pytorch简单实现神经网络功能

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

#定义一个MLP网络
class MLP(nn.Module):
    '''
    网络里面主要是要定义__init__()、forward()
    '''
    def __init__(self):
        '''
        这里定义网络有哪些层(比如nn.Linear,Conv2d……)[可不含激活函数]
        '''
        super().__init__()#调用Module（父）初始化
        self.hidden = nn.Linear(5,10)
        self.out = nn.Linear(10,2)
    
    def forward(self,x):
        '''
        这里定义前向传播的顺序,即__init__()中定义的层是按怎样的顺序进行连接以及传播的[在这里加上激活函数,以构造复杂函数,提高拟合能力]
        '''
        return self.out(F.relu(self.hidden(x)))

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

net = nn.Sequential(nn.Linear(5,10),nn.ReLU(),nn.Linear(10,2))

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

net = MLP()
x = torch.randn((15,5))#15个samples，5个输入属性
out = net(x)
#也可调用forward->"out = net.forward(x)"
print(out)
#print(out.shape)

tensor([[-0.0760, -0.1026],
        [-0.3277, -0.2332],
        [-0.0314, -0.1921],
        [ 0.0131, -0.1473],
        [-0.0650, -0.2310],
        [ 0.3009, -0.5510],
        [ 0.1491, -0.0928],
        [-0.1438, -0.1304],
        [-0.1945, -0.1944],
        [ 0.1088, -0.2249],
        [ 0.0016, -0.2334],
        [ 0.1401, -0.3709],
        [-0.1864, -0.1764],
        [ 0.0775, -0.0160],
        [ 0.0150, -0.3198]], grad_fn=<AddmmBackward>)

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

class MLP2(nn.Module):
    def __init__(self):
        super().__init__()
        self.net = nn.Sequential(nn.Linear(5,10),nn.ReLU(),nn.Linear(10,5))
        self.out = nn.Linear(5,4)
   
    def forward(self,x):
        return self.out(F.relu(self.net(x)))

net2 = nn.Sequential(MLP2(),nn.ReLU(),nn.Linear(4,2))
net2.eval()
# eval()等价print(net2)

Sequential(
  (0): MLP2(
    (net): Sequential(
      (0): Linear(in_features=5, out_features=10, bias=True)
      (1): ReLU()
      (2): Linear(in_features=10, out_features=5, bias=True)
    )
    (out): Linear(in_features=5, out_features=4, bias=True)
  )
  (1): ReLU()
  (2): Linear(in_features=4, out_features=2, bias=True)
)

#访问权重和偏差
print(net2[2].weight)#注意weight是parameter类型，.data访问数值
print(net2[2].bias.data)

#输出所有权重、偏差
print(*[(name,param) for name,param in net2[2].parameters()])

shared = nn.Linear(8,8)

net = nn.Sequential(nn.Linear(5,8),nn.ReLU(),shared,nn.ReLU(),shared)
print(net[2].weight.data[0])
net[2].weight.data[0][0] = 100
print(net[2].weight.data[0][0])
print(net[2].weight.data[0] == net[4].weight.data[0])
net.eval()

def init_Linear(m):
    if type(m) == nn.Linear:
        nn.init.normal_(m.weight,mean = 0,std = 0.01)   #将权重按照均值为0，标准差为0.01的正态分布进行初始化
        nn.init.zeros_(m.bias)  #将偏差置为0

def init_const(m):
    if type(m) == nn.Linear:
        nn.init.constant_(m.weight,42)   #将权重全部置为42
        
def my_init(m):
    if type(m) == nn.Linear:
        '''
        对weight和bias自定义初始化
        '''
        pass

#如何调用？
net2.apply(init_const)  #在net2中进行遍历，对每个Linear执行初始化

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

#a
class decentralized(nn.Module):
    def __init__(self):
        super().__init__()
    
    def forward(self,x):
        return x-x.mean()

#b
class my_Linear(nn.Module):
    def __init__(self,dim_in,dim_out):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(dim_in,dim_out))  #由于x行数为dim_out，列数为dim_in，要做乘法，权重行列互换
        self.bias = nn.Parameter(torch.randn(dim_out))
    
    def forward(self,x):
        return F.relu(torch.matmul(x,self.weight.data)+self.bias.data)

tmp = my_Linear(5,3)
print(tmp.weight)

#存取任意torch类型变量

x = torch.randn((20,20))
torch.save(x,'X')   #存
y = torch.load('X') #取

#存储网络

torch.save(net2.state_dict(),'Past_parameters') #把所有参数全部存储
clone = nn.Sequential(MLP2(),nn.ReLU(),nn.Linear(4,2))   #存储时同时存储网络定义（网络结构）
clone.load_state_dict(torch.load('Past_parameters'))
clone.eval()