Repository navigation
np.vectorize handling of timedelta64 arrays #5180
Description
It seems that np.vectorize casts / unboxes timedelta64 values and does so kind of inconsistently depending on the unit. I am not sure that this was intentional, but makes it difficult to work with such arrays; in particular, it is not possible to vectorize a function which takes timedelta64 value;
>>> np.__version__
'1.9.0'
>>> f = lambda x: print(type(x), x)
>>> xs = np.arange(1, 3).astype('timedelta64[h]')
>>> xs
array([1, 2], dtype='timedelta64[h]')
with [h] and [s] unit:
>>> _ = np.vectorize(f)(xs)
<class 'numpy.timedelta64'> 1 hours
<class 'datetime.timedelta'> 1:00:00
<class 'datetime.timedelta'> 2:00:00
>>> _ = np.vectorize(f)(xs.astype('timedelta64[s]'))
<class 'numpy.timedelta64'> 3600 seconds
<class 'datetime.timedelta'> 1:00:00
<class 'datetime.timedelta'> 2:00:00
>>> _ = np.vectorize(f, otypes='O')(xs.astype('timedelta64[s]'))
<class 'datetime.timedelta'> 1:00:00
<class 'datetime.timedelta'> 2:00:00
with [ns] unit:
>>> _ = np.vectorize(f)(xs.astype('timedelta64[ns]'))
<class 'numpy.timedelta64'> 3600000000000 nanoseconds
<class 'int'> 3600000000000
<class 'int'> 7200000000000
>>> _ = np.vectorize(f, otypes='O')(xs.astype('timedelta64[ns]'))
<class 'int'> 3600000000000
<class 'int'> 7200000000000
Activity
seberg commentedon Jan 3, 2019
Closing in favor of similar/duplicate (newer) issue gh-12550