Softplus without Variable

autograd
1

Traceback (most recent call last):
File “”, line 1, in
RuntimeError: softplus(): argument ‘input’ (position 1) must be Variable, not torch.FloatTensor

Why softplus doesn’t work without Variable(…) wrapper?
Doesn’t torch.nn.functional.softplus(Variable(tensor)).data seem ugly?

Pytorch versions: 0.3.0.post4, 0.2.0.post4’

2

nn.funtional are supposed to work with torch Variable not the plain Tensor. Ref here.

3

So the only way out at the moment is to use F.softplus(Variable(x)).data until Variable and Tensor will be merged, because torch.log(torch.exp(x) + 1) is not as numerically stable as F.softplus?

4

I suppose so. Or you can find the source code of softplus and see how it is implemented.

5

Variable and Tensor will be merged into one class in near future.