Introducing pytorch to generate a random 5x3 tensor

>>> from __future__ import print_function
>>> import torch
>>> x = torch.rand(5, 3)
>>> print(x)
tensor([[0.5555, 0.7301, 0.5655],
        [0.9998, 0.1754, 0.7808],
        [0.5512, 0.8162, 0.6148],
        [0.8618, 0.3293, 0.6236],
        [0.2787, 0.0943, 0.2074]])

Declare a 5x3 tensor in which all elements are initialized to zero

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

Construct the tensor directly from the data, where the data is generally a python array

>>> x = torch.tensor([5.5, 3])
>>> print(x)
tensor([5.5000, 3.0000])

Similar to creating a new tensor from an existing tensor, the shape and data type of the new and old tensors are the same unless the dtype is overridden

>>> x = x.new_ones(5, 3, dtype=torch.double) 
>>> print(x)
tensor([[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]], dtype=torch.float64)
>>> y = torch.rand_like(x, dtype=torch.float)
>>> print(y)
tensor([[0.6934, 0.9637, 0.0594],
        [0.0863, 0.6638, 0.4728],
        [0.3416, 0.0892, 0.1761],
        [0.6831, 0.6404, 0.8307],
        [0.6254, 0.4180, 0.2174]])

Tensor size, numpy is shape

>>> print(x.size())
torch.Size([5, 3])

Tensor Addition

>>> x=torch.rand(5, 3)
>>> y = torch.zeros(5, 3)
>>> print(x + y)
tensor([[0.8991, 0.9222, 0.2050],
        [0.2478, 0.7688, 0.4156],
        [0.4055, 0.9526, 0.2559],
        [0.9481, 0.8576, 0.4816],
        [0.0767, 0.3346, 0.0922]])

>>> print(torch.add(x, y))
tensor([[0.8991, 0.9222, 0.2050],
        [0.2478, 0.7688, 0.4156],
        [0.4055, 0.9526, 0.2559],
        [0.9481, 0.8576, 0.4816],
        [0.0767, 0.3346, 0.0922]])

>>> result = torch.empty(5, 3)
>>> torch.add(x, y, out=result)
tensor([[0.8991, 0.9222, 0.2050],
        [0.2478, 0.7688, 0.4156],
        [0.4055, 0.9526, 0.2559],
        [0.9481, 0.8576, 0.4816],
        [0.0767, 0.3346, 0.0922]])

>>> y.add_(x)
tensor([[0.8991, 0.9222, 0.2050],
        [0.2478, 0.7688, 0.4156],
        [0.4055, 0.9526, 0.2559],
        [0.9481, 0.8576, 0.4816],
        [0.0767, 0.3346, 0.0922]])

Element access within a tensor is consistent with numpy, such as all elements where the second dimension superscript and superscript of output tensor y are 1

>>> print(y[:, 1])
tensor([0.9222, 0.7688, 0.9526, 0.8576, 0.3346])

The iew function changes tensor's shape, similar to numpy's reshape

>>> x = torch.randn(4, 4)
>>> y = x.view(16)  # Becomes a tensor of 1x16
>>> z = x.view(-1, 8)  # When the second dimension becomes 8, the first dimension automatically calculates the tensor, resulting in a 2x8 tensor
>>> print(x.size(), y.size(), z.size())
torch.Size([4, 4]) torch.Size([16]) torch.Size([2, 8])

A vector with only one element, take this element

>>> x = torch.randn(1)
>>> print(x)
tensor([0.8542])
>>> print(x.item())
0.8541867136955261

Convert to numpy array

>>> x = torch.rand(5, 3)
>>> x.numpy()
array([[0.9320856 , 0.473859  , 0.6787642 ],
       [0.14365482, 0.1112923 , 0.8280207 ],
       [0.4609589 , 0.51031697, 0.15313298],
       [0.18854082, 0.4548    , 0.49709243],
       [0.8351501 , 0.6160053 , 0.61391556]], dtype=float32)

All cpu tensors except CharTensor are converted from numpy to tensor

import numpy as np
a = np.ones(5)
b = torch.from_numpy(a)
np.add(a, 1, out=a)
print(a)
print(b)

Move tensor between cpu and gpu.

if torch.cuda.is_available():
    device = torch.device("cuda")          # a CUDA device object
    y = torch.ones_like(x, device=device)  # Create directly on the GPU device
    x = x.to(device)                       # or just use strings ``.to("cuda")``
    z = x + y
    print(z)
    print(z.to("cpu", torch.double))       # ``.to`` can also change dtype together!

The loss function measures the distance between the input and the target

from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F

class Net(nn.Module):
    ## Define the structure of the network
    def __init__(self):
        super(Net, self).__init__()
        ## input is channel 1, output 6 channels with 3x3 convulutionanl kernel
        self.conv1 = nn.Conv2d(1, 6, 3) 
        self.conv2 = nn.Conv2d(6, 16, 3)
        # an affine operation: y = Wx + b,  # 6*6 from image dimension
        self.fc1 = nn.Linear(16*6*6, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)
    
    ## Forward propagation, function name must be forward   
    def forward(self, x):
        # Max pooling over a (2, 2) window
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
        # If the size is a square you can only specify a single number
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        x = x.view(-1, self.num_flat_features(x))
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x
    
    def num_flat_features(self, x):
        size = x.size()[1:] # all dimensions except the batch dimension
        num_features = 1
        for s in size:
            num_features *= s
        return num_features
    
## Create a new Net object
net = Net() 
print(net)  
params = list(net.parameters())
print(len(params))
print(params[0].size())  # conv1's .weight

# Declare a 4D tensor of 1x1x32x32 as input to the network
input = torch.randn(1, 1, 32, 32) 
# input = torch.randn(1, 1, 32, 32)
output = net(input) 

# net.zero_grad()
# out.backward(torch.randn(1, 10))
target = torch.randn(10) 
target = target.view(1, -1) 
criterion = nn.MSELoss()
loss = criterion(output, target)
print(loss)

print(loss.grad_fn)  # MSELoss
print(loss.grad_fn.next_functions[0][0])  # Linear
print(loss.grad_fn.next_functions[0][0].next_functions[0][0])  # ReLU

net.zero_grad()     # zeroes the gradient buffers of all parameters

print('conv1.bias.grad before backward')
print(net.conv1.bias.grad)

loss.backward()

print('conv1.bias.grad after backward')
print(net.conv1.bias.grad)

learning_rate = 0.01
for f in net.parameters():
    f.data.sub_(f.grad.data * learning_rate)

However, torch has implemented various weight updates, such as SGD, Nesterov-SGD, Adam, RMSProp, etc., which can be called directly

import torch.optim as optim

# create your optimizer
optimizer = optim.SGD(net.parameters(), lr=0.01)

# in your training loop:
optimizer.zero_grad()   # zero the gradient buffers
output = net(input)
loss = criterion(output, target)
loss.backward()
optimizer.step()    # Does the update