Pytoch foundation - (1) initialization of tensors

Import pytorch library

import torch

• Initialization tensor

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"  #  Set up the running equipment 
x = torch.tensor([[1, 2, 3], [4, 5, 6]], dtype=torch.float32, device=DEVICE, requires_grad=True)

• Initialize other methods

 1. produce 0-1 Between Uniform distribution Of 2x3 Tensor  

x = torch.rand((2, 3))
print(x)

2. produce 0-1 Between Normal distribution Of 2x3 Tensor

x = torch.randn((2, 3))
print(x)

3. produce 3-10 Between Random integers Of 2x3 Tensor

x = torch.randint(3, 10, (2, 3))
print(x)

4. Produce and input Tensor with the same format  

input = torch.rand((3, 3))
x = torch.rand_like(input)

5. The initial value generated is 0, In steps of 1, Sequences that do not contain termination values  

x = torch.arange(start=0, end=10, step=1)

 6. The initial value generated is 0, The end value is 9, The total number of steps is 11 Sequence

 x = torch.linspace(start=0, end=9, steps=11)

 7. The diagonal is 0-1 Uniformly distributed tensor

x = torch.diag(torch.rand(5))

•   Convert data type

We often use numpy data type , But it can't be directly in torch Middle operation , Therefore, you need to convert the data type .

import numpy as np

x = np.zeros((5, 5))
print(x)
print(x.dtype)

x_torch = torch.from_numpy(x)   #  from numpy become torch Tensor type 
print(x_torch)
print(x_torch.dtype)

x_back = x_torch.numpy()   #  from torch Tensor reversion numpy type 
print(x_back)
print(x_back.dtype)

Be careful ：

appear  Numpy is not available When reporting a mistake , yes numpy and pytorch The version of does not match .