神经网络架构参考：2-1 卷积篇

提示词：

给出{xxx}的网络结构表格，包含层名称、类型、输入大小（HWC），输出大小（HWC）、核尺寸、步长、参数数量

AlexNet

层名称	类型	输入大小（HWC）	输出大小（HWC）	核尺寸	步长	参数数量
输入层	输入	227x227x3	-	-	-	0
Conv1	卷积层	227x227x3	55x55x96	11x11	4	961111*3 + 96 = 34944
MaxPool1	最大池化层	55x55x96	27x27x96	3x3	2	0
LRN1	局部响应归一化	27x27x96	27x27x96	-	-	-
Conv2	卷积层	27x27x96	27x27x256	5x5	1	25655*96 + 256 = 614656
MaxPool2	最大池化层	27x27x256	13x13x256	3x3	2	0
LRN2	局部响应归一化	13x13x256	13x13x256	-	-	-
Conv3	卷积层	13x13x256	13x13x384	3x3	1	38433*256 + 384 = 885120
Conv4	卷积层	13x13x384	13x13x384	3x3	1	38433*384 + 384 = 1327488
Conv5	卷积层	13x13x384	13x13x256	3x3	1	25633*384 + 256 = 884992
MaxPool3	最大池化层	13x13x256	6x6x256	3x3	2	0
FC6	全连接层	6x6x256	4096	-	-	66256*4096 + 4096 = 37752832
FC7	全连接层	4096	4096	-	-	4096*4096 + 4096 = 16781312
FC8	全连接层	4096	1000	-	-	4096*1000 + 1000 = 4194304

PyTorch 源码

import torch
import torch.nn as nn
import torch.nn.functional as F
class AlexNet(nn.Module):
    def __init__(self, num_classes=1000):
        super(AlexNet, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 96, kernel_size=11, stride=4, padding=2),  # Conv1
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),  # MaxPool1
            nn.LocalResponseNorm(size=5, alpha=0.0001, beta=0.75, k=2),  # LRN1
            nn.Conv2d(96, 256, kernel_size=5, padding=2),  # Conv2
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),  # MaxPool2
            nn.LocalResponseNorm(size=5, alpha=0.0001, beta=0.75, k=2),  # LRN2
            nn.Conv2d(256, 384, kernel_size=3, padding=1),  # Conv3
            nn.ReLU(inplace=True),
            nn.Conv2d(384, 384, kernel_size=3, padding=1),  # Conv4
            nn.ReLU(inplace=True),
            nn.Conv2d(384, 256, kernel_size=3, padding=1),  # Conv5
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),  # MaxPool3
        )
        self.classifier = nn.Sequential(
            nn.Dropout(),
            nn.Linear(256 * 6 * 6, 4096),  # FC6
            nn.ReLU(inplace=True),
            nn.Dropout(),
            nn.Linear(4096, 4096),  # FC7
            nn.ReLU(inplace=True),
            nn.Linear(4096, num_classes),  # FC8
        )
    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), 256 * 6 * 6)
        x = self.classifier(x)
        return x
# 创建AlexNet模型实例
model = AlexNet(num_classes=1000)
print(model)

LENET5

网络结构

层名称	类型	输入大小 (HWC)	输出大小 (HWC)	核尺寸	步长	参数数量
输入层	输入	32x32x1	32x32x1	-	-	0
C1	卷积层	32x32x1	28x28x6	5x5	1	(5x5x1+1)x6 = 156
S2	下采样层	28x28x6	14x14x6	2x2	2	0
C3	卷积层	14x14x6	10x10x16	5x5	1	(5x5x6+1)x16 = 2416
S4	下采样层	10x10x16	5x5x16	2x2	2	0
C5	卷积层	5x5x16	1x1x120	5x5	1	(5x5x16+1)x120 = 48120
F6	全连接层	1x1x120	1x1x84	-	-	120x84 + 84 = 10164
输出层	全连接层	1x1x84	1x1x10	-	-	84x10 + 10 = 850

PyTorch 代码

import torch
import torch.nn as nn
import torch.nn.functional as F
class LeNet5(nn.Module):
    def __init__(self, num_classes=10):
        super(LeNet5, self).__init__()
        # Convolutional layer (C1)
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5, padding=2)
        # Subsampling layer (S2)
        self.pool1 = nn.AvgPool2d(kernel_size=2, stride=2)
        # Convolutional layer (C3)
        self.conv2 = nn.Conv2d(in_channels=6, out_channels=16, kernel_size=5)
        # Subsampling layer (S4)
        self.pool2 = nn.AvgPool2d(kernel_size=2, stride=2)
        # Convolutional layer (C5)
        self.conv3 = nn.Conv2d(in_channels=16, out_channels=120, kernel_size=5)
        # Fully connected layer (F6)
        self.fc1 = nn.Linear(in_features=120, out_features=84)
        # Output layer
        self.fc2 = nn.Linear(in_features=84, out_features=num_classes)
    def forward(self, x):
        # C1
        x = self.conv1(x)
        x = F.relu(x)
        # S2
        x = self.pool1(x)
        # C3
        x = self.conv2(x)
        x = F.relu(x)
        # S4
        x = self.pool2(x)
        # C5
        x = self.conv3(x)
        x = F.relu(x)
        # Flatten the output for the fully connected layer
        x = x.view(-1, self.num_flat_features(x))
        # F6
        x = self.fc1(x)
        x = F.relu(x)
        # Output layer
        x = self.fc2(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
# Example of creating the LeNet5 model
model = LeNet5(num_classes=10)
print(model)
# Example input tensor (batch size of 1, 1 channel, 32x32 image)
input_tensor = torch.randn(1, 1, 32, 32)
# Forward pass through the model
output = model(input_tensor)
print(output)

VGG16

层名称	类型	输入大小 (HWC)	输出大小 (HWC)	核尺寸	步长	参数数量
Input	-	224x224x3	-	-	-	0
Conv1_1	Conv2D	224x224x3	224x224x64	3x3	1	1792
Conv1_2	Conv2D	224x224x64	224x224x64	3x3	1	36928
MaxPool1	MaxPooling2D	224x224x64	112x112x64	2x2	2	0
Conv2_1	Conv2D	112x112x64	112x112x128	3x3	1	73856
Conv2_2	Conv2D	112x112x128	112x112x128	3x3	1	147584
MaxPool2	MaxPooling2D	112x112x128	56x56x128	2x2	2	0
Conv3_1	Conv2D	56x56x128	56x56x256	3x3	1	295168
Conv3_2	Conv2D	56x56x256	56x56x256	3x3	1	590080
Conv3_3	Conv2D	56x56x256	56x56x256	3x3	1	590080
MaxPool3	MaxPooling2D	56x56x256	28x28x256	2x2	2	0
Conv4_1	Conv2D	28x28x256	28x28x512	3x3	1	1180160
Conv4_2	Conv2D	28x28x512	28x28x512	3x3	1	2359808
Conv4_3	Conv2D	28x28x512	28x28x512	3x3	1	2359808
MaxPool4	MaxPooling2D	28x28x512	14x14x512	2x2	2	0
Conv5_1	Conv2D	14x14x512	14x14x512	3x3	1	2359808
Conv5_2	Conv2D	14x14x512	14x14x512	3x3	1	2359808
Conv5_3	Conv2D	14x14x512	14x14x512	3x3	1	2359808
MaxPool5	MaxPooling2D	14x14x512	7x7x512	2x2	2	0
Flatten	Flatten	7x7x512	25088	-	-	0
FC6	Dense	25088	4096	-	-	102760448
FC7	Dense	4096	4096	-	-

PyTorch 代码

import torch
import torch.nn as nn
class VGG16(nn.Module):
    def __init__(self, num_classes=1000):
        super(VGG16, self).__init__()
        self.features = nn.Sequential(
            # Conv1
            nn.Conv2d(3, 64, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(64, 64, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            
            # Conv2
            nn.Conv2d(64, 128, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(128, 128, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            
            # Conv3
            nn.Conv2d(128, 256, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, 256, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, 256, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            
            # Conv4
            nn.Conv2d(256, 512, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2),
            
            # Conv5
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2, stride=2)
        )
        
        self.classifier = nn.Sequential(
            nn.Linear(512 * 7 * 7, 4096),
            nn.ReLU(inplace=True),
            nn.Dropout(),
            nn.Linear(4096, 4096),
            nn.ReLU(inplace=True),
            nn.Dropout(),
            nn.Linear(4096, num_classes)
        )
        
    def forward(self, x):
        x = self.features(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x
# 实例化模型
model = VGG16(num_classes=1000)
print(model)

Inception

层名称	类型	输入大小（HWC）	输出大小（HWC）	核尺寸	步长	参数数量
Conv2d_1a_3x3	卷积层	299x299x3	149x149x32	3x3	2	864
Conv2d_2a_3x3	卷积层	149x149x32	147x147x32	3x3	1	9216
Conv2d_2b_3x3	卷积层	147x147x32	147x147x64	3x3	1	18432
MaxPool_3a_3x3	最大池化层	147x147x64	73x73x64	3x3	2	0
Conv2d_3b_1x1	卷积层	73x73x64	73x73x80	1x1	1	5120
Conv2d_4a_3x3	卷积层	73x73x80	71x71x192	3x3	1	138240
MaxPool_5a_3x3	最大池化层	71x71x192	35x35x192	3x3	2	0
Mixed_5b	Inception模块	35x35x192	35x35x256	-	-	-
Mixed_5c	Inception模块	35x35x256	35x35x288	-	-	-
Mixed_5d	Inception模块	35x35x288	35x35x288	-	-	-
Mixed_6a	Inception模块	35x35x288	17x17x768	-	2	-
Mixed_6b	Inception模块	17x17x768	17x17x768	-	-	-
Mixed_6c	Inception模块	17x17x768	17x17x768	-	-	-
Mixed_6d	Inception模块	17x17x768	17x17x768	-	-	-
Mixed_6e	Inception模块	17x17x768	17x17x768	-	-	-
Mixed_7a	Inception模块	17x17x768	8x8x1280	-	2	-
Mixed_7b	Inception模块	8x8x1280	8x8x2048	-	-	-
Mixed_7c	Inception模块	8x8x2048	8x8x2048	-	-	-

以Mixed_5b为例，列出其内部结构。

层名称	类型	输入大小（HWC）	输出大小（HWC）	核尺寸	步长	参数数量
Mixed_5b/1x1	卷积层	35x35x192	35x35x64	1x1	1	12288
Mixed_5b/3x3/1x1	卷积层	35x35x192	35x35x64	1x1	1	12288
Mixed_5b/3x3/3x3	卷积层	35x35x64	35x35x96	3x3	1	63360
Mixed_5b/5x5/1x1	卷积层	35x35x192	35x35x16	1x1	1	3072
Mixed_5b/5x5/5x5	卷积层	35x35x16	35x35x16	5x5	1	3072
Mixed_5b/pool	池化层	35x35x192	35x35x32	-	1	0
Mixed_5b/output	Concatenate	-	35x35x256	-	-	-

PyTorch 源码

以下是使用PyTorch构建InceptionV3模型的一部分源码。这个源码展示了如何定义Inception模块和一些辅助函数，但不包括整个网络的所有细节。完整的InceptionV3模型定义会更长，这里只提供了核心部分。

import torch
import torch.nn as nn
import torch.nn.functional as F
class InceptionA(nn.Module):
    def __init__(self, in_channels):
        super(InceptionA, self).__init__()
        self.branch1x1 = BasicConv2d(in_channels, 64, kernel_size=1)
        self.branch5x5_1 = BasicConv2d(in_channels, 48, kernel_size=1)
        self.branch5x5_2 = BasicConv2d(48, 64, kernel_size=5, padding=2)
        self.branch3x3dbl_1 = BasicConv2d(in_channels, 64, kernel_size=1)
        self.branch3x3dbl_2 = BasicConv2d(64, 96, kernel_size=3, padding=1)
        self.branch3x3dbl_3 = BasicConv2d(96, 96, kernel_size=3, padding=1)
        self.branch_pool = BasicConv2d(in_channels, 32, kernel_size=1)
    def forward(self, x):
        branch1x1 = self.branch1x1(x)
        branch5x5 = self.branch5x5_1(x)
        branch5x5 = self.branch5x5_2(branch5x5)
        branch3x3dbl = self.branch3x3dbl_1(x)
        branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl)
        branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl)
        branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)
        branch_pool = self.branch_pool(branch_pool)
        outputs = [branch1x1, branch5x5, branch3x3dbl, branch_pool]
        return torch.cat(outputs, 1)
class BasicConv2d(nn.Module):
    def __init__(self, in_channels, out_channels, **kwargs):
        super(BasicConv2d, self).__init__()
        self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs)
        self.bn = nn.BatchNorm2d(out_channels, eps=0.001)
    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        return F.relu(x, inplace=True)
class InceptionV3(nn.Module):
    def __init__(self, num_classes=1000):
        super(InceptionV3, self).__init__()
        # Stem
        self.Conv2d_1a_3x3 = BasicConv2d(3, 32, kernel_size=3, stride=2)
        self.Conv2d_2a_3x3 = BasicConv2d(32, 32, kernel_size=3)
        self.Conv2d_2b_3x3 = BasicConv2d(32, 64, kernel_size=3, padding=1)
        # ... additional stem layers ...
        # Inception modules
        self.Mixed_5b = InceptionA(256)
        self.Mixed_5c = InceptionA(288)
        # ... additional Inception modules ...
        # Auxiliary Logits
        self.AuxLogits = None
        # ... auxiliary logits layers ...
        # Final Logits
        self.Mixed_7c = InceptionA(768)
        # ... additional final layers ...
        self.fc = nn.Linear(2048, num_classes)
    def forward(self, x):
        # Stem
        x = self.Conv2d_1a_3x3(x)
        x = self.Conv2d_2a_3x3(x)
        x = self.Conv2d_2b_3x3(x)
        # ... additional stem layers ...
        # Inception modules
        x = self.Mixed_5b(x)
        x = self.Mixed_5c(x)
        # ... additional Inception modules ...
        # Auxiliary Logits
        if self.AuxLogits is not None:
            aux = self.AuxLogits(x)
        else:
            aux = None
        # Final Logits
        x = self.Mixed_7c(x)
        # ... additional final layers ...
        x = F.adaptive_avg_pool2d(x, (1, 1))
        x = torch.flatten(x, 1)
        x = self.fc(x)
        return x, aux
# Example usage:
# model = InceptionV3(num_classes=1000)

Resnet18

层名称	类型	输入大小（HWC）	输出大小（HWC）	核尺寸	步长	参数数量
Conv1	卷积层	224x224x3	112x112x64	7x7	2	9472
BatchNorm1	批归一化层	112x112x64	112x112x64	-	-	256
ReLU1	激活层	112x112x64	112x112x64	-	-	0
MaxPool1	最大池化层	112x112x64	56x56x64	3x3	2	0
ResidualBlock1_1	残差块	56x56x64	56x56x64	-	-	8448
ResidualBlock1_2	残差块	56x56x64	56x56x64	-	-	8448
ResidualBlock2_1	残差块	56x56x64	28x28x128	-	2	43008
ResidualBlock2_2	残差块	28x28x128	28x28x128	-	-	43008
ResidualBlock3_1	残差块	28x28x128	14x14x256	-	2	172448
ResidualBlock3_2	残差块	14x14x256	14x14x256	-	-	172448
AvgPool	平均池化层	14x14x256	7x7x256	7x7	2	0
Flatten	展平层	7x7x256	12544	-	-	0
FC	全连接层	12544	1000	-	-	12545000
Softmax	Softmax层	1000	1000	-	-	0

每个残差块的结构：

阶段	残差块	层名称	类型	输入大小（HWC）	输出大小（HWC）	核尺寸	步长	参数数量
1	1	conv1	卷积	224x224x64	112x112x64	7x7	2	9408
		conv2	卷积	112x112x64	112x112x64	3x3	1	18432
		skip1	卷积	224x224x64	112x112x64	1x1	2	256
1	2	conv1	卷积	112x112x64	112x112x64	3x3	1	18432
		conv2	卷积	112x112x64	112x112x64	3x3	1	18432
2	1	conv1	卷积	112x112x64	56x56x128	3x3	2	73984
		conv2	卷积	56x56x128	56x56x128	3x3	1	147584
		skip1	卷积	112x112x64	56x56x128	1x1	2	832
2	2	conv1	卷积	56x56x128	56x56x128	3x3	1	147584
		conv2	卷积	56x56x128	56x56x128	3x3	1	147584
3	1	conv1	卷积	56x56x128	28x28x256	3x3	2	295168
		conv2	卷积	28x28x256	28x28x256	3x3	1	589824
		skip1	卷积	56x56x128	28x28x256	1x1	2	3328
3	2	conv1	卷积	28x28x256	28x28x256	3x3	1	589824
		conv2	卷积	28x28x256	28x28x256	3x3	1	589824
4	1	conv1	卷积	28x28x256	14x14x512	3x3	2	1180928
		conv2	卷积	14x14x512	14x14x512	3x3	1	2359296

PyTorch 代码

import torch
import torch.nn as nn
import torch.nn.functional as F
# 定义基本残差块
class BasicBlock(nn.Module):
    expansion = 1
    def __init__(self, in_channels, out_channels, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, 
                               stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, 
                               stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(out_channels)
        self.downsample = downsample
    def forward(self, x):
        identity = x
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        if self.downsample is not None:
            identity = self.downsample(x)
        out += identity
        out = self.relu(out)
        return out
# 定义ResNet网络
class ResNet(nn.Module):
    def __init__(self, block, layers, num_classes=1000):
        super(ResNet, self).__init__()
        self.in_channels = 64
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)
        # 初始化权重
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)
    def _make_layer(self, block, out_channels, blocks, stride=1):
        downsample = None
        if stride != 1 or self.in_channels != out_channels * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.in_channels, out_channels * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(out_channels * block.expansion),
            )
        layers = []
        layers.append(block(self.in_channels, out_channels, stride, downsample))
        self.in_channels = out_channels * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.in_channels, out_channels))
        return nn.Sequential(*layers)
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)
        return x
# 实例化ResNet-16模型
def resnet16(pretrained=False, **kwargs):
    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
    if pretrained:
        # 这里没有预训练权重，如果需要预训练，可以在这里加载
        pass
    return model
# 创建模型实例
model = resnet16()
print(model)