从零实现深度神经网络的正向传播

500个数据，20个特征，标签为3分类，

网络架构：第一层13个神经元，第二层8个，第三层输出

第一层激活函数relu，第二层 。

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

torch.random.manual_seed(22)
x = torch.rand((500,20),dtype=torch.float32)
y = torch.randint(low=0,high=3,size=(500,1),dtype=torch.float32)

class Model(nn.Module):
    def __init__(self,in_features=10,out_features=2):
        """
        :param in_features: input feature count
        :param out_features: output feature count
        """
        super(Model, self).__init__()# copy init in nn.Module,
        self.linear1 = nn.Linear(in_features,13)
        self.linear2 = nn.Linear(13,8)
        self.output = nn.Linear(8,out_features)
    def forward(self,x):
        """
        nn forward propagat
        Run each layer and add activation function
        :param x:
        :return:
        """
        z1 = self.linear1(x)
        relu1 = torch.relu(z1)
        z2 = self.linear2(relu1)
        sigma1 = torch.sigmoid(z2)
        out = self.output(sigma1)
        softmax1 = F.softmax(out,dim=1)
        return softmax1

if __name__ == '__main__':
    # instantiation nn
    torch.random.manual_seed(22)
    net = Model(in_features=x.shape[1],out_features=len(y.unique()))
    print(net.forward(x))
    print(net.linear1.weight.shape)
    print(net.linear2.weight.shape)
    print(net.output.weight.shape)

关于super，Model是nn.Module的子类，会继承其父类的方法（init之外的方法)，但是无法继承父类的init之内的属性，所以这里要用super。

nn.apply() # 对神经网络中的所有层，init中的所有对象都执行相同的操作

nn.parameters() # 神经网络生成的所有参数