1. LeNet（1998）
1 """ 2 note: 3 LeNet: 4 输入体：32*32*1 5 卷积核：5*5 6 步长：1 7 填充：无 8 池化：2*2 9 代码旁边的注释：卷积或者池化后的数据的尺寸 10 """ 11 import torch 12 import torch.nn as nn 13 14 15 class LeNet(nn.Module): 16 def __init__(self): 17 super(LeNet,self).__init__() 18 layer1 = nn.Sequential() 19 layer1.add_module('conv1',nn.Conv2d(1,6,5,1,padding=0))# 没有填充 ，b,6,28*28 20 layer1.add_module('pool1',nn.MaxPool2d(2,2)) # 6,14*14 (28-2)/2+1 = 14 21 self.layer1= layer1 22 23 layer2 = nn.Sequential() 24 layer2.add_module('conv2', nn.Conv2d(6, 16, 5, 1, padding=0)) # 没有填充 b,16,10*10 25 layer2.add_module('pool2', nn.MaxPool2d(2, 2)) # 16,5*5 26 self.layer2 = layer2 27 28 layer3 = nn.Sequential() 29 layer3.add_module('fc1',nn.Linear(400,120)) 30 layer3.add_module('fc2',nn.Linear(120,84)) 31 layer3.add_module('fc3',nn.Linear(84,10)) 32 self.layer3 = layer3 33 34 def forward(self,x): 35 x = self.layer1(x) 36 x = self.layer2(x) 37 x = x.view(x.size(0),-1) # 将多维数据排列成一行：1*400（16*5*5） 38 x = self.layer3(x) 39 return x
2.AlexNet（2012）:层数更深，同时第一次引入了激活层ReLU，在全连接层引入了Dropout层防止过拟合
3.VGGNet（2014）：有16~19层网络，使用了3*3的卷积滤波器和2*2的池化层。只是对网络层进行不断的堆叠，并没有太大的创新，增加深度缺失可以一定程度改善模型效果。
4.GoogleLeNet:(InceptionNet)(2014):比VGGNet更深的网络结构，一共22层，但是它的参数比AlexNet少了12倍，同时有很高的计算效率，因为它采用了一种有效的Inception模块，而且它也没有全连接层。Inception模块设计了一个局部的网络拓扑结构，然后将这些模块堆叠在一起形成一个抽象层次的网络结构。具体来说是运用几个并行的滤波器对输入进行卷积核池化，这些滤波器有不同的感受野，最后将输出的结果按深度拼接在一起形成输出层。缺点：参数太多，导致计算复杂。这些模块增加了一些1*1的卷积层来降低输入层的维度，使网络参数减少，从而减少网络的复杂性。
1 """ 2 GooglNet的Inceoption模块，整个GoogleNet都是由这些Inception模块组成的 3 nn.BatchNorm1d:在每个小批量数据中，计算输入各个维度的均值和标注差。 4 num_features:期望输入大小：batch_size * num_features 5 torch.cat:将不同尺度的卷积深度相加，只是深度不同，数据体大小是一样的 6 (0)表示增加行，（1）表示增加列 7 """ 8 9 import torch.nn as nn 10 import torch 11 import torch.nn.functional as F 12 13 14 class BasicConv2d(nn.Module): 15 def __init__(self,in_channels,out_channles,**kwargs): 16 super(BasicConv2d,self).__init__() 17 self.conv = nn.Conv2d(in_channels,out_channles,bias=False,**kwargs) 18 self.bn = nn.BatchNorm1d(out_channles,eps=0.001) 19 20 def forward(self,x): 21 x = self.conv(x) 22 x = self.bn(x) 23 return F.relu(x,inplace = True) 24 25 26 class Inception(nn.Module): 27 def __init__(self,in_channles,pool_features): 28 super(Inception,self).__init__() 29 self.branch1x1 = BasicConv2d(in_channles,64,kernel_size = 1) 30 31 self.branch5x5_1 = BasicConv2d(in_channles,48,kernel_size = 1) 32 self.branch5x5_2 = BasicConv2d(48,64,kernel_size = 5,padding = 2) 33 34 self.branch3x3dbl_1 = BasicConv2d(in_channles,64,kernel_size = 1) 35 self.branch3x3dbl_2 = BasicConv2d(64,96,kernel_size = 3,padding = 1) 36 #self.branch3x3dbl_3 = BasicConv2d(96,96,kernel_size = 3,padding = 1) 37 38 self.branch_pool = BasicConv2d(in_channles,pool_features,kenel_size = 1) 39 40 def forward(self, x): 41 branch1x1 = self.branch1x1(x) 42 43 branch5x5 = self.branch5x5_1(x) # 核是1 44 branch5x5 = self.branch5x5_2(branch5x5) #核是5 45 46 branch3x3 = self.branch3x3dbl_1(x) # 核是1 47 branch3x3 = self.branch3x3dbl_2(branch3x3) 48 49 branch_pool = F.avg_pool2d(x,kernel_size = 3,stride = 1,padding = 1) 50 branch_pool = self.branch_pool(branch_pool) 51 52 outputs = [branch1x1,branch5x5,branch3x3,branch_pool] 53 return torch.cat(outputs,1)
5.ResNet(2015)
在不断加深神经网络的时候，会出现准确率先上升然后达到饱和，再持续增加深度会导致模型准确率下降，这并不是过拟合问题，因为不仅在验证集上误差增加，训练集本身误差也会增加，假设一个比较浅的网络达到了饱和的准确率，那么在后面加上几个恒等的映射层，误差不会增加，也就是说更深的模型起码不会使得模型效果下降。假设某个神经网络的输入是x，期望输出值是H（x），如果直接把输入x传到输出作为初始结果，那么此时需要学习的目标就是F(x) = H（x）- x，即残差。ResNet相当于将学习目标改变了，不再学习一个完整的输出H（x），而是学习输出和输入的差别 H（x）-x
1 import torch 2 import torch.nn as nn 3 4 5 def conv3x3(in_planes,out_plans,stride = 1): 6 return nn.Conv2d( 7 in_planes, 8 out_plans,kernel_size=3, 9 stride=stride, 10 padding=1, 11 bias = False 12 ) 13 14 15 class BasicBlock(nn.Module): 16 def __init__(self,inplanes,planes,stride = 1,downsample = None): 17 super(BasicBlock,self).__init__() 18 self.conv1 = conv3x3(inplanes,planes,stride) 19 self.bn1 = nn.BatchNorm2d(planes) 20 self.relu = nn.ReLU(inplace=True) 21 self.conv2 = conv3x3(planes,planes) 22 self.bn2 = nn.BatchNorm2d(planes) 23 self.downsample = downsample 24 self.stride = stride 25 26 def forward(self,x): 27 residual = x 28 out = self.conv1(x) 29 30 out = self.bn1(out) 31 out = self.relu(out) 32 33 out = self.conv2(out) 34 out = self.bn2(out) 35 36 if self.downsample is not None: 37 residual = self.downsample(x) 38 39 out += residual 40 out = self.relu(out) 41 return out