Max of a tensor and a scalar

1

Is there a PyTorch analogue of Lua’s torch.cmax(tensor, value)?

I checked the docs torch.max seems to want another tensor as an argument and doesn’t cut it.

There are workarounds, but an easy direct way of doing it would be more concise and probably efficient.

Thanks!

2 Likes
2

One way to accomplish this would be with the torch.max() function you mentioned:

X = th.FloatTensor([
    [-1, 2],
    [2, -1]
])
scalar = th.FloatTensor([1])

th.max(X, scalar.expand_as(X))

gives you

 1  2
 2  1
[torch.FloatTensor of size 2x2]

Please keep in mind that I’m playing with PyTorch for few days only, so that may not be the optimal solution.

EDIT I remember one of the devs mentioned they’re considering adding a autograd.Scalar, I guess it would make things nicer in this case.

1 Like
3

Yep, that’s workaround I’m currently using, but it’s a little too verbose IMO compared to cmax.

5 
torch.clamp(tensor, max=value) # cmin
torch.clamp(tensor, min=value) # cmax

Edited to reflect what vadim said below

10 Likes
6

Seems it’s the other way around:

torch.clamp(tensor, min=value) is cmax and
torch.clamp(tensor, max=value) is cmin.

It works but is a little confusing at first.

7

cmax/cmin was removed to simplify the API, after some comments we received.
https://github.com/pytorch/pytorch/pull/455

8

Maybe someone is looking for this. Is there a way to do this without that much code?

def min_(value1, value2):
    if is_number(value1) and isinstance(value2, torch.Tensor):
        return torch.clamp(value2, max=value1)
    elif is_number(value2) and isinstance(value1, torch.Tensor):
        return torch.clamp(value1, max=value2)
    elif isinstance(value1, torch.Tensor) and isinstance(value2, torch.Tensor):
        return torch.min(value1, value2)


def max_(value1, value2):
    if is_number(value1) and isinstance(value2, torch.Tensor):
        return torch.clamp(value2, min=value1)
    elif is_number(value2) and isinstance(value1, torch.Tensor):
        return torch.clamp(value1, min=value2)
    elif isinstance(value1, torch.Tensor) and isinstance(value2, torch.Tensor):
        return torch.max(value1, value2)

min_(2, torch.tensor(3))
min_(torch.tensor(3), 2)
min_(torch.tensor(3), torch.tensor(2))