torch.max and torch.min inconsistent on cpu/gpu for tensors with zero elements. #17750
Description
On cpu:
g = torch.rand(0,4)
torch.max(g)
Behaviour:
RuntimeError: invalid argument 1: tensor must have one dimension at /pytorch/aten/src/TH/generic/THTensorEvenMoreMath.cpp:59
On Gpu:
g = torch.rand(0,4).cuda()
torch.max(g)
Behaviour:
tensor(-inf, device='cuda:0')
Expected Behaviour on GPU:
Some error or an empty tensor as i feel the max/min of a set should belong in the set so -inf is wrong.
PyTorch Version :'1.0.1.post2'