How to assign NaN to tensor element?

Question

I want to assign NaN to a tensor element.

import torch
x = torch.tensor([1, 2, 3])
x[x == 2] = None

I have the error:

TypeError: can't assign a NoneType to a torch.LongTensor

I need it to make sure that some later sophisticated calculations are not made for certain values of x.

Osmunda

Please ask a consistent question: you say that you want to assign Nan, but your code specifically tries to assign None. – Bulge

Answer 1

The following code will set the desired value to nan:

import torch
x = torch.tensor([1, 2, 3]).float()
x[x == 2] = float('nan')

Spyglass

Answer 2

I had never seen float('nan') before! Thanks for the other answer. To add up to it, I wanted to emphasize that there are different versions of NaN, defined in different packages. For example, pandas and numpy have multiple versions. I wanted to verify that they all represent a similar object.

The takeaways are:

1. Equality for torch tensors works, but not when they contain nans,
2. All nans can be used when checking for closeness of tensors.
3. All nans can be used to create torch tensors. They are all compatible.

---

Here is the experiment:

import torch
import numpy

tensor_float_nan = torch.tensor([(float('nan'))])
tensor_np_nan = torch.tensor([(numpy.nan)])
tensor_np_nan_var = torch.tensor([(numpy.NaN)])
tensor_np_nan_again = torch.tensor([numpy.nan])

>>> print(f"Check if equality works for tensors:  "
          f"{(torch.tensor([1.]) == torch.tensor([1.])).item()}.")
True

>>> print(f"Check if float nan is the same as itself:  "
          f"{(tensor_float_nan == tensor_float_nan).item()}.")
False

>>> print(f"Check if float nan is the same as np nan: "
          f"{(tensor_float_nan == tensor_np_nan).item()}.")
False

>>> print(f"Check if float nan is the same as np nan var: "
          f"{(tensor_float_nan == tensor_np_nan_var).item()}.")
False

>>> print(f"Check if float nan is the same as np nan again: "
          f"{(tensor_float_nan == tensor_np_nan_again).item()}.")
False

>>> print(f"Check if np nan is the same as np nan var: "
          f"{(tensor_np_nan == tensor_np_nan_var).item()}.")
False

>>> print(f"Check if np nan is the same as np nan again: "
          f"{(tensor_np_nan == tensor_np_nan_again).item()}.")
False

>>> print(f"Check if np nan var is the same as np nan again: "
          f"{(tensor_np_nan_var == tensor_np_nan_again).item()}.")
False

>>> print(f"Check all close between float nan and np nan: "
      f"{torch.allclose(tensor_float_nan, tensor_np_nan, equal_nan=True)}.")
False

>>> print(f"Check all close between float nan and np nan var: "
      f"{torch.allclose(tensor_float_nan, tensor_np_nan_var, equal_nan=True)}.")
True

>>> print(f"Check all close between float nan and np nan again: "
      f"{torch.allclose(tensor_float_nan, tensor_np_nan_again, equal_nan=True)}.")
True

>>> print(f"Check all close between np nan and np nan var: "
      f"{torch.allclose(tensor_np_nan, tensor_np_nan_var, equal_nan=True)}.")
True

>>> print(f"Check all close between np nan and np nan again: "
      f"{torch.allclose(tensor_np_nan, tensor_np_nan_again, equal_nan=True)}.")
True

>>> print(f"Check all close between np nan var and np nan again: "
      f"{torch.allclose(tensor_np_nan_var, tensor_np_nan_again, equal_nan=True)}.")
True

Accipiter