The problem caused by the parameter dim of torch.norm

[triumph-wangyuyang]

🐛 Describe the bug

pytorch1.12.1

code 1 : The test code dim=False, but the thrown exception describes dim as a list.

    import torch
    input = torch.rand([5, 5, 5], dtype=torch.float64)
    out = torch.norm(input, p='fro', dim=False, keepdim=False, out=None, dtype=None)
    print(out)

result:    TypeError: frobenius_norm(): argument 'dim' (position 2) must be tuple of ints, not list


code 2 : The test code dim=False, the exception thrown is correct.

    import torch
    input = torch.rand([5, 5, 5], dtype=torch.float64)
    out = torch.norm(input, p='fro', dim="False", keepdim=False, out=None, dtype=None)
    print(out)

result: TypeError: frobenius_norm(): argument 'dim' (position 2) must be tuple of ints, not str

pytorch1.8.0

code 1 : In the pytorch1.8 version, dim is still set to False, and the program can run successfully, because the program automatically treats False as a value of 0..

      import torch
      input = torch.rand([5, 5, 5], dtype=torch.float64)
      out = torch.norm(input, p='fro', dim=False, keepdim=False, out=None, dtype=None)
      print(out)

result:    tensor([[1.3522, 1.1940, 0.7555, 1.3410, 1.3593],
              [1.0926, 1.0986, 1.3429, 1.0170, 1.1915],
              [1.1807, 1.4252, 1.5389, 1.2257, 1.5626],
              [1.5249, 0.9933, 1.2927, 1.7106, 1.6482],
              [1.1870, 1.0370, 1.6912, 1.0616, 1.0801]], dtype=torch.float64)

Question 1: Why is False treated as a list in the first test code?
Question 2: Why can’t False be treated as a value of 0 on a higher version, because when compiled according to the bottom layer of python, False will be set to 0, so why does the program fail?

Versions

pytorch: 1.12.1
pytorch: 1.8.0
Python version: 3.8
CUDA/cuDNN version: cuDNN 11.3
GPU models and configuration: RTX3060
Operating System：Windows

[samdow]

Hi, in general we use github issues only for bug reports or feature requests. Please use the forum to ask questions: https://discuss.pytorch.org/

This is being called an list instead of a bool because of the transform that norm does before calling the C++ function. Specifically it does
_dim = [i for i in range(ndim)]
and passes that to the function, seen in the source for norm

With these older versions, there may have been unintended behavior that we allowed. The user behavior when passing False here is ambiguous (did they really mean False or did they mean for it to be treated as an int). If they did intend for it to be treated as an int, they can cast it themselves

[triumph-wangyuyang]

Hi, in general we use github issues only for bug reports or feature requests. Please use the forum to ask questions: https://discuss.pytorch.org/

This is being called an list instead of a bool because of the transform that norm does before calling the C++ function. Specifically it does _dim = [i for i in range(ndim)] and passes that to the function, seen in the source for norm

With these older versions, there may have been unintended behavior that we allowed. The user behavior when passing False here is ambiguous (did they really mean False or did they mean for it to be treated as an int). If they did intend for it to be treated as an int, they can cast it themselves

Thank you for your reply, but the same code has different performances in different versions, which is quite confusing, especially this kind of underlying problem.