torch.bincount errors on empty input

[davidbau]

Issue description

torch.bincount gives an error when given an empty tensor, rather than just returning zero counts like numpy.bincount. Trips up edge cases when counting filtered data.

Code example

>>> import torch, numpy
>>> numpy.bincount([1]), numpy.bincount([]), numpy.bincount([], minlength=10)
(array([0, 1]), array([], dtype=int64), array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0]))

>>> torch.bincount(torch.LongTensor([]))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: bincount only supports 1-d non-negative integral inputs.

>>> torch.bincount(torch.LongTensor([]), minlength=10)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: bincount only supports 1-d non-negative integral inputs.

[zou3519]

@davidbau Numpy doesn't document this behavior docs here for the numpy.bincount([]) case.

The docs suggest the output should be

out : ndarray of ints
The result of binning the input array. The length of out is equal to np.amax(x)+1.

but if x is [], the size of the output isn't well defined. Do you think torch.bincount(torch.LongTensor([])) = torch.tensor([]) makes sense?

[davidbau]

Producing an empty vector from an empty count makes sense to me, although I find that in practice, I'm always using minlength, in which case the right output shape is unambiguous.