PyTorch .nan_to_num()

Priyanshjain10's avatar
Priyanshjain10's avatar
Priyanshjain10
Published Oct 31, 2025

The .nan_to_num() function in PyTorch replaces NaN (Not a Number), positive infinity, and negative infinity values in a tensor with user-defined finite numbers. This is useful for handling invalid or extreme values in numerical computations.

  • Career path

    Data Scientist: Machine Learning Specialist

    Machine Learning Data Scientists solve problems at scale, make predictions, find patterns, and more! They use Python, SQL, and algorithms.
    • Includes 27 Courses
    • With Professional Certification
    • Beginner Friendly.
      95 hours

Syntax

torch.nan_to_num(input, nan=0.0, posinf=None, neginf=None, *, out=None) → Tensor

Parameters:

  • input (Tensor): The input tensor to process.
  • nan (Number, optional): The value to replace NaN with. Default is 0.0.
  • posinf (Number, optional): The value to replace positive infinity with. If None, positive infinities are replaced with the greatest finite value representable by the input’s dtype. Default is None.
  • neginf (Number, optional): The value to replace negative infinity with. If None, negative infinities are replaced with the lowest finite value representable by the input’s dtype. Default is None.
  • out (Tensor, optional): The output tensor to store the result. Default is None.

Return value:

A tensor with NaN, positive infinity, and negative infinity values replaced according to the specified parameters.

Example 1: Replacing NaN Values

This example demonstrates how to clean a tensor that contains NaN values by replacing them with zeros, which is a common preprocessing step before mathematical operations:

import torch


# Create a tensor with NaN values
x = torch.tensor([1.0, float('nan'), 3.0, float('nan'), 5.0])
print("Original tensor:", x)


# Replace NaN with 0.0
result = torch.nan_to_num(x, nan=0.0)
print("After nan_to_num:", result)

The output of this code is:

Original tensor: tensor([1., nan, 3., nan, 5.])
After nan_to_num: tensor([1., 0., 3., 0., 5.])

Example 2: Replacing Infinity Values

This example shows how to replace both positive and negative infinity values with finite numbers to prevent overflow or undefined results in computations:

import torch


# Create a tensor with infinity values
x = torch.tensor([1.0, float('inf'), -float('inf'), 4.0])
print("Original tensor:", x)


# Replace infinities with specific values
result = torch.nan_to_num(x, nan=0.0, posinf=100.0, neginf=-100.0)
print("After nan_to_num:", result)

The output of this code is:

Original tensor: tensor([1., inf, -inf, 4.])
After nan_to_num: tensor([1., 100., -100., 4.])

Example 3: Handling Mixed Invalid Values

Here, the tensor contains a mix of NaN, positive infinity, and negative infinity. All invalid values are replaced in one pass, making the tensor safe for numerical operations like normalization or model input:

import torch


# Create a tensor with various invalid values
x = torch.tensor([2.0, float('nan'), float('inf'), -float('inf'), 7.0])
print("Original tensor:", x)


# Replace all invalid values
result = torch.nan_to_num(x, nan=-1.0, posinf=999.0, neginf=-999.0)
print("After nan_to_num:", result)

The output of this code is:

Original tensor: tensor([2., nan, inf, -inf, 7.])
After nan_to_num: tensor([2., -1., 999., -999., 7.])

All contributors

Learn PyTorch on Codecademy

  • Career path

    Data Scientist: Machine Learning Specialist

    Machine Learning Data Scientists solve problems at scale, make predictions, find patterns, and more! They use Python, SQL, and algorithms.
    • Includes 27 Courses
    • With Professional Certification
    • Beginner Friendly.
      95 hours