resnet結(jié)構(gòu)是一種可以到達(dá)比較深的層數(shù)的網(wǎng)絡(luò),在機(jī)器學(xué)習(xí)中有著很多應(yīng)用。作為機(jī)器學(xué)習(xí)最常用的語言,python實(shí)現(xiàn)resnet結(jié)構(gòu)的介紹還是比較多的,今天小編就整理了一篇python實(shí)現(xiàn)resnet結(jié)構(gòu)的文章,希望給小伙伴們帶來幫助。
1.ResNet的創(chuàng)新
現(xiàn)在重新稍微系統(tǒng)的介紹一下ResNet網(wǎng)絡(luò)結(jié)構(gòu)。 ResNet結(jié)構(gòu)首先通過一個(gè)卷積層然后有一個(gè)池化層,然后通過一系列的殘差結(jié)構(gòu),最后再通過一個(gè)平均池化下采樣操作,以及一個(gè)全連接層的得到了一個(gè)輸出。ResNet網(wǎng)絡(luò)可以達(dá)到很深的層數(shù)的原因就是不斷的堆疊殘差結(jié)構(gòu)而來的。
1)亮點(diǎn)
網(wǎng)絡(luò)中的亮點(diǎn) :
- 超深的網(wǎng)絡(luò)結(jié)構(gòu)( 突破1000 層)
- 提出residual 模塊
- 使用Batch Normalization 加速訓(xùn)練( 丟棄dropout)
但是,一般來說,并不是一直的加深神經(jīng)網(wǎng)絡(luò)的結(jié)構(gòu)就會(huì)得到一個(gè)更好的結(jié)果,一般太深的網(wǎng)絡(luò)會(huì)出現(xiàn)過擬合的現(xiàn)象嚴(yán)重,可能還沒有一些淺層網(wǎng)絡(luò)要好。
2)原因
其中有兩個(gè)原因:
- 梯度消失或梯度爆炸
當(dāng)層數(shù)過多的時(shí)候,假設(shè)每一層的誤差梯度都是一個(gè)小于1的數(shù)值,當(dāng)進(jìn)行方向傳播的過程中,每向前傳播一層,都要乘以一個(gè)小于1的誤差梯度,當(dāng)網(wǎng)絡(luò)越來越深時(shí),所成的小于1的系數(shù)也就越來越多,此時(shí)梯度便越趨近于0,這樣梯度便會(huì)越來越小。這便會(huì)造成梯度消失的現(xiàn)象。
而當(dāng)所成的誤差梯度是一個(gè)大于1的系數(shù),而隨著網(wǎng)絡(luò)層數(shù)的加深,梯度便會(huì)越來越大,這便會(huì)造成梯度爆炸的現(xiàn)象。
- 退化問題(degradation problem)
當(dāng)解決了梯度消失或者梯度爆炸的問題之后,其實(shí)網(wǎng)絡(luò)的效果可能還是不盡如意,還可能有退化問題。為此,ResNet提出了殘差結(jié)構(gòu)來解決這個(gè)退化問題。 也正是因?yàn)橛羞@個(gè)殘差的結(jié)構(gòu),所以才可以搭建這么深的網(wǎng)絡(luò)。
2.ResNet的結(jié)構(gòu)
殘差結(jié)構(gòu)如圖所示
作圖是針對(duì)ResNet-18/34層淺層網(wǎng)絡(luò)的結(jié)構(gòu),右圖是ResNet-50/101/152層深層網(wǎng)絡(luò)的結(jié)構(gòu),其中注意:主分支與shortcut 的輸出特征矩陣shape。
一下表格為網(wǎng)絡(luò)的一些主要參數(shù)
可以看見,不同層數(shù)的網(wǎng)絡(luò)結(jié)構(gòu)其實(shí)框架是類似的,不同的至少堆疊的殘差結(jié)構(gòu)的數(shù)量。
1)淺層的殘差結(jié)構(gòu)
需要注意,有些殘差結(jié)構(gòu)的ShortCut是實(shí)線,而有的是虛線,這兩者是不同的。對(duì)于左圖來說,ShortCut是實(shí)線,這表明輸入與輸出的shape是一樣的,所以可以直接的進(jìn)行相加。而對(duì)于右圖來說,其輸入的shape與輸出的shape是不一樣的,這時(shí)候需要調(diào)整步長stribe與kernel size來使得兩條路(主分支與捷徑分支)所處理好的shape是一模一樣的。
2)深層的殘差結(jié)構(gòu)
同樣的,需要注意,主分支與shortcut 的輸出特征矩陣shape必須相同,同樣的通過步長來調(diào)整。
但是注意原論文中:
右側(cè)虛線殘差結(jié)構(gòu)的主分支上、第一個(gè)1x1卷積層的步距是2,第二個(gè)3x3卷積層的步距是1.
而在pytorch官方實(shí)現(xiàn)的過程中是第一個(gè)1x1卷積層的步距是1,第二個(gè)3x3卷積層步距是2,這樣能夠在ImageNet的top1上提升大概0.5%的準(zhǔn)確率。
所以在conv3_x,conv4_x,conv5_x中所對(duì)應(yīng)的殘差結(jié)構(gòu)的第一層,都是指虛線的殘差結(jié)構(gòu),其他的殘差結(jié)構(gòu)是實(shí)線的殘差結(jié)構(gòu)。
3)總結(jié)
對(duì)于每個(gè)大模塊中的第一個(gè)殘差結(jié)構(gòu),需要通過虛線分支來調(diào)整殘差結(jié)構(gòu)的輸入與輸出是同一個(gè)shape。此時(shí)使用了下采樣的操作函數(shù)。
對(duì)于每個(gè)大模塊中的其他剩余的殘差結(jié)構(gòu),只需要通過實(shí)線分支來調(diào)整殘差網(wǎng)絡(luò)結(jié)構(gòu),因?yàn)槠漭敵龊洼斎氡旧砭褪峭粋€(gè)shape的。
對(duì)于第一個(gè)大模塊的第一個(gè)殘差結(jié)構(gòu),其第二個(gè)3x3的卷積中,步長是1的,而其他的三個(gè)大模塊的步長均為2.
在每一個(gè)大模塊的維度變換中,主要是第一個(gè)殘差結(jié)構(gòu)使得shape減半,而模塊中其他的殘差結(jié)構(gòu)都是沒有改變shape的。也真因?yàn)闆]有改變shape,所以這些殘差結(jié)構(gòu)才可以直接的通過實(shí)線進(jìn)行相加。
3.Batch Normalization
Batch Normalization的目的是使我們的一批(Batch)特征矩陣feature map滿足均值為0,方差為1的分布規(guī)律。
其中:
μ,σ_2在正向傳播過程中統(tǒng)計(jì)得到
γ,β在反向傳播過程中訓(xùn)練得到
Batch Normalization是google團(tuán)隊(duì)在2015年論文《Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift》提出的。通過該方法能夠加速網(wǎng)絡(luò)的收斂并提升準(zhǔn)確率。
具體的相關(guān)原理見:Batch Normalization詳解以及pytorch實(shí)驗(yàn)
4.參考代碼
import torch import torch.nn as nn # 分類數(shù)目 num_class = 5 # 各層數(shù)目 resnet18_params = [2, 2, 2, 2] resnet34_params = [3, 4, 6, 3] resnet50_params = [3, 4, 6, 3] resnet101_params = [3, 4, 23, 3] resnet152_params = [3, 8, 36, 3] # 定義Conv1層 def Conv1(in_planes, places, stride=2): return nn.Sequential( nn.Conv2d(in_channels=in_planes,out_channels=places,kernel_size=7,stride=stride,padding=3, bias=False), nn.BatchNorm2d(places), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=3, stride=2, padding=1) ) # 淺層的殘差結(jié)構(gòu) class BasicBlock(nn.Module): def __init__(self,in_places,places, stride=1,downsampling=False, expansion = 1): super(BasicBlock,self).__init__() self.expansion = expansion self.downsampling = downsampling # torch.Size([1, 64, 56, 56]), stride = 1 # torch.Size([1, 128, 28, 28]), stride = 2 # torch.Size([1, 256, 14, 14]), stride = 2 # torch.Size([1, 512, 7, 7]), stride = 2 self.basicblock = nn.Sequential( nn.Conv2d(in_channels=in_places, out_channels=places, kernel_size=3, stride=stride, padding=1, bias=False), nn.BatchNorm2d(places), nn.ReLU(inplace=True), nn.Conv2d(in_channels=places, out_channels=places, kernel_size=3, stride=1, padding=1, bias=False), nn.BatchNorm2d(places * self.expansion), ) # torch.Size([1, 64, 56, 56]) # torch.Size([1, 128, 28, 28]) # torch.Size([1, 256, 14, 14]) # torch.Size([1, 512, 7, 7]) # 每個(gè)大模塊的第一個(gè)殘差結(jié)構(gòu)需要改變步長 if self.downsampling: self.downsample = nn.Sequential( nn.Conv2d(in_channels=in_places, out_channels=places*self.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(places*self.expansion) ) self.relu = nn.ReLU(inplace=True) def forward(self, x): # 實(shí)線分支 residual = x out = self.basicblock(x) # 虛線分支 if self.downsampling: residual = self.downsample(x) out += residual out = self.relu(out) return out # 深層的殘差結(jié)構(gòu) class Bottleneck(nn.Module): # 注意:默認(rèn) downsampling=False def __init__(self,in_places,places, stride=1,downsampling=False, expansion = 4): super(Bottleneck,self).__init__() self.expansion = expansion self.downsampling = downsampling self.bottleneck = nn.Sequential( # torch.Size([1, 64, 56, 56]),stride=1 # torch.Size([1, 128, 56, 56]),stride=1 # torch.Size([1, 256, 28, 28]), stride=1 # torch.Size([1, 512, 14, 14]), stride=1 nn.Conv2d(in_channels=in_places,out_channels=places,kernel_size=1,stride=1, bias=False), nn.BatchNorm2d(places), nn.ReLU(inplace=True), # torch.Size([1, 64, 56, 56]),stride=1 # torch.Size([1, 128, 28, 28]), stride=2 # torch.Size([1, 256, 14, 14]), stride=2 # torch.Size([1, 512, 7, 7]), stride=2 nn.Conv2d(in_channels=places, out_channels=places, kernel_size=3, stride=stride, padding=1, bias=False), nn.BatchNorm2d(places), nn.ReLU(inplace=True), # torch.Size([1, 256, 56, 56]),stride=1 # torch.Size([1, 512, 28, 28]), stride=1 # torch.Size([1, 1024, 14, 14]), stride=1 # torch.Size([1, 2048, 7, 7]), stride=1 nn.Conv2d(in_channels=places, out_channels=places * self.expansion, kernel_size=1, stride=1, bias=False), nn.BatchNorm2d(places * self.expansion), ) # torch.Size([1, 256, 56, 56]) # torch.Size([1, 512, 28, 28]) # torch.Size([1, 1024, 14, 14]) # torch.Size([1, 2048, 7, 7]) if self.downsampling: self.downsample = nn.Sequential( nn.Conv2d(in_channels=in_places, out_channels=places*self.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(places*self.expansion) ) self.relu = nn.ReLU(inplace=True) def forward(self, x): # 實(shí)線分支 residual = x out = self.bottleneck(x) # 虛線分支 if self.downsampling: residual = self.downsample(x) out += residual out = self.relu(out) return out class ResNet(nn.Module): def __init__(self,blocks, blockkinds, num_classes=num_class): super(ResNet,self).__init__() self.blockkinds = blockkinds self.conv1 = Conv1(in_planes = 3, places= 64) # 對(duì)應(yīng)淺層網(wǎng)絡(luò)結(jié)構(gòu) if self.blockkinds == BasicBlock: self.expansion = 1 # 64 -> 64 self.layer1 = self.make_layer(in_places=64, places=64, block=blocks[0], stride=1) # 64 -> 128 self.layer2 = self.make_layer(in_places=64, places=128, block=blocks[1], stride=2) # 128 -> 256 self.layer3 = self.make_layer(in_places=128, places=256, block=blocks[2], stride=2) # 256 -> 512 self.layer4 = self.make_layer(in_places=256, places=512, block=blocks[3], stride=2) self.fc = nn.Linear(512, num_classes) # 對(duì)應(yīng)深層網(wǎng)絡(luò)結(jié)構(gòu) if self.blockkinds == Bottleneck: self.expansion = 4 # 64 -> 64 self.layer1 = self.make_layer(in_places = 64, places= 64, block=blocks[0], stride=1) # 256 -> 128 self.layer2 = self.make_layer(in_places = 256,places=128, block=blocks[1], stride=2) # 512 -> 256 self.layer3 = self.make_layer(in_places=512,places=256, block=blocks[2], stride=2) # 1024 -> 512 self.layer4 = self.make_layer(in_places=1024,places=512, block=blocks[3], stride=2) self.fc = nn.Linear(2048, num_classes) self.avgpool = nn.AvgPool2d(7, stride=1) # 初始化網(wǎng)絡(luò)結(jié)構(gòu) for m in self.modules(): if isinstance(m, nn.Conv2d): # 采用了何凱明的初始化方法 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, in_places, places, block, stride): layers = [] # torch.Size([1, 64, 56, 56]) -> torch.Size([1, 256, 56, 56]), stride=1 故w,h不變 # torch.Size([1, 256, 56, 56]) -> torch.Size([1, 512, 28, 28]), stride=2 故w,h變 # torch.Size([1, 512, 28, 28]) -> torch.Size([1, 1024, 14, 14]),stride=2 故w,h變 # torch.Size([1, 1024, 14, 14]) -> torch.Size([1, 2048, 7, 7]), stride=2 故w,h變 # 此步需要通過虛線分支,downsampling=True layers.append(self.blockkinds(in_places, places, stride, downsampling =True)) # torch.Size([1, 256, 56, 56]) -> torch.Size([1, 256, 56, 56]) # torch.Size([1, 512, 28, 28]) -> torch.Size([1, 512, 28, 28]) # torch.Size([1, 1024, 14, 14]) -> torch.Size([1, 1024, 14, 14]) # torch.Size([1, 2048, 7, 7]) -> torch.Size([1, 2048, 7, 7]) # print("places*self.expansion:", places*self.expansion) # print("block:", block) # 此步需要通過實(shí)線分支,downsampling=False, 每個(gè)大模塊的第一個(gè)殘差結(jié)構(gòu)需要改變步長 for i in range(1, block): layers.append(self.blockkinds(places*self.expansion, places)) return nn.Sequential(*layers) def forward(self, x): # conv1層 x = self.conv1(x) # torch.Size([1, 64, 56, 56]) # conv2_x層 x = self.layer1(x) # torch.Size([1, 256, 56, 56]) # conv3_x層 x = self.layer2(x) # torch.Size([1, 512, 28, 28]) # conv4_x層 x = self.layer3(x) # torch.Size([1, 1024, 14, 14]) # conv5_x層 x = self.layer4(x) # torch.Size([1, 2048, 7, 7]) x = self.avgpool(x) # torch.Size([1, 2048, 1, 1]) / torch.Size([1, 512]) x = x.view(x.size(0), -1) # torch.Size([1, 2048]) / torch.Size([1, 512]) x = self.fc(x) # torch.Size([1, 5]) return x def ResNet18(): return ResNet(resnet18_params, BasicBlock) def ResNet34(): return ResNet(resnet34_params, BasicBlock) def ResNet50(): return ResNet(resnet50_params, Bottleneck) def ResNet101(): return ResNet(resnet101_params, Bottleneck) def ResNet152(): return ResNet(resnet152_params, Bottleneck) if __name__=='__main__': # model = torchvision.models.resnet50() # 模型測試 # model = ResNet18() # model = ResNet34() # model = ResNet50() # model = ResNet101() model = ResNet152() # print(model) input = torch.randn(1, 3, 224, 224) out = model(input) print(out.shape)
以上就是pytorch實(shí)現(xiàn)ResNet結(jié)構(gòu)的詳細(xì)內(nèi)容,更多機(jī)器學(xué)習(xí)的資料請(qǐng)關(guān)注W3Cschool其它相關(guān)文章!