torch.sign() Method in PyTorch

The torch.sign() method returns a new tensor with the sign of each element in the input tensor. For a given element ( x ), the output is defined as:

1. 1 if  x > 0
2. 0 if  x = 0
3. -1 if  x < 0

If input is complex, the sign is defined as input / |input|.

It is an element-wise operation, so the output tensor preserves the shape of the input tensor.

Syntax

torch.sign(input, out=None)

Parameters

Argument	Description
input (Tensor)	It represents an input tensor of any numerical data type, such as torch.float32 and torch.int64.
out (Tensor, optional)	It represents an optional output tensor to store the result. If you don’t provide, it will create a new tensor.

Basic usage with floating-point tensors

Let’s define the 1D tensor and compute the sign of elements in a floating-point tensor.

import torch

tensor = torch.tensor([-2.1, 0.0, 1.9, -1.1])

signed_tensor = torch.sign(tensor)

print(signed_tensor)

# Output: tensor([-1.,  0.,  1., -1.])

From the above code, you can see that it replaces – (negative) value with -1, a + (positive) value with 1, and 0 with 0.

Integer tensors

In the above section, we saw a floating-point tensor; now, we will define a 2D integer tensor and apply the sign() method on it.

import torch

tensor_2d = torch.tensor([[-11, 0],
                          [19, -21]])

signed_2d_tensor = torch.sign(tensor_2d)

print(signed_2d_tensor)

# Output:
# tensor([[-1,  0],
#         [ 1, -1]])

GPU Support

You can also run the torch.sign() method on a CUDA-enabled GPU like this:

import torch

if torch.cuda.is_available():
   x = torch.tensor([-1.5, 0.0, 2.3]).cuda()
   result = torch.sign(x)
   print(result)  # Output: tensor([-1., 0., 1.], device='cuda:0')

Combining with torch.where() for conditional logic

We can use torch.sign() method with torch.where() to implement conditional transformations. We will conditionally replace negative values with -10 and positive or zero values with 10.

import torch

tensor = torch.tensor([-1.0, 0.0, 5.0])

# Replace positive values with 10, others with -10
output_tensor = torch.where(torch.sign(tensor) > 0, torch.tensor(10.0), torch.tensor(-10.0))

print(output_tensor)  

# Output: tensor([-10., -10.,  10.])

That’s all!