Pytoch Foundation

install Pytorch

pip install torch

One 、 What is? Pytorch

Pytorch It's based on Numpy Scientific computing package for , It provides two functions to its users

• As Numpy The replacement of , Provide users with access to GPU Powerful capabilities
• As a platform for deep learning , Provide users with maximum flexibility and speed

Two 、 Basic element operations

• Tensor tensor ： Be similar to Numpy Medium ndarray data structure , The biggest difference is Tensor You can use GPU The acceleration function of

1、 Operation of creating matrix

• Create an uninitialized matrix

import torch

x = torch.empty(5, 3)
print(x)

• Create a matrix with initialization
  The value conforms to the standard Gaussian distribution

import torch

x = torch.rand(5, 3)
print(x)

Be careful : Compare whether there is an initialized matrix , When declaring an uninitialized matrix , It itself does not contain any exact value , When creating an uninitialized matrix , Any value in the memory allocated to the matrix is assigned to the matrix , Essentially meaningless data

• Create an all zero matrix and specify the type of data element as long

import torch

x = torch.zeros(5, 3, dtype=torch.long)
print(x)

• Create tensors from data

import torch

x = torch.tensor([2.5, 3.5, 2.5])
print(x)

• Create a new tensor of the same size from an existing tensor

x = torch.zeros(5, 3, dtype=torch.long)

#  utilize new_ones Method to get a tensor 
x = x.new_ones(5, 3, dtype=torch.double)
print(x)

#  utilize randn_like Method to get a new tensor with the same tensor size , And use random initialization to assign values 
y = torch.randn_like(x, dtype=torch.float)
print(y)

• Get the size of the tensor

x.size()

Note that the return value is a tuple , Support all tuple operations

3、 ... and 、 Basic operation

• Addition operation

result = torch.empty(5, 3)
x = torch.rand(5, 3)
y = torch.rand(5, 3)

print(x + y)  #  The first addition 

print(torch.add(x, y))  #  The second addition 1
torch.add(x, y, out=result)  #  The second addition 2
print(result)

y.add_(x)  #  Third addition 
print(y)

The results are the same

• Slice of tensor

x = torch.randn(4, 4)
print(x[:, :1])

• Change the shape of the tensor

x = torch.randn(4, 4)
# torch.view() The operation needs to ensure that the total number of data elements remains unchanged 
y = x.view(16)
# -1 Number of automatic matches 
z = x.view(-1, 8)
print(x.size(), y.size(), z.size())

• If there is only one element in the tensor , It can be used .iem() Take out the value , As python Of number type

x = torch.randn(1)
print(x)
print(x.item())

Four 、Torch Tensor and Numpy array Mutual conversion between

• Torch Tensor and Numpy array Share the underlying memory space , So change one of the values , The other will change with it

a = torch.ones(5)
print(a)
b = a.numpy()
print(b, type(b))

a = numpy.ones(5)
b = torch.from_numpy(a)
print(a)
print(b)

5、 ... and 、GPU and CPU Transfer between

x = torch.randn(1)
#  If installed on the server GPU and CUDA
if torch.cuda.is_available():
    #  Define a device object , This is designated as CUDA, namely GPU
    device = torch.device("cuda")
    #  Directly in GPU Create a Tensor
    y = torch.ones_like(x, device=device)
    #  Will be in CPU above x Tensor moves to GPU above 
    x = x.to(device)
    # x and y All in GPU above , In order to add 
    z = x + y
    #  The tensor at this time z stay GPU above 
    print(z)
    #  Can also be z Transferred to the CPU above , And specify the element type of the tensor 
    print(z.to("cpu",torch.double))