您现在的位置是:首页 >技术交流 >将DenseNet换成Resnet——更换深度学习骨干网络网站首页技术交流

将DenseNet换成Resnet——更换深度学习骨干网络

是啊洋呀 2024-06-15 12:01:02
简介将DenseNet换成Resnet——更换深度学习骨干网络

最近我在学习一个手写公式识别的网络,这个网络的backbone使用的是DenseNet,我想将其换成ResNet
至于为什么要换呢,因为我还没换过骨干网络,就像单纯拿来练练手,增加我对网络的熟悉程度,至于会不会对模型的性能有所提升,这我不知道。废话不多说,直接开干

这个网络中使用的是DenseNet-100,这里的100是这么来的
100=(16+16+16)2+1(77的卷积)+3(transition layer)

论文中给出的DenseNet代码如下

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from thop import profile


# DenseNet-B
#Bottleneck又称为瓶颈层,因为其长得像一个瓶子,两头大中间细,其主要目的是为了减少计算量
class Bottleneck(nn.Module):
    def __init__(self, nChannels, growthRate, use_dropout):
        super(Bottleneck, self).__init__()
        interChannels = 4 * growthRate
        self.bn1 = nn.BatchNorm2d(interChannels)#对输入的数据进行批量标准化
        self.conv1 = nn.Conv2d(nChannels, interChannels, kernel_size=1, bias=False)#一层1*1的卷积层,nChannels为输入通道数,interChannels为输出通道数
        self.bn2 = nn.BatchNorm2d(growthRate)#归一化
        self.conv2 = nn.Conv2d(interChannels, growthRate, kernel_size=3, padding=1, bias=False)#一层3*3的卷积层
        self.use_dropout = use_dropout#每次训练时随机丢掉一些神经元防止过拟合
        self.dropout = nn.Dropout(p=0.2)#20%的神经元被丢弃

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)), inplace=True)#对第一层的卷积进行relu操作
        if self.use_dropout:
            out = self.dropout(out)#20%不更新
        out = F.relu(self.bn2(self.conv2(out)), inplace=True)#对第二层的卷积进行relu操作
        if self.use_dropout:
            out = self.dropout(out)##20%不更新
        out = torch.cat((x, out), 1)#将输入的初始量和经过特征提取的量相加
#        print(out.shape)
        return out


# single layer 进行了一层3*3的特征提取,并与初始量进行相加
#这段代码没用上-----------------------------------------------------------------------
class SingleLayer(nn.Module):
    def __init__(self, nChannels, growthRate, use_dropout):
        super(SingleLayer, self).__init__()
        self.bn1 = nn.BatchNorm2d(nChannels)
        self.conv1 = nn.Conv2d(nChannels, growthRate, kernel_size=3, padding=1, bias=False)
        self.use_dropout = use_dropout
        self.dropout = nn.Dropout(p=0.2)

    def forward(self, x):
        out = self.conv1(F.relu(x, inplace=True))
        if self.use_dropout:
            out = self.dropout(out)
        out = torch.cat((x, out), 1)
        return out
# ----------------------------------------------------------------------------------


# transition layer
class Transition(nn.Module):
    def __init__(self, nChannels, nOutChannels, use_dropout):
        super(Transition, self).__init__()
        self.bn1 = nn.BatchNorm2d(nOutChannels)#标准化
        self.conv1 = nn.Conv2d(nChannels, nOutChannels, kernel_size=1, bias=False)#1*1卷积,调整维度
        self.use_dropout = use_dropout
        self.dropout = nn.Dropout(p=0.2)

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)), inplace=True)
        # print('relu:')
        # print(out.shape)
        if self.use_dropout:
            out = self.dropout(out)
        out = F.avg_pool2d(out, 2, ceil_mode=True)#平局池化,使用2*2的核图像变小一般,奇数进一
        # print('pool')
        # print(out.shape)
        return out


class DenseNet(nn.Module):
    def __init__(self, params):
        super(DenseNet, self).__init__()
        growthRate = params['densenet']['growthRate']#获取字典中densenet-growthRate的值(论文配置为24)
        reduction = params['densenet']['reduction']#论文为0.5
        bottleneck = params['densenet']['bottleneck']#使用瓶颈层
        use_dropout = params['densenet']['use_dropout']#使用正则化

        nDenseBlocks = 16
        nChannels = 2 * growthRate
        self.conv1 = nn.Conv2d(params['encoder']['input_channel'], nChannels, kernel_size=7, padding=3, stride=2, bias=False)#densenet的第一层*7*7的卷积核,输入通道数为1,输出通道数为48
        self.dense1 = self._make_dense(nChannels, growthRate, nDenseBlocks, bottleneck, use_dropout)#相当于使用了48个7*7的卷积核来学习特征,那么自然有48个通道
        nChannels += nDenseBlocks * growthRate#densenet输入通道等于块数乘每个块的输入通道数
        nOutChannels = int(math.floor(nChannels * reduction))#输出通道数每次是输入通道数的一半
        self.trans1 = Transition(nChannels, nOutChannels, use_dropout)
        nChannels = nOutChannels
        self.dense2 = self._make_dense(nChannels, growthRate, nDenseBlocks, bottleneck, use_dropout)#第二层densenet

        nChannels += nDenseBlocks * growthRate
        nOutChannels = int(math.floor(nChannels * reduction))
        self.trans2 = Transition(nChannels, nOutChannels, use_dropout)

        nChannels = nOutChannels
        self.dense3 = self._make_dense(nChannels, growthRate, nDenseBlocks, bottleneck, use_dropout)

    def _make_dense(self, nChannels, growthRate, nDenseBlocks, bottleneck, use_dropout):
        layers = []
        for i in range(int(nDenseBlocks)):#这是一个包含16个块的densenet网络
            if bottleneck:#使用的均是瓶颈层
                layers.append(Bottleneck(nChannels, growthRate, use_dropout))
            else:
                layers.append(SingleLayer(nChannels, growthRate, use_dropout))

            nChannels += growthRate#输出通道数增加
        return nn.Sequential(*layers)#将多个层按顺序连起来

    def forward(self, x):
        # print('原始')#torch.Size([4, 1, 128, 316])
        # print(x.shape)
        out = self.conv1(x)#经过一个7*7的卷积核,输出从1通道变为48通道,图片大小减半
        # print('经过7*7卷积')#torch.Size([4, 48, 64, 158])
        # print(out.shape)
        out = F.relu(out, inplace=True)
        # print('RELU')#torch.Size([4, 48, 64, 158])
        # print(out.shape)
        out = F.max_pool2d(out, 2, ceil_mode=True)#维数不变,尺寸减半
        # print('池化')#torch.Size([4, 48, 32, 79])
        # print(out.shape)
        out = self.dense1(out)#维数增加,尺寸减半
        # print('dense1层')#torch.Size([4, 432, 32, 79])
        # print(out.shape)
        out = self.trans1(out)#通道数、feature map均减半
        # print('tramns1层')#torch.Size([4, 216, 16, 40])
        # print(out.shape)
        out = self.dense2(out)
        # print('dense2层')#torch.Size([4, 600, 16, 40])
        # print(out.shape)
        out = self.trans2(out)
        # print('tramns2层')#torch.Size([4, 300, 8, 20])
        # print(out.shape)
        out = self.dense3(out)
        # print('dense3层')#torch.Size([4, 684, 8, 20])
        # print(out.shape)
        return out

这里我就在想将其换成resnet看看效果如何,这里我首先将其换成resnet50

这里更换非常值得一提,更换骨干网络,说简单非常简单,只要把resnet和densenet想象成两个黑盒子,只要输入和输出的维数一样就算更换成功,但是必须对两个网络有一个大致的了解才能进行更换

这里resnet个densenet的输入是相同的(因为就是直接输入图片,不需要进行其他处理),那就需要将输出的维度弄相同即可,这里我将我的代码densenet输入和输出维度打印在终端上,可以看出来,这里这四个维度分别是:[batch_size,通道数,图片宽度,图片高度]

在这里插入图片描述
由此可以看出,输入张量经过densenet网络后,通道数从1变为684,图片的宽高变缩小了16倍(向上取整)

那么接下来要做的事情就简单了,我只需将resnet对输入张量的操作也变为通道数从1变为684,图片的宽高变缩小了16倍即可

首先我运行resnet50的代码将张量通过resnet50的输入输出信息打印到终端上,看一下张量进入resnet后维度进行了怎样的变化
在这里插入图片描述

这里发现resnet-50将张量的通道数从1变为2048,图片的宽高变缩小了32倍

那么只需要将通道数从2048变为684,让图片的宽高缩小为16倍即可达到目的

这里我对resnet的代码做了如下修改

1.首先修改输出通道数

我在resnet输出的最后,加入了一个1*1的卷积层,将通道数从2048降到了684

        self.layer5 =  nn.Conv2d(2048, 684, kernel_size=1, bias=False)
        self.mybn1 = nn.BatchNorm2d(684)  # 标准化
2.修改特征图宽高

将第二个卷积层的步长从2变为1,这样feature map的宽高就少进行了一次缩小操作,即可将缩小32倍修改为缩小16倍

#self.layer2 = self._make_layer(block, 128, layers[1], stride=2)修改前
self.layer2 = self._make_layer(block, 128, layers[1], stride=1)#修改后

再次运行代码,即可变为将通道数从1变为684,图片的宽高变缩小了16倍的操作了,也就是成功将骨干网络将Densenet换成resnet-50,这里我的的手写公式的代码就可以正常训练了

在这里插入图片描述

最后给出完整版的修改后的resnet-50的代码

import torch.nn as nn

class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
                               padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride
        self.dropout = nn.Dropout(p=0.2)

    def forward(self, x):
        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        out = self.relu(out)
        out = self.dropout(out)  ##20%不更新
        return out

class ResNet50(nn.Module):

    def __init__(self, block, layers, num_classes=1000):
        super(ResNet50, self).__init__()
        self.inplanes = 64
        self.conv1 = nn.Conv2d(1, 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=1)
        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)
        self.layer5 =  nn.Conv2d(2048, 684, kernel_size=1, bias=False)
        self.mybn1 = nn.BatchNorm2d(684)  # 标准化

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )
        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes))
        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.layer5(x)
        x=self.mybn1(x)
        # print('###############')
        # print(x.shape)
        return x

总结:

更换骨干网络其实挺简单的,说白了就是将两个网络的输入维数和输出维数调成一样的即可,但是想要调成一样的,要求对两个骨干网络的代码和原理都比较熟悉才行,我这里更换骨干网络花了接近2整天才完成,前一天半主要是学习resnet和densnet的代码和原理,最后半天进行代码的修改。

风语者!平时喜欢研究各种技术,目前在从事后端开发工作,热爱生活、热爱工作。