PyTorch .copysign()

PyTorch’s .copysign() function creates a new tensor by combining the magnitude of the first input with the sign of the second input, element-wise.

Note: If the first input contains booleans, they are cast to floats (True becomes 1.0, False becomes 0.0). -0.0 or -0 in the second input is treated as a negative sign, while 0.0 or 0 is treated as positive, and these signs are preserved when copied.

Syntax

torch.copysign(input, other, *, out=None)

Parameters:

• input (Tensor): The input tensor providing magnitudes.
• other (Tensor): The tensor whose sign will be copied to input. Must be broadcastable to the shape of input.
• out (Tensor, optional): The tensor to store the output. Must have the same shape as the broadcasted result.

Note: The tensors must be broadcastable in order to perform the operation. Otherwise, this will result in a RuntimeError.

Return value:

A tensor where each element has the absolute value of the corresponding element in input and the sign of the corresponding element in other.

Example 1

This example shows how to copy signs from the signs tensor to the magnitudes tensor using .copysign():

import torch

# Initializing tensors
magnitudes = torch.tensor([-5.0, True, -8.0])
signs = torch.tensor([-0.0, 0.0, 3.0])

# Using torch.copysign to copy the sign from 'signs' to 'magnitudes'
result = torch.copysign(magnitudes, signs)
print(result)

Copy to clipboard
Copy to clipboard

The output of this code is:

tensor([-5., 1., 8.])

Copy to clipboard
Copy to clipboard

Example 2

This example shows how to copy the sign from a one-dimensional signs tensor to a two-dimensional magnitudes tensor using broadcasting. Each row of magnitudes is aligned with the corresponding value in signs, and the sign is applied column-wise:

import torch

torch.manual_seed(0)

# Create a 3x3 tensor of magnitudes
magnitudes = torch.rand(3, 3)
print("Magnitudes:\n", magnitudes)

# 1D tensor whose signs will be broadcast across rows
signs = torch.tensor([0.6323, 0.3489, -0.4017])

# Apply copysign with broadcasting
result = torch.copysign(magnitudes, signs)
print("\nResult:\n", result)

Copy to clipboard
Copy to clipboard

The output of this code is:

Magnitudes:
 tensor([[0.4963, 0.7682, 0.0885],
        [0.1320, 0.3074, 0.6341],
        [0.4901, 0.8964, 0.4556]])

Result:
 tensor([[ 0.4963,  0.7682, -0.0885],
        [ 0.1320,  0.3074, -0.6341],
        [ 0.4901,  0.8964, -0.4556]])

Copy to clipboard
Copy to clipboard