Difference in shape of tensor torch.Size([]) and torch.Size([1]) in pytorch

You can play with tensors having the single scalar value like this:

import torch

t = torch.tensor(1)
print(t, t.shape) # tensor(1) torch.Size([])

t = torch.tensor([1])
print(t, t.shape) # tensor([1]) torch.Size([1])

t = torch.tensor([[1]])
print(t, t.shape) # tensor([[1]]) torch.Size([1, 1])

t = torch.tensor([[[1]]])
print(t, t.shape) # tensor([[[1]]]) torch.Size([1, 1, 1])

t = torch.unsqueeze(t, 0)
print(t, t.shape) # tensor([[[[1]]]]) torch.Size([1, 1, 1, 1])

t = torch.unsqueeze(t, 0)
print(t, t.shape) # tensor([[[[[1]]]]]) torch.Size([1, 1, 1, 1, 1])

t = torch.unsqueeze(t, 0)
print(t, t.shape) # tensor([[[[[[1]]]]]]) torch.Size([1, 1, 1, 1, 1, 1])

#squize dimension with id 0
t = torch.squeeze(t,dim=0)
print(t, t.shape) # tensor([[[[[1]]]]]) torch.Size([1, 1, 1, 1, 1])

#back to beginning.
t = torch.squeeze(t)
print(t, t.shape) # tensor(1) torch.Size([])

print(type(t)) # <class 'torch.Tensor'>
print(type(t.data)) # <class 'torch.Tensor'>

Tensors, do have a size or shape. Which is the same. Which is actually a class torch.Size. You can write help(torch.Size) to get more info. Any time you write t.shape, or t.size() you will get that size info.

The idea of tensors is they can have different compatible size dimension for the data inside it including torch.Size([]).

Any time you unsqueeze a tensor it will add another dimension of 1. Any time you squeeze a tensor it will remove dimensions of 1, or in general case all dimensions of one.

First one has 0 size dimension, second one has 1 dimension, PyTorch tries to make both compatible (0 size can be regarded similarly to float or a-like although I haven't really met the case where it's explicitly needed, except what @javadr shown in his answer below).

Usually you would use list to initialize it though, see here for more information.

Look at the documentation of tensor in pytorch:

Docstring:
tensor(data, dtype=None, device=None, requires_grad=False, pin_memory=False) -> Tensor

Constructs a tensor with :attr:`data`.

then it describes what the data is:

Args:
    data (array_like): Initial data for the tensor. Can be a list, tuple,
        NumPy ``ndarray``, scalar, and other types.

As you can see the data could be a scalar (which is a data with the dimension of zero).

Thus, in response to your question the tensor(58) is a tensor with dimension 0 and the tensor([58]) is a tensor with dimension 1.