacher]

I don't mean a timedelta64 object, I mean the timedelta64 dtype. AFAIK the only paths to the timedelta64[ns] dtype go either through an instance (as in In[4]) or through the string 'm8[ns]'.

[jreback]

we are talking about M8 here, right? (not m8)

[ischwabacher]

Whoops, yes. The point still stands.

[Wilfred]

OK, perhaps the best solution would be to throw an explicit error about parameterized types, so the user knows how to fix. Something like:

Series([], dtype=np.datetime64)
TypeError: Can't use parameterised dtypes. Choose an concrete dtype (e.g. "datetime64[ns]" instead of "datetime") or let pandas infer the type.

[ischwabacher]

I was under the impression that from pandas's point of view, there is one true datetime64 dtype and that dtype is datetime64[ns], and any other datetime64 dtype should be cast appropriately. So I would actually expect that you would get this:

In [5]: pd.Series([1], dtype='M8')
Out[5]: 
0   1970-01-01 00:00:00.000000001
dtype: datetime64[ns]

In [6]: pd.Series([1], dtype='M8[s]')
Out[6]: 
0   1970-01-01 00:00:01
dtype: datetime64[ns]

In [7]: pd.Series([1], dtype='m8[s]')
Out[7]: 
0   00:00:01
dtype: timedelta64[ns]

In [8]: pd.Series([1], dtype='m8[Y]')
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-21-f94f55f94a5f> in <module>()
----> 1 pd.Series([1], dtype='m8[Y]')
...
TypeError: cannot convert dtype [timedelta64[Y]] to nanoseconds

Currently all of these raise.