分类神经网络3：DenseNet模型复现

目录

DenseNet网络架构

DenseNet部分实现代码

DenseNet网络架构

论文原址：https://arxiv.org/pdf/1608.06993.pdf

稠密连接神经网络（DenseNet）实质上是ResNet的进阶模型（了解ResNet模型请点击），二者均是通过建立前面层与后面层之间的“短路连接”，但不同的是，DenseNet建立的是前面所有层与后面层的密集连接，其一大特点是通过特征在通道上的连接来实现特征重用，这让DenseNet在参数和计算成本更少的情形下实现比ResNet更优的性能。DenseNet 网络的模型结构如下：

DenseNet 的网络结构主要由DenseBlock和Transition Layer组成。

DenseBlock：密集连接机制。互相连接所有的层，即每一层的输入都来自于它前面所有层的特征图，每一层的输出均会直接连接到它后面所有层的输入，这可以实现特征重用（即对不同“级别”的特征——不同表征进行总体性地再探索），提升效率。具体的连接方式如下图示：

在同一个DenseBlock当中，特征层的高宽不会发生改变，但是通道数会发生改变。可以看出，DenseBlock中采用了BN+ReLU+Conv的结构，然而一般网络是用Conv+BN+ReLU的结构。这是由于卷积层的输入包含了它前面所有层的输出特征，它们来自不同层的输出，因此数值分布差异比较大，所以它们在输入到下一个卷积层时，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。通常为了减少参数，一般还会先加一个1x1 卷积来减少参数量。所以DenseBlock中的每一层采用BN+ReLU+1x1Conv 、Conv+BN+ReLU+3x3 Conv的结构。

Transition Layer：用于将不同DenseBlock之间进行连接，整合上一个DenseBlock获得的特征，并且缩小上一个DenseBlock的宽高，达到下采样的效果，实质上起到压缩模型的作用。Transition Layer中一般包含一个1x1卷积（用于调整通道数）和2x2平均池化（用于降低特征图大小），结构为BN+ReLU+1x1 Conv+2x2 AvgPooling。

DenseNet网络的具体配置信息如下：

可以看出，一个DenseNet中一般有3个或4个DenseBlock，最后的DenseBlock后连接了一个最大池化层，然后是一个包含1000个类别的全连接层，通过softmax激活函数得到类别属性。

DenseNet部分实现代码

直接上干货

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp
from collections import OrderedDict
 
__all__ = ["densenet121", "densenet161", "densenet169", "densenet201"]
 
 
class DenseLayer(nn.Module):
    def __init__(self, num_input_features, growth_rate, bn_size, drop_rate, memory_efficient = False):
        super(DenseLayer,self).__init__()
        self.norm1 = nn.BatchNorm2d(num_input_features)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv1 = nn.Conv2d(num_input_features, bn_size * growth_rate, kernel_size=1, stride=1, bias=False)
 
        self.norm2 = 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 = float(drop_rate)
        self.memory_efficient = memory_efficient
 
    def bn_function(self, inputs):
        concated_features = torch.cat(inputs, 1)
        bottleneck_output = self.conv1(self.relu1(self.norm1(concated_features)))
        return bottleneck_output
 
    def any_requires_grad(self, input):
        for tensor in input:
            if tensor.requires_grad:
                return True
        return False
 
    @torch.jit.unused
    def call_checkpoint_bottleneck(self, input):
        def closure(*inputs):
            return self.bn_function(inputs)
 
        return cp.checkpoint(closure, *input)
 
    def forward(self, input):
        if isinstance(input, torch.Tensor):
            prev_features = [input]
        else:
            prev_features = input
 
        if self.memory_efficient and self.any_requires_grad(prev_features):
            if torch.jit.is_scripting():
                raise Exception("Memory Efficient not supported in JIT")
 
            bottleneck_output = self.call_checkpoint_bottleneck(prev_features)
        else:
            bottleneck_output = self.bn_function(prev_features)
 
        new_features = self.conv2(self.relu2(self.norm2(bottleneck_output)))
        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):
    def __init__(self,num_layers,num_input_features,bn_size,growth_rate,
                 drop_rate,memory_efficient = False,):
        super(DenseBlock,self).__init__()
        for i in range(num_layers):
            layer = DenseLayer(
                num_input_features + 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():
            new_features = layer(features)
            features.append(new_features)
        return torch.cat(features, 1)
 
 
class Transition(nn.Sequential):
    """
    Densenet Transition Layer:
        1 × 1 conv
        2 × 2 average pool, stride 2
    """
    def __init__(self, num_input_features, num_output_features):
        super(Transition,self).__init__()
        self.norm = nn.BatchNorm2d(num_input_features)
        self.relu = nn.ReLU(inplace=True)
        self.conv = nn.Conv2d(num_input_features, num_output_features, kernel_size=1, stride=1, bias=False)
        self.pool = nn.AvgPool2d(kernel_size=2, stride=2)
 
 
class DenseNet(nn.Module):
    def __init__(self,growth_rate = 32,num_init_features = 64,block_config = None,num_classes = 1000,
                 bn_size = 4,drop_rate = 0.,memory_efficient = False,):
 
        super(DenseNet,self).__init__()
 
        # First convolution
        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)),
                ]
            )
        )
 
        # Each denseblock
        num_features = num_init_features
        for i, num_layers in enumerate(block_config):
            block = DenseBlock(
                num_layers=num_layers,
                num_input_features=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)
            num_features = num_features + num_layers * growth_rate
            if i != len(block_config) - 1:
                trans = Transition(num_input_features=num_features, num_output_features=num_features // 2)
                self.features.add_module("transition%d" % (i + 1), trans)
                num_features = num_features // 2
 
        # Final batch norm
        self.features.add_module("norm5", nn.BatchNorm2d(num_features))
        # Linear layer
        self.classifier = nn.Linear(num_features, num_classes)
 
        # Official init from torch repo.
        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)
        out = F.relu(features, inplace=True)
        out = F.adaptive_avg_pool2d(out, (1, 1))
        out = torch.flatten(out, 1)
        out = self.classifier(out)
        return out
 
 
def densenet121(num_classes):
    """Densenet-121 model"""
    return DenseNet(32, 64, (6, 12, 24, 16),num_classes=num_classes)
 
def densenet161(num_classes):
    """Densenet-161 model"""
    return DenseNet(48, 96, (6, 12, 36, 24),  num_classes=num_classes)
 
def densenet169(num_classes):
    """Densenet-169 model"""
    return DenseNet(32, 64, (6, 12, 32, 32),   num_classes=num_classes)
 
def densenet201(num_classes):
    """Densenet-201 model"""
    return DenseNet(32, 64, (6, 12, 48, 32), num_classes=num_classes)
 
 
 
if __name__=="__main__":
    # from torchsummaryX import summary
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    input = torch.ones(2, 3, 224, 224).to(device)
    net = densenet121(num_classes=4)
    net = net.to(device)
    out = net(input)
    print(out)
    print(out.shape)
    # summary(net, torch.ones((1, 3, 224, 224)).to(device))

希望对大家能够有所帮助呀！