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

#定義一個MLP網絡
class MLP(nn.Module):
    '''
    網絡里面主要是要定義__init__()、forward()
    '''
    def __init__(self):
        '''
        這里定義網絡有哪些層(比如nn.Linear,Conv2d……)[可不含激活函數]
        '''
        super().__init__()#調用Module（父）初始化
        self.hidden = nn.Linear(5,10)
        self.out = nn.Linear(10,2)
    
    def forward(self,x):
        '''
        這里定義前向傳播的順序,即__init__()中定義的層是按怎樣的順序進行連接以及傳播的[在這里加上激活函數,以構造復雜函數,提高擬合能力]
        '''
        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個samples，5個輸入屬性
out = net(x)
#也可調用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()等價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)
)

#訪問權重和偏差
print(net2[2].weight)#注意weight是parameter類型，.data訪問數值
print(net2[2].bias.data)

#輸出所有權重、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)   #將權重按照均值為0，標準差為0.01的正態分布進行初始化
        nn.init.zeros_(m.bias)  #將偏差置為0

def init_const(m):
    if type(m) == nn.Linear:
        nn.init.constant_(m.weight,42)   #將權重全部置為42
        
def my_init(m):
    if type(m) == nn.Linear:
        '''
        對weight和bias自定義初始化
        '''
        pass

#如何調用？
net2.apply(init_const)  #在net2中進行遍歷，對每個Linear執行初始化

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行數為dim_out，列數為dim_in，要做乘法，權重行列互換
        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類型變量

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

#存儲網絡

torch.save(net2.state_dict(),'Past_parameters') #把所有參數全部存儲
clone = nn.Sequential(MLP2(),nn.ReLU(),nn.Linear(4,2))   #存儲時同時存儲網絡定義（網絡結構）
clone.load_state_dict(torch.load('Past_parameters'))
clone.eval()