"nan" behaves differently in CPU and GPU #1810
Description
Should I report this to torch repository?
>>> torch.topk(torch.FloatTensor([-1,float("nan"),2, 1]), 1)
(
2
[torch.FloatTensor of size 1],
2
[torch.LongTensor of size 1])
>>> torch.topk(torch.cuda.FloatTensor([-1,float("nan"),2]), 1)
(
nan
[torch.cuda.FloatTensor of size 1 (GPU 0)],
9.2054e+18
[torch.cuda.LongTensor of size 1 (GPU 0)])In CPU, topk seems to ignore "nan" and its indices from result.
In GPU, topk with "nan" is kind of undefined behavior?
and other examples
>>> torch.max(torch.cuda.FloatTensor([float("nan"), 1]))
1.0
>>> torch.max(torch.FloatTensor([float("nan"), 1]))
nan
>>> torch.min(torch.cuda.FloatTensor([float("nan"), 1]))
1.0
>>> torch.min(torch.FloatTensor([float("nan"), 1]))
nan>>> torch.sort(torch.FloatTensor([float("nan"), 1]))
(
nan
1
[torch.FloatTensor of size 2],
0
1
[torch.LongTensor of size 2])
>> torch.sort(torch.cuda.FloatTensor([float("nan"), 1]))
(
1
nan
[torch.cuda.FloatTensor of size 2 (GPU 0)],
1
0
[torch.cuda.LongTensor of size 2 (GPU 0)])