最新！最全！深度学习特征提取主干网络总结（内含基本单元结构代码）
接下来为每个模型提供基本结构的代码，并对其亮点进行简要描述。以下代码是简化的示例：
1. ResNet
亮点：
-
残差连接： ResNet 引入了残差块，通过跳跃连接和恒等映射解决了梯度消失和爆炸问题，使得网络可以更轻松地学习恒等映射。
-
适应深度： 允许构建非常深的神经网络，成为深度学习中的里程碑。
import torch
import torch.nn as nn
class ResidualBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride=1):
super(ResidualBlock, 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.shortcut = nn.Sequential()
if stride != 1 or in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(out_channels)
)
def forward(self, x):
residual = self.shortcut(x)
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out += residual
out = self.relu(out)
return out
2. MobileNetV1
亮点：
-
深度可分离卷积： MobileNetV1 使用深度可分离卷积层，减少模型参数和计算量，适应移动设备和嵌入式系统。
-
高效性能： 设计旨在在保持高准确性的同时，降低模型大小和计算复杂度，适用于资源受限的环境。
import torch
import torch.nn as nn
class DepthwiseSeparableConv(nn.Module):
def __init__(self, in_channels, out_channels, stride):
super(DepthwiseSeparableConv, self).__init__()
self.depthwise = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=stride, padding=1, groups=in_channels, bias=False)
self.pointwise = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0, bias=False)
self.bn = nn.BatchNorm2d(out_channels)
def forward(self, x):
x = self.depthwise(x)
x = self.pointwise(x)
x = self.bn(x)
return x
class MobileNetV1(nn.Module):
def __init__(self):
super(MobileNetV1, self).__init__()
self.model = nn.Sequential(
nn.Conv2d(3, 32, kernel_size=3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(32),
nn.ReLU(inplace=True),
DepthwiseSeparableConv(32, 64, 1),
DepthwiseSeparableConv(64, 128, 2),
DepthwiseSeparableConv(128, 128, 1),
# ... Repeat layers as needed
)
def forward(self, x):
return self.model(x)
3. MobileNetV2
亮点：
-
倒残差块： MobileNetV2 引入了倒残差块，包括扩张卷积、融合层和线性瓶颈，提高了性能。
-
动态调整结构： MobileNetV2通过改变扩张比例和网络宽度，可以在不同的资源和准确性要求之间进行平衡。
import torch
import torch.nn as nn
class InvertedResidual(nn.Module):
def __init__(self, in_channels, out_channels, stride, expand_ratio):
super(InvertedResidual, self).__init__()
self.use_res_connect = stride == 1 and in_channels == out_channels
hidden_dim = int(round(in_channels * expand_ratio))
layers = []
if expand_ratio != 1:
layers.append(nn.Conv2d(in_channels, hidden_dim, kernel_size=1, bias=False))
layers.append(nn.BatchNorm2d(hidden_dim))
layers.append(nn.ReLU6(inplace=True))
layers.extend([
nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=stride, padding=1, groups=hidden_dim, bias=False),
nn.BatchNorm2d(hidden_dim),
nn.ReLU6(inplace=True),
nn.Conv2d(hidden_dim, out_channels, kernel_size=1, bias=False),
nn.BatchNorm2d(out_channels)
])
self.conv = nn.Sequential(*layers)
def forward(self, x):
if self.use_res_connect:
return x + self.conv(x)
else:
return self.conv(x)
4. ShuffleNetV2
亮点：
-
通道重排： ShuffleNetV2 使用通道重排（channel shuffling）来降低通道之间的相关性，减少计算量。
-
组卷积： 使用组卷积结构，进一步减小计算复杂度。
import torch
import torch.nn as nn
class ShuffleUnit(nn.Module):
def __init__(self, in_channels, out_channels, mid_channels, stride):
super(ShuffleUnit, self).__init__()
self.stride = stride
self.groups = 4
if stride == 1:
mid_channels = mid_channels // 2
self.branch1 = nn.Sequential(
nn.Conv2d(in_channels, mid_channels, kernel_size=1, bias=False),
nn.BatchNorm2d(mid_channels),
nn.ReLU(inplace=True),
)
self.branch2 = nn.Sequential(
nn.Conv2d(mid_channels, mid_channels, kernel_size=3, stride=stride, padding=1, groups=mid_channels, bias=False),
nn.BatchNorm2d(mid_channels),
nn.Conv2d(mid_channels, out_channels - mid_channels, kernel_size=1, bias=False),
nn.BatchNorm2d(out_channels - mid_channels),
nn.ReLU(inplace=True),
)
def forward(self, x):
if self.stride == 1:
x1, x2 = x.chunk(2, dim=1)
out = torch.cat([x1, self.branch1(x2), self.branch2(x2)], dim=1)
else:
out = torch.cat([self.branch1(x), self.branch2(x)], dim=1)
return out
5. EfficientNet
亮点：
-
复合缩放： EfficientNet 使用复合缩放（compound scaling）策略，通过调整网络深度、宽度和分辨率来平衡准确性和计算复杂度。
-
MBConv结构： 使用了MBConv（Mobile Inverted Bottleneck Convolution）结构，结合了轻量级的深度可分离卷积和通道注意力机制。
import torch
import torch.nn as nn
class MBConvBlock(nn.Module):
def __init__(self, in_channels, out_channels, expand_ratio, kernel_size, stride, se_ratio):
super(MBConvBlock, self).__init__()
self.use_se = se_ratio is not None and 0 < se_ratio <= 1
self.use_residual = stride == 1 and in_channels == out_channels
# Expansion phase
hidden_dim = int(round(in_channels * expand_ratio))
self.expand_conv = nn.Conv2d(in_channels, hidden_dim, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(hidden_dim)
self.swish = Swish()
# Depthwise convolution phase
self.dw_conv = nn.Conv2d(hidden_dim, hidden_dim, kernel_size=kernel_size, stride=stride, padding=kernel_size // 2, groups=hidden_dim, bias=False)
self.bn2 = nn.BatchNorm2d(hidden_dim)
# Squeeze and Excitation phase
if self.use_se:
num_squeezed_channels = max(1, int(in_channels * se_ratio))
self.se = SqueezeExcitation(hidden_dim, num_squeezed_channels)
else:
self.se = None
# Output phase
self.project_conv = nn.Conv2d(hidden_dim, out_channels, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(out_channels)
def forward(self, x):
# Expansion
x = self.expand_conv(x)
x = self.bn1(x)
x = self.swish(x)
# Depthwise convolution
x = self.dw_conv(x)
x = self.bn2(x)
x = self.swish(x)
# Squeeze and Excitation
if self.use_se:
x = self.se(x)
# Output
x = self.project_conv(x)
x = self.bn3(x)
# Skip connection and drop connect
if self.use_residual:
x = x + input
return x
6. GhostNet
亮点：
-
Ghost Module： GhostNet 使用 Ghost Module 结构，通过分组卷积和深度可分离卷积的结合，降低计算复杂度。
-
线性变换： 使用线性变换来进行信息的跨通道传递，减少信息的损失。
import torch
import torch.nn as nn
class GhostModule(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=1, ratio=2, dw_size=3, stride=1):
super(GhostModule, self).__init__()
self.primary_conv = nn.Sequential(
nn.Conv2d(in_channels, int(in_channels / ratio), kernel_size=kernel_size, stride=stride, padding=(kernel_size - 1) // 2, bias=False),
nn.BatchNorm2d(int(in_channels / ratio)),
nn.ReLU(inplace=True),
)
self.cheap_operation = nn.Sequential(
nn.Conv2d(int(in_channels / ratio), out_channels - in_channels, kernel_size=dw_size, stride=1, padding=dw_size // 2, groups=int(in_channels / ratio), bias=False),
nn.BatchNorm2d(out_channels - in_channels),
nn.ReLU(inplace=True),
)
def forward(self, x):
x1 = self.primary_conv(x)
x2 = self.cheap_operation(x1)
out = torch.cat([x, x2], 1)
return out
7. GoogLeNet (Inception)
亮点：
-
Inception Module： GoogLeNet 使用 Inception Module 结构，通过并行的卷积操作和池化操作来捕捉不同尺度的特征。
-
全局平均池化： 使用全局平均池化来减小参数量，提高计算效率。
import torch
import torch.nn as nn
class InceptionModule(nn.Module):
def __init__(self, in_channels, out1, mid2, out2, mid3, out3, out4):
super(InceptionModule, self).__init__()
self.branch1 = nn.Sequential(
nn.Conv2d(in_channels, out1, kernel_size=1),
nn.BatchNorm2d(out1),
nn.ReLU(inplace=True)
)
self.branch2 = nn.Sequential(
nn.Conv2d(in_channels, mid2, kernel_size=1),
nn.BatchNorm2d(mid2),
nn.ReLU(inplace=True),
nn.Conv2d(mid2, out2, kernel_size=3, padding=1),
nn.BatchNorm2d(out2),
nn.ReLU(inplace=True)
)
self.branch3 = nn.Sequential(
nn.Conv2d(in_channels, mid3, kernel_size=1),
nn.BatchNorm2d(mid3),
nn.ReLU(inplace=True),
nn.Conv2d(mid3, out3, kernel_size=5, padding=2),
nn.BatchNorm2d(out3),
nn.ReLU(inplace=True)
)
self.branch4 = nn.Sequential(
nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
nn.Conv2d(in_channels, out4, kernel_size=1),
nn.BatchNorm2d(out4),
nn.ReLU(inplace=True)
)
def forward(self, x):
branch1 = self.branch1(x)
branch2 = self.branch2(x)
branch3 = self.branch3(x)
branch4 = self.branch4(x)
outputs = [branch1, branch2, branch3, branch4]
return torch.cat(outputs, 1)
8. DenseNet
亮点：
-
密集连接： DenseNet 使用密集连接结构，通过将前一层的所有特征图与当前层的输入连接在一起，促使信息充分流动。
-
过渡层： 使用过渡层来控制特征图的维度，平衡参数量和计算复杂度。
import torch
import torch.nn as nn
class Bottleneck(nn.Module):
def __init__(self, in_channels, growth_rate):
super(Bottleneck, self).__init__()
self.bn1 = nn.BatchNorm2d(in_channels)
self.relu1 = nn.ReLU(inplace=True)
self.conv1 = nn.Conv2d(in_channels, 4 * growth_rate, kernel_size=1, bias=False)
self.bn2 = nn.BatchNorm2d(4 * growth_rate)
self.relu2 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(4 * growth_rate, growth_rate, kernel_size=3, padding=1, bias=False)
def forward(self, x):
out = self.conv1(self.relu1(self.bn1(x)))
out = self.conv2(self.relu2(self.bn2(out)))
out = torch.cat([x, out], 1)
return out
class DenseBlock(nn.Module):
def __init__(self, in_channels, growth_rate, num_layers):
super(DenseBlock, self).__init__()
layers = []
for i in range(num_layers):
layers.append(Bottleneck(in_channels + i * growth_rate, growth_rate))
self.layers = nn.Sequential(*layers)
def forward(self, x):
return self.layers(x)
9. HRNet (High-Resolution Network)
亮点：
-
多分辨率融合： HRNet 通过保持高分辨率特征图的信息，使用多分辨率的特征融合策略，更全面地捕捉图像的细节。
-
适应性结构： 允许根据任务需求调整网络结构，通过增加或减少分辨率分支的数量来适应不同场景。
import torch
import torch.nn as nn
class BasicBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride=1):
super(BasicBlock, self).__init__()
# ... (详细结构)
def forward(self, x):
# ... (详细结构)
return x
class HighResolutionBlock(nn.Module):
def __init__(self, in_channels, out_channels, num_modules, num_blocks, stride):
super(HighResolutionBlock, self).__init__()
self.blocks = nn.ModuleList([
BasicBlock(in_channels, out_channels, stride),
*[BasicBlock(out_channels, out_channels) for _ in range(num_blocks - 1)]
])
def forward(self, x):
for block in self.blocks:
x = block(x)
return x
10. MultiHeadAttention
import torch
import torch.nn as nn
class MultiHeadAttention(nn.Module):
def __init__(self, embed_size, num_heads):
super(MultiHeadAttention, self).__init__()
self.embed_size = embed_size
self.num_heads = num_heads
self.head_dim = embed_size // num_heads
self.values = nn.Linear(embed_size, embed_size, bias=False)
self.keys = nn.Linear(embed_size, embed_size, bias=False)
self.queries = nn.Linear(embed_size, embed_size, bias=False)
self.fc_out = nn.Linear(num_heads * self.head_dim, embed_size)
def forward(self, values, keys, query, mask):
# 自注意力机制的前向传播
N = query.shape[0]
value_len, key_len, query_len = values.shape[1], keys.shape[1], query.shape[1]
values = self.values(values).view(N, value_len, self.num_heads, self.head_dim)
keys = self.keys(keys).view(N, key_len, self.num_heads, self.head_dim)
queries = self.queries(query).view(N, query_len, self.num_heads, self.head_dim)
energy = torch.einsum("nqhd,nkhd->nhqk", [queries, keys])
if mask is not None:
energy = energy.masked_fill(mask == 0, float("-1e20"))
attention = torch.nn.functional.softmax(energy / (self.embed_size ** (1 / 2)), dim=3)
out = torch.einsum("nhql,nlhd->nqhd", [attention, values]).reshape(N, query_len, self.num_heads * self.head_dim)
out = self.fc_out(out)
return out