Handling of pandas datetimes in numpy

[hjstein]

I'm finding a regression from numpy v1.14.3 to numpy v1.15.4. The issue is that in v1.14.3, busday_count() would operate on datetime64 types, and in v1.15.4, this is no longer the case. This comes up when trying to use busday_count() on dates from a pandas dataframe that stores them as datetimes.

Reproducing code example:

import numpy as np
import pandas as pd
df = pd.DataFrame([["2011-01-01", 14], ["2012-01-01", 25]], columns=["Date", "Amount"])
d1 = np.busday_count(df["Date"].iloc[0], df["Date"].iloc[1])
df["Date"] = pd.to_datetime(df["Date"])
d2 = np.busday_count(df["Date"].iloc[0].date(), df["Date"].iloc[1].date())
d3 = np.busday_count(df["Date"].iloc[0], df["Date"].iloc[1])

Error message:

In v1.14.0 and v1.14.3, the assignments to d1, d2 and d3 succeed and yield the same result. In v1.15.4, the assignment to d3 gives an error message:

in
----> 1 d3 = np.busday_count(df["Date"].iloc[0], df["Date"].iloc[1])

TypeError: Iterator operand 0 dtype could not be cast from dtype('<M8[us]') to dtype('<M8[D]') according to the rule 'safe'

Numpy/Python version information:

Versions on system where error occurs:

import sys, numpy, pandas; print(numpy.version, pandas.version, sys.version)
1.15.4 0.23.4 3.6.7 |Anaconda custom (64-bit)| (default, Dec 10 2018, 20:35:02) [MSC v.1915 64 bit (AMD64)]

Versions on system where error does not occur:

import sys, numpy, pandas; print(numpy.version, pandas.version, sys.version)
1.14.3 0.23.0 3.6.5 |Anaconda, Inc.| (default, Apr 29 2018, 16:14:56)
[GCC 7.2.0]

[jbrockmendel]

This issue may be clarified by taking the pandas out of of it:

import numpy as np
from datetime import datetime

left = datetime(2011, 1, 1)
right = datetime(2012, 1, 1)

>>> np.busday_count(left, right)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<__array_function__ internals>", line 5, in busday_count
TypeError: Iterator operand 0 dtype could not be cast from dtype('<M8[us]') to dtype('<M8[D]') according to the rule 'safe'

(numpy 1.19.5)