【图像分类】【深度学习】【Pytorch版本】 DenseNet模型算法详解

【图像分类】【深度学习】【Pytorch版本】 DenseNet模型算法详解

---

前言

DenseNet是康奈尔大学的Huang, Gao等人在《Densely Connected Convolutional Networks【CVPR-2017】》【论文地址】一文中提出的模型，受ResNet基本思路启发，将前面所有网络层输出特征与后面网络层输出特征进行密集连接(而非相加)从而实现特征重用，减少了参数量和计算成本并达到了显著的效果。

---

DenseNet讲解

随着卷积神经网络的网络层越来越深入，出现了新的研究问题：当输入(前向传播)或梯度(反向传播)的信息经过许多层时，信息可能会在到达网络末尾或起始时消失并“冲洗掉”，即网络退化问题。尽管不同的方法在网络拓扑和训练过程方面有所不同，但是都具有一个关键特点：这些方法创建了从浅层到深层的短路径。

从深度方向研究，ResNet【参考】解决了深层网络梯度消失问题；从宽度方向研究，GoogleNet【参考】解决了宽度增加时的计算和参数量的问题。DenseNet则是是从特征的角度入手，通过对特征的极致利用达到更好的效果和减少参数。

DenseNet架构将这种见解(关键特点)提炼为简单的连接模式：为了确保最大程度的信息在网络中各层之间流动，将所有层(相匹配的特征图大小)彼此直接连接。为了保留前馈特性，每个层都从所有先前的图层获取附加输入，并将其自身的特征图传递给所有后续层。

与ResNets将特征传递到图层之前通过求和来组合特征相比，DenseNet通过级联特征来组合它们。因此，第$L$层具有$L$个输入，由所有前面的卷积块的特征图组成。

Dense Block(稠密块)

DenseBlock包含很多互相连接的DenseLayer层，每个层的输出特征图大小相同，这样才可以在通道上进行连结，具体来说就是每一层的输入都来自于它前面所有层的特征，每一层的输出均会直接连接到它后面所有层的输入，层与层之间采用密集连接方式。

所以对于$L$层的DenseBlock，采用密集连接方式共包含$\frac{L\left(L+1\right)}{2}$个连接(等差数列求和公式)，相比$L$层的ResNetBlock则只包含$\left(L?1\right)\times 2+1$个连接。DenseNet与ResNet最主要的区别在于DenseBlock是直接concat来自不同层的特征图实现特征重用，即对不同“级别”的不同表征进行总体性地再探索，提升效率；ResNetBlock是直接add上一层的特征图实现特征融合，将之前层次的信息直接传递给后续层次，解决梯度消失问题。
ResNet和DenseNet对比如下图所示：

• ResNet: 传统的卷积前馈网络将第${?}^{\mathrm{th}}$层的输出$x_{?}$作为输入连接到第$(?+1{)}^{\mathrm{th}}$层得到输出$x_{?+1}=H_{?+1}(x_{?})$，ResNet增加了一个跳跃连接(通过一个恒等函数绕过非线性转换)，得到输出$x_{?+1}=H_{?+1}(x_{?})+x_{?}$。
• DenseNet: 为了进一步改进层之间的信息流，将任何层直接连接到所有后续层，将第$1^{\mathrm{th}}$,$2^{\mathrm{th}}$,$3^{\mathrm{th}}$…${?}^{\mathrm{th}}$层的输出$x_1$,$x_2$,$x_3$…$x_{?}$作为输入连接到第${(\ell
  • 1)^{{\rm{th}}}}$层得到输出${x_{\ell + 1}} = {H_{\ell + 1}}(\left[ {{x_1},{x_2},{x_3}…{x_\ell }} \right])$

$H_{?+1}(?)$定义为三个连续操作的复合函数：批量归一化层(BN)、激活函数(ReLU)[6]和卷积层(Conv)。

增长率 如果每一个DenseLayer$H_{?}(?)$生成$k$个特征图，则第$?$层具有$k_0+k\left(??1\right)$个特征图，其中$k_0$是输入DenseBlock的特征图数量，$k$指的是网络的超参数增长率。每个DenseLayer都可以访问其块中的所有先前的特征图(即网络的“集体知识”)，这些特征图被视为网络的全局状态，每个DenseLayer都将自己$k$个特征图添加到全局状态，因此增长率控制着每一层为全局状态贡献多少新信息。写入后的全局状态可以在网络中的任何位置进行访问，并且与传统的网络体系结构不同，无需将其逐层复制。

DenseNet与现有网络体系结构之间的一个重要区别是DenseNet具有通道非常狭窄的DenseLayer，通常DenseBlock起始的DenseLayer就是通道就十分狭窄($k=12$)，因为要考虑后续DenseLayer的增长。

Dense Layer(稠密层)

DenseLayer中采用的BN+ReLU+Conv的结构模式，不同于常见的是Conv+BN+ReLU，这是因为稠密层的输入包含了之前所有稠密层的输出特征，这些来自不同层的输出数值分布差异比较大，因此在输入到DenseLayer的Conv层之前，必须先经过BN层将其数值进行标准化，然后再进行卷积操作。
DenseLayer中采用BN+ReLU+1x1 Conv+BN+ReLU+3x3 Conv的结构：

瓶颈层 尽管每个DenseLayer仅产生$k$个输出特征图，但由于密集连接模式，DenseLayer通常会有更多的输入。因此为了减少参数，DenseLayer在3×3卷积之前引入1×1卷积作为瓶颈层，以减少输入特征图的数量，从而提高计算效率。

Transition Layer 过渡层

TransitionLayer进行卷积和池化以起到整合压缩的作用：卷积网络的重要组成部分是降低特征图尺寸的下采样层。为了便于DenseNet结构进行下采样，将DenseNet划分为为多个DenseBlock，TransitionLayer主要用于连接两个相邻的DenseBlock，整合前一个DenseBlock输出的特征，通过下采样缩小前一个DenseBlock的征图尺寸。
TransitionLayer中采用BN+ReLU+1x1 Conv+ 2x2 AvgPooling的结构：

压缩 为了进一步提高DenseNet的紧凑性，可以减少TransitionLayer的特征图数量。如果DenseBlock输出$m$个特征图，则让之后的TransitionLayer输出${\theta }_m$个特征图。其中$0<{\theta }_m\le 1$称为压缩因子。当${\theta }_m=1$时，TransitionLayer的特征图数量与DenseBlock的保持不变。

DenseNet模型结构

下图是原论文给出的关于DenseNet模型结构的详细示意图：

Resnet在图像分类中分为两部分：backbone部分： 主要由卷积层、池化层(汇聚层)、稠密块和过渡层组成，分类器部分：由全局平均池化层和全连接层组成 。

---

DenseNet Pytorch代码

稠密层Dense Layer： BN+ReLU+1x1 Conv+BN+ReLU+3x3 Conv

class DenseLayer(nn.Module):
    """Basic unit of DenseBlock (DenseLayer) """
    def __init__(self, input_c, growth_rate, bn_size, drop_rate):
        super(DenseLayer, self).__init__()

        self.bn1 = nn.BatchNorm2d(input_c)
        self.relu1 = nn.ReLU()
        self.conv1 = nn.Conv2d(input_c, bn_size * growth_rate,
                               kernel_size=1, stride=1, bias=False)

        self.bn2 = nn.BatchNorm2d(bn_size * growth_rate)
        self.relu2 = nn.ReLU()
        self.conv2 = nn.Conv2d(bn_size * growth_rate, growth_rate,
                               kernel_size=3, stride=1, padding=1, bias=False)
        self.drop_rate = drop_rate

    def forward(self, x):
        # 1×1卷积 bottleneck瓶颈层
        output = self.bn1(x)
        output = self.relu1(output)
        output = self.conv1(output)
        # 3×3卷积
        output = self.bn2(output)
        output = self.relu2(output)
        output = self.conv2(output)

        if self.drop_rate > 0:
            output = F.dropout(output, p=self.drop_rate)
        return torch.cat([x, output], 1)

稠密块DenseBlock： 由多个稠密层组成

class DenseBlock(nn.Module):
    def __init__(self, num_layers, input_c, bn_size, growth_rate, drop_rate):
        super(DenseBlock, self).__init__()
        for i in range(num_layers):
            layer = DenseLayer(input_c + i * growth_rate,
                                growth_rate=growth_rate,
                                bn_size=bn_size,
                                drop_rate=drop_rate)
            self.add_module("denselayer%d" % (i + 1), layer)
    def forward(self, init_features):
        features = [init_features]
        for name, layer in self.items():
            # 当前DenseLayer的输出特征
            new_features = layer(features)
            # 拼接所有DenseLayer的输出特征
            features.append(new_features)
        return torch.cat(features, 1)

过渡层TransitionLayer ： BN+ReLU+1x1 Conv+ 2x2 AvgPooling

class Transition(nn.Module):
    def __init__(self, input_c, output_c):
        super(Transition, self).__init__()
        self.bn = nn.BatchNorm2d(input_c)
        self.relu = nn.ReLU(inplace=True)
        # 1×1卷积
        self.conv = nn.Conv2d(input_c, output_c,
                              kernel_size=1, stride=1, bias=False)
        # 2×2池化
        self.pool = nn.AvgPool2d(2, stride=2)

    def forward(self, input):
        output = self.bn(input)
        output = self.relu(output)
        output = self.conv(output)
        output = self.pool(output)
        return output

---

完整代码

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchsummary import summary
class DenseLayer(nn.Module):
    """Basic unit of DenseBlock (DenseLayer) """
    def __init__(self, input_c, growth_rate, bn_size, drop_rate):
        super(DenseLayer, self).__init__()

        self.bn1 = nn.BatchNorm2d(input_c)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv1 = nn.Conv2d(input_c, bn_size * growth_rate,
                               kernel_size=1, stride=1, bias=False)

        self.bn2 = nn.BatchNorm2d(bn_size * growth_rate)
        self.relu2 = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(bn_size * growth_rate, growth_rate,
                               kernel_size=3, stride=1, padding=1, bias=False)
        self.drop_rate = drop_rate

    def forward(self, inputs):
        # 1×1卷积 bottleneck瓶颈层
        output = self.bn1(inputs)
        output = self.relu1(output)
        output = self.conv1(output)
        # 3×3卷积
        output = self.bn2(output)
        output = self.relu2(output)
        output = self.conv2(output)

        if self.drop_rate > 0:
            output = F.dropout(output, p=self.drop_rate)
        return output

class DenseBlock(nn.ModuleDict):
    def __init__(self, num_layers, input_c, bn_size, growth_rate, drop_rate):
        super(DenseBlock, self).__init__()
        for i in range(num_layers):
            layer = DenseLayer(input_c + i * growth_rate,
                                growth_rate=growth_rate,
                                bn_size=bn_size,
                                drop_rate=drop_rate)
            self.add_module("denselayer%d" % (i + 1), layer)
    def forward(self, init_features):
        features = [init_features]
        for name, layer in self.items():
            concat_features = torch.cat(features, 1)
            # 当前DenseLayer的输出特征
            new_features = layer(concat_features)
            # 收集所有DenseLayer的输出特征
            features.append(new_features)
        return torch.cat(features, 1)

class Transition(nn.Module):
    def __init__(self, input_c, output_c):
        super(Transition, self).__init__()
        self.bn = nn.BatchNorm2d(input_c)
        self.relu = nn.ReLU(inplace=True)
        # 1×1卷积
        self.conv = nn.Conv2d(input_c, output_c,
                              kernel_size=1, stride=1, bias=False)
        # 2×2池化
        self.pool = nn.AvgPool2d(2, stride=2)

    def forward(self, input):
        output = self.bn(input)
        output = self.relu(output)
        output = self.conv(output)
        output = self.pool(output)
        return output


class DenseNet(nn.Module):
    def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16), num_init_features=64, bn_size=4,
                 compression_rate=0.5, drop_rate=0, num_classes=1000):
        super(DenseNet, self).__init__()

        # 前部 conv+bn+relu+pool
        self.features = nn.Sequential(
            # 第一层
            nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False),
            nn.BatchNorm2d(num_init_features),
            nn.ReLU(inplace=True),
            # 第二层
            nn.MaxPool2d(3, stride=2, padding=1)
        )

        # 中部 DenseBlock
        num_features = num_init_features
        for i, num_layers in enumerate(block_config):
            block = DenseBlock(num_layers=num_layers,
                               input_c=num_features,
                               bn_size=bn_size,
                               growth_rate=growth_rate,
                               drop_rate=drop_rate)
            # 新增DenseBlock
            self.features.add_module("denseblock%d" % (i + 1), block)
            # 更新通道数
            num_features = num_features + num_layers * growth_rate

            # 除去最后一层DenseBlock不需要加Transition来连接两个相邻的DenseBlock
            if i != len(block_config) - 1:
                transition = Transition(input_c=num_features, output_c=int(num_features * compression_rate))
                # 添加Transition
                self.features.add_module("transition%d" % (i + 1), transition)
                # 更新通道数
                num_features = int(num_features * compression_rate)

        # 后部 bn+ReLU
        self.tail = nn.Sequential(
            nn.BatchNorm2d(num_features),
            nn.ReLU(inplace=True)
        )

        # 分类器 classification
        self.classifier = nn.Linear(num_features, num_classes)

        # 初始化权重
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)
            elif isinstance(m, nn.Linear):
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        features = self.features(x)
        tail_output = self.tail(features)
        # 平均池化
        out = F.adaptive_avg_pool2d(tail_output, (1, 1))
        out = torch.flatten(out, 1)
        out = self.classifier(out)
        return out

def densenet121(**kwargs):
    # Top-1 error: 25.35%
    # 'densenet121': 'https://download.pytorch.org/models/densenet121-a639ec97.pth'
    return DenseNet(growth_rate=32,
                    block_config=(6, 12, 24, 16),
                    num_init_features=64,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)
def densenet169(**kwargs):
    # Top-1 error: 24.00%
    # 'densenet169': 'https://download.pytorch.org/models/densenet169-b2777c0a.pth'
    return DenseNet(growth_rate=32,
                    block_config=(6, 12, 32, 32),
                    num_init_features=64,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)

def densenet201(**kwargs):
    # Top-1 error: 22.80%
    # 'densenet201': 'https://download.pytorch.org/models/densenet201-c1103571.pth'
    return DenseNet(growth_rate=32,
                    block_config=(6, 12, 48, 32),
                    num_init_features=64,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)


def densenet161(**kwargs):
    # Top-1 error: 22.35%
    # 'densenet161': 'https://download.pytorch.org/models/densenet161-8d451a50.pth'
    return DenseNet(growth_rate=48,
                    block_config=(6, 12, 36, 24),
                    num_init_features=96,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)

if __name__ == '__main__':
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    model = densenet121().to(device)
    summary(model, input_size=(3, 224, 224))

summary可以打印网络结构和参数，方便查看搭建好的网络结构。

---

附加代码

由于需要进行多次Concatnate操作，数据需要被复制多次，显存容易增加得很快，需要一定的显存优化技术。这里博主提供了优化后的代码。

import re

from collections import OrderedDict
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp
from torchsummary import summary

class _DenseLayer(nn.Module):
    def __init__(self,
                 input_c,
                 growth_rate,
                 bn_size,
                 drop_rate,
                 memory_efficient= False):
        super(_DenseLayer, self).__init__()

        self.add_module("norm1", nn.BatchNorm2d(input_c))
        self.add_module("relu1", nn.ReLU(inplace=True))

        self.add_module("conv1", nn.Conv2d(in_channels=input_c,
                                           out_channels=bn_size * growth_rate,
                                           kernel_size=1,
                                           stride=1,
                                           bias=False))

        self.add_module("norm2", nn.BatchNorm2d(bn_size * growth_rate))
        self.add_module("relu2", nn.ReLU(inplace=True))
        self.add_module("conv2", nn.Conv2d(bn_size * growth_rate,
                                           growth_rate,
                                           kernel_size=3,
                                           stride=1,
                                           padding=1,
                                           bias=False))
        self.drop_rate = drop_rate
        self.memory_efficient = memory_efficient

    def bn_function(self, inputs):
        # 1×1卷积
        bottleneck_output = self.conv1(self.relu1(self.norm1(inputs)))
        return bottleneck_output

    # 所有连接都能保留梯度,否则不进行梯度更新
    def any_requires_grad(inputs):
        for tensor in inputs:
            if tensor.requires_grad:
                return True
        return False

    # 本函数主要是为了网络训练阶段高效内存管理
    def call_checkpoint_bottleneck(self, inputs):
        def closure(*inp):
            return self.bn_function(inp)
        return cp.checkpoint(closure, *inputs)

    def forward(self, inputs):
        if self.memory_efficient and self.any_requires_grad(inputs):
            # 即非部署阶段
            if torch.jit.is_scripting():
                raise Exception("memory efficient not supported in JIT")
            bottleneck_output = self.call_checkpoint_bottleneck(inputs)
        else:
            bottleneck_output = self.bn_function(inputs)

        # 3×3卷积
        new_features = self.conv2(self.relu2(self.norm2(bottleneck_output)))
        # 随机失活
        if self.drop_rate > 0:
            new_features = F.dropout(new_features,
                                     p=self.drop_rate,
                                     training=self.training)
        return new_features

class _DenseBlock(nn.ModuleDict):
    _version = 2

    def __init__(self,
                 num_layers,
                 input_c,
                 bn_size,
                 growth_rate,
                 drop_rate,
                 memory_efficient=False):
        super(_DenseBlock, self).__init__()
        for i in range(num_layers):
            layer = _DenseLayer(input_c + i * growth_rate,
                                growth_rate=growth_rate,
                                bn_size=bn_size,
                                drop_rate=drop_rate,
                                memory_efficient=memory_efficient)
            self.add_module("denselayer%d" % (i + 1), layer)

    def forward(self, init_features):
        features = [init_features]
        for name, layer in self.items():
            concat_features = torch.cat(features, 1)
            new_features = layer(concat_features)
            features.append(new_features)
        return torch.cat(features, 1)

class _Transition(nn.Sequential):
    def __init__(self,
                 input_c,
                 output_c):
        super(_Transition, self).__init__()
        self.add_module("norm", nn.BatchNorm2d(input_c))
        self.add_module("relu", nn.ReLU(inplace=True))
        self.add_module("conv", nn.Conv2d(input_c,
                                          output_c,
                                          kernel_size=1,
                                          stride=1,
                                          bias=False))
        self.add_module("pool", nn.AvgPool2d(kernel_size=2, stride=2))


class DenseNet(nn.Module):
    def __init__(self,
                 growth_rate=32,
                 block_config=(6, 12, 24, 16),
                 num_init_features=64,
                 bn_size=4,
                 compression_rate=0.5,
                 drop_rate=0,
                 num_classes=1000,
                 memory_efficient= False):
        super(DenseNet, self).__init__()

        # 前部 conv+bn+relu+pool
        self.features = nn.Sequential(OrderedDict([
            ("conv0", nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)),
            ("norm0", nn.BatchNorm2d(num_init_features)),
            ("relu0", nn.ReLU(inplace=True)),
            ("pool0", nn.MaxPool2d(kernel_size=3, stride=2, padding=1)),
        ]))

        # 中部 DenseBlock
        num_features = num_init_features
        for i, num_layers in enumerate(block_config):
            block = _DenseBlock(num_layers=num_layers,
                                input_c=num_features,
                                bn_size=bn_size,
                                growth_rate=growth_rate,
                                drop_rate=drop_rate,
                                memory_efficient=memory_efficient)
            self.features.add_module("denseblock%d" % (i + 1), block)
            # 更新通道数
            num_features = num_features + num_layers * growth_rate
            # 除去最后一层DenseBlock不需要加Transition来连接两个相邻的DenseBlock
            if i != len(block_config) - 1:
                trans = _Transition(input_c=num_features,
                                    output_c=num_features // 2)
                self.features.add_module("transition%d" % (i + 1), trans)
                # 更新通道数
                num_features = int(num_features * compression_rate)

        # 后部 bn+ReLU
        self.tail = nn.Sequential(
            nn.BatchNorm2d(num_features),
            nn.ReLU(inplace=True)
        )

        # 分类器 classification
        self.classifier = nn.Linear(num_features, num_classes)

        # 初始化权重
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)
            elif isinstance(m, nn.Linear):
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        features = self.features(x)
        tail_output = self.tail(features)
        # 平均池化
        out = F.adaptive_avg_pool2d(tail_output, (1, 1))
        out = torch.flatten(out, 1)
        out = self.classifier(out)
        return out


def densenet121(**kwargs):
    # Top-1 error: 25.35%
    # 'densenet121': 'https://download.pytorch.org/models/densenet121-a639ec97.pth'
    return DenseNet(growth_rate=32,
                    block_config=(6, 12, 24, 16),
                    num_init_features=64,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)


def densenet169(**kwargs):
    # Top-1 error: 24.00%
    # 'densenet169': 'https://download.pytorch.org/models/densenet169-b2777c0a.pth'
    return DenseNet(growth_rate=32,
                    block_config=(6, 12, 32, 32),
                    num_init_features=64,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)


def densenet201(**kwargs):
    # Top-1 error: 22.80%
    # 'densenet201': 'https://download.pytorch.org/models/densenet201-c1103571.pth'
    return DenseNet(growth_rate=32,
                    block_config=(6, 12, 48, 32),
                    num_init_features=64,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)


def densenet161(**kwargs):
    # Top-1 error: 22.35%
    # 'densenet161': 'https://download.pytorch.org/models/densenet161-8d451a50.pth'
    return DenseNet(growth_rate=48,
                    block_config=(6, 12, 36, 24),
                    num_init_features=96,
                    bn_size=4,
                    compression_rate=0.5,
                    **kwargs)

if __name__ == '__main__':
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    model = densenet121().to(device)
    summary(model, input_size=(3, 224, 224))

---

总结

尽可能简单、详细的介绍了稠密结构的原理和在卷积神经网络中的作用，讲解了DenseNet模型的结构和pytorch代码。