Timedelta does not allow assignment from float

When trying gh-4434, I ran into a force-cast difference between timedelta force casting and timedelta item assignment. There is also a discrepancy for float64 vs. python floats:

In [12]: np.array(3.).astype('timedelta64[D]')
Out[12]: array(datetime.timedelta(3), dtype='timedelta64[D]')

In [13]: np.array(3., dtype='timedelta64[D]')
ValueError: Could not convert object to NumPy timedelta

and:

In [17]: np.empty(1, dtype='timedelta64[D]')[[0]] = np.float64(3.)

In [18]: np.empty(1, dtype='timedelta64[D]')[0] = np.float64(3.)
ValueError: Could not convert object to NumPy timedelta

I could easily hack around this by adding an ellipsis, but this all seems a bit weird…

[heavelock]

To all those previous observations I would also like to add that casting existing array of floats to np.timedelta64 results in flooring the values in the array, without throwing any error. Quite unfortunate.

np.__version__ # '1.19.4'
np.array([1.111]).astype('timedelta64[s]')  # array([1], dtype='timedelta64[s]')
np.array([1.999]).astype('timedelta64[s]')  # array([1], dtype='timedelta64[s]')

[eric-wieser]

Then I don’t think that’s any worse than array(1 .5).astype(int), which also silently rounds