Consider the following code:
from torch.autograd import Variable
t = torch.Tensor([2])
bool(torch.max(t) < 2)
Out[4]:
False
bool(torch.max(t) < 3)
Out[5]:
True
However, If you do the same with a Variable:
bool(torch.max(v) < 2)
Out[6]:
True
bool(torch.max(v) < 3)
Out[7]:
True
In the mean time I call v.data to overcome this.
It is because torch.max(v) returns another Variable containing a Tensor with a single element (so that you can use it and autograd will work as expected) while torch.max(t) returns a python number.
For conditions (that are not differentiable and thus you don’t want to keep a Variable) you can either do torch.max(v.data) or torch.max(v).data[0].
Well that makes sense. Thanks.