【Pytorch】Tensor. Use and understanding of continguous()

Official document address ：https://pytorch.org/docs/stable/generated/torch.Tensor.contiguous.html?highlight=torch%20contiguous#torch.Tensor.contiguous

form ：Tensor.contiguous(memory_format=torch.contiguous_format) → Tensor

effect ： Returns a continuous memory tensor , It contains the same data as the self tensor . If self tensor It is already the specified memory format , Then this function returns self tensor.

Parameters ：memory_format (torch.memory_format, optional) The memory format required for the returned tensor . The default value is ：torch.contiguous_format.

understand ：
This involves pytorch The relationship between the description and storage of tensors . some tensor operation , Will not change tensor Content itself , Instead, it redefines the correspondence between subscripts and elements . The corresponding relationship between subscript and element is Metadata . For example, there are ：

• transpose()
• view()

With transpose() For example , Its function is to tensor Do transpose , This operation will not generate a new tensor. The following code , after transpose After the operation , Changed the b,a And it's changing .

import torch

a = torch.randn(2, 3)
b = a.transpose(0, 1)
print("before modify")
print('a:', a)
print('b:', b)
b[1, 1] = 5
print("end modify")
print('a:', a)
print('b:', b)

The output is as follows ：

before modify
a: tensor([[-1.0912, -0.0348, -0.8400],
        [-0.2839,  0.0912, -1.3085]])
b: tensor([[-1.0912, -0.2839],
        [-0.0348,  0.0912],
        [-0.8400, -1.3085]])
end modify
a: tensor([[-1.0912, -0.0348, -0.8400],
        [-0.2839,  5.0000, -1.3085]])
b: tensor([[-1.0912, -0.2839],
        [-0.0348,  5.0000],
        [-0.8400, -1.3085]])

Then at this time , We just think a yes contiguous Of ,b It's not . Of course, we can also use the following functions to view ：

print(a.is_contiguous())  # True
print(b.is_contiguous())  # False

So use contiguous() The role of Namely ： Break the connection between these two tensors , The new tensor opens up a new memory space in memory .