PyTorch .frexp()

In PyTorch, the .frexp() function takes a tensor as input and returns a tuple storing two tensors: mantissa and exponent. The range of mantissa is the open interval (-1, 1). The original input can be reconstructed as :

$$\text{input} = \text{mantissa}\times 2^{\text{exponent}}$$

Syntax

torch.frexp(input, *, out=None) -> (Tensor mantissa, Tensor exponent)

Parameters:

• input: The input tensor.
• out (Optional): A tuple of two tensors (mantissa, exponent) that will store the result. If provided, the output is written in place.

Return value:

It returns a tuple containing two tensors - mantissa and exponent.

Example

This example uses the .frexp() function on a tensor. The result contains two tensors mantissa and exponent (which can give numerical stability when floating point numbers are too small or too large):

import torch

x = torch.tensor([1. , 4. , 5. , 9. ])
mantissa, exponent = torch.frexp(x)

print(mantissa)
print(exponent)

Copy to clipboard
Copy to clipboard

The output of this code is:

tensor([0.5000, 0.5000, 0.6250, 0.5625])
tensor([1, 3, 3, 4], dtype=torch.int32)

Copy to clipboard
Copy to clipboard