np.ndarray[object] - Timedelta raises

[jbrockmendel]

arr = np.array([pd.Timestamp.now(), pd.Timedelta('2D')])

>>> arr - pd.Timedelta('1D')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: unsupported operand type(s) for -: 'numpy.ndarray' and 'Timedelta'

It should attempt to operate element-wise.

[jorisvandenbossche]

I suppose this is because of

pandas/pandas/_libs/tslibs/timedeltas.pyx

Lines 539 to 544 in 27ebb3e

	elif hasattr(other, 'dtype'):
	# nd-array like
	if other.dtype.kind not in ['m', 'M']:
	# raise rathering than letting numpy return wrong answer
	return NotImplemented
	return op(self.to_timedelta64(), other)

Any idea what the "wrong" answer would be? (with timedelta.timedelta instead of Timedelta that seems to work just fine, so I assume with Timedelta it will be the same)

[jbrockmendel]

No idea what the wrong answer would be. This should be easy to fix; if no one else picks it up I'll take care of it once the current PR queue settles down.

[jorisvandenbossche]

Yes, PR with a fix is certainly welcome I think

[illegalnumbers]

Is this still an issue? I wasn't able to repro from master.

[jbrockmendel]

Is this still an issue? I wasn't able to repro from master.

What platform etc? I still get it on OSX in both py27 and py37.

[illegalnumbers]

OSX 10.11.6 with Python 3.6. I just pulled up a REPL and imported pandas from a compile I did yesterday from master and didn't get an exception from the example code posted. Specifically Python 3.6.6 (default, Jul 23 2018, 11:08:18) [GCC 4.2.1 Compatible Clang 6.0.0 (tags/RELEASE_600/final)] on darwin

[illegalnumbers]

I also didn't see the issue from the latest install from pip either. Both times I just got

>>> arr = np.array([pd.Timestamp.now(), pd.Timedelta('2D')])
>>> arr
array([Timestamp('2018-07-24 10:49:41.898067'),
       Timedelta('2 days 00:00:00')], dtype=object)

[jbrockmendel]

Did you try subtracting a Timedelta from arr?

[illegalnumbers]

Ah! 🤦‍♂️ yea I missed that part in the example. I repro'd the bug with that on master and latest pip. So with this then how should I go about the fix? It's not operating element wise on the array because the timedeltas.pyx isn't returning that it is a timedelta correctly? or...?

[jbrockmendel]

how should I go about the fix?

Take a look at the code block Joris quoted above. At the moment that lets 'm' and 'M' dtypes through but stops everything else. The fix will involve letting 'o' dtypes through (and making sure they are handled correctly)