Search code examples
pythonpandasnumpydatetimenumba

Python numpy: cannot convert datetime64[ns] to datetime64[D] (to use with Numba)

I want to pass a datetime array to a Numba function (which cannot be vectorised and would otherwise be very slow). I understand Numba supports numpy.datetime64. However, it seems it supports datetime64[D] (day precision) but not datetime64[ns] (nanosecond precision) (I learnt this the hard way: is it documented?).

I tried to convert from datetime64[ns] to datetime64[D], but can't seem to find a way!

I have summarised my problem with the minimal code below. If you run testdf(mydates), which is datetime64[D], it works fine. If you run testdf(dates_input), which is datetime64[ns], it doesn't. Note that this example simply passes the dates to the Numba function, which doesn't (yet) do anything with them. I try to convert dates_input to datetime64[D], but the conversion doesn't work. In my original code I read from a SQL table into a pandas dataframe, and need a column which changes the day of each date to the 15th.

import numba
import numpy as np
import pandas as pd
import datetime

mydates =np.array(['2010-01-01','2011-01-02']).astype('datetime64[D]')
df=pd.DataFrame()
df["rawdate"]=mydates
df["month_15"] = df["rawdate"].apply(lambda r: datetime.date( r.year, r.month,15 ) )

dates_input = df["month_15"].astype('datetime64[D]')
print dates_input.dtype # Why datetime64[ns] and not datetime64[D] ??


@numba.jit(nopython=True)
def testdf(dates):
    return 1

print testdf(mydates)

The error I get if I run testdf(dates_input) is:

numba.typeinfer.TypingError: Failed at nopython (nopython frontend)
Var 'dates' unified to object: dates := {pyobject}
Solution

  • Note (2023-05-30): This answer only works for pandas version <2. Pandas 2.0.0 was released on 2023-04-03. See relevant changelog entry.

    Series.astype converts all date-like objects to datetime64[ns].

    To convert to datetime64[D], use values to obtain a NumPy array before calling astype:

    dates_input = df["month_15"].values.astype('datetime64[D]')

    Note that NDFrames (such as Series and DataFrames) can only hold datetime-like objects as objects of dtype datetime64[ns]. The automatic conversion of all datetime-likes to a common dtype simplifies subsequent date computations. But it makes it impossible to store, say, datetime64[s] objects in a DataFrame column. Pandas core developer, Jeff Reback explains,

    "We don't allow direct conversions because its simply too complicated to keep anything other than datetime64[ns] internally (nor necessary at all)."

    Also note that even though df['month_15'].astype('datetime64[D]') has dtype datetime64[ns]:

    In [29]: df['month_15'].astype('datetime64[D]').dtype
Out[29]: dtype('<M8[ns]')

    when you iterate through the items in the Series, you get pandas Timestamps, not datetime64[ns]s.

    In [28]: df['month_15'].astype('datetime64[D]').tolist()
Out[28]: [Timestamp('2010-01-15 00:00:00'), Timestamp('2011-01-15 00:00:00')]

    Therefore, it is not clear that Numba actually has a problem with datetime64[ns], it might just have a problem with Timestamps. Sorry, I can't check this -- I don't have Numba installed.

    However, it might be useful for you to try

    testf(df['month_15'].astype('datetime64[D]').values)

    since df['month_15'].astype('datetime64[D]').values is truly a NumPy array of dtype datetime64[ns]:

    In [31]: df['month_15'].astype('datetime64[D]').values.dtype
Out[31]: dtype('<M8[ns]')

    If that works, then you don't have to convert everything to datetime64[D], you just have to pass NumPy arrays -- not Pandas Series -- to testf.