经典卷积神经网络-ResNet
一、背景介绍
残差神经网络(ResNet)是由微软研究院的何恺明、张祥雨、任少卿、孙剑等人提出的。ResNet 在2015 年的ILSVRC（ImageNet Large Scale Visual Recognition Challenge）中取得了冠军。残差神经网络的主要贡献是发现了“退化现象（Degradation）”，并针对退化现象发明了 “快捷连接（Shortcut connection）”，极大的消除了深度过大的神经网络训练困难问题。神经网络的“深度”首次突破了100层、最大的神经网络甚至超过了1000层。
二、ResNet网络结构
2.0 残差块
配合吴恩达深度学习视频中的图片进行讲解：
如图所示，Residual block就是将 a [ l ] a^{[l]} a[l]传送到 z l + 2 z^{l+2} zl+2上，其相加之后再进行激活得到 a [ l + 2 ] a^{[l+2]} a[l+2]。这一步骤称为"skip connection"，即指 a [ l ] a^{[l]} a[l]跳过一层或好几层，从而将信息传递到神经网络的更深层。所以构建一个ResNet网络就是通过将很多这样的残差块堆积在一起，形成一个深度神经网络。
那么引入残差块为什么有效呢？ 一个直观上的理解：
如图所示，假设我们给我们的神经网络再增加两层，我们要得到 a [ l + 2 ] a^{[l+2]} a[l+2]，我们通过增加一个残差块来完成。这时 a [ l + 2 ] = g ( z [ l + 2 ] + a [ l ] ) = g ( w [ l + 2 ] a [ l + 1 ] + a [ l ] ) a^{[l+2]}=g(z^{[l+2]}+a^{[l]})=g(w^{[l+2]}a^{[l+1]}+a^{}[l]) a[l+2]=g(z[l+2]+a[l])=g(w[l+2]a[l+1]+a[l])，如果我们应用了L2正则化，此时权重参数会减小，我们可以极端的假设 w [ l + 2 ] = 0 ， b [ l + 2 ] = 0 w^{[l+2]}=0，b^{[l+2]}=0 w[l+2]=0，b[l+2]=0，那么得到 a [ l + 2 ] = g ( a [ l ] ) = a [ l ] a^{[l+2]}=g(a^{[l]})=a^{[l]} a[l+2]=g(a[l])=a[l]（因为使用的是ReLU激活函数，非负的值激活后为原来的值， a [ l ] a^{[l]} a[l]已经经过ReLU激活过了，所以全为非负值）。这意味着，即使给神经网络增加了两层，它的效果并不逊色于更简单的神经网络。所以给大型的神经网络添加残差块来增加网络深度，并不会影响网络的表现。如果我们增加的这两层碰巧能学习到一些有用的信息，那么它就比原来的神经网络表现的更好。
论文中ResNet层数在34及以下和50及以上时采用的是不同的残差块。下面我们分别介绍：
2.1 ResNet-34
如上图所示是ResNet-34以下采用的残差块，我们将其称作BasicBlock。
ResNet-34的网络结构如下：图中实线表示通道数没有变化，虚线表示通道数发生了变化。
其具体的网络结构如下表：
2.2 ResNet-50
如图所示是ResNet-50以上采用的残差块，我们将其称作Bottleneck，使用了1 × 1的卷积来进行通道数的改变，减小计算量。其具体的网络结构见上面的表。
三、论文部分解读
-
论文中的第一张图就表明了更深层的“plain”神经网络（即不用Residual Learning）的错误率在训练集和测试集上甚至比层数少的“plain”神经网络还要高，所以就引出了问题：训练很深的网络是一个很难的问题。
-
论文中的这张图就表明了，使用Residual Learning后训练更深层的神经网络效果会变得更好。
-
论文中处理残差连接中输入和输出不对等的解决方法：
- 添加一些额外的0，使得输入和输出的形状可以对应起来可以做相加。
- 输入和输出不对等的时候使用1 × 1的卷积操作来做投影（目前都是这种方法）
- 所有的地方都使用1 × 1的卷积来做投影，计算量太大没必要
-
论文最主要的是提出了residual结构（残差结构），并搭建超深的网络结构(突破1000层)
-
使用了大量BN来加速训练(丢弃dropout)
-
论文中对CIFAR10数据集做了大量实验验证其效果，最后还将ResNet用到了目标检测领域
三、ResNet-18的Pytorch实现
import torch
import torch.nn as nn
import torchsummary# BasicBlock
class BasicBlock(nn.Module):def __init__(self, in_channels, out_channels, stride=1):super(BasicBlock, self).__init__()# 第一个卷积有可能要进行下采样 即将输出通道翻倍 输出数据大小全部减半# 所以我们让第一个卷积的stride设置为可传入的参数# 如果要进行BN 卷积就不需要加偏置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)# 第二个卷积保持尺寸和通道数self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)self.bn2 = nn.BatchNorm2d(out_channels)# 如果最后输入和输出的通道数不一样 那么就用1 × 1的卷积调节输入 方便后面能和输出相加# shortcut操作也是downsample(下采样) 即将输出通道翻倍 输出数据大小全部减半# 这一步主要做的就是为了能让最后的输出数据和输入数据"连接"上 即相加self.shortcut = nn.Sequential()if stride != 1 or in_channels != out_channels:self.shortcut = nn.Sequential(nn.Conv2d(in_channels, out_channels, stride=stride, kernel_size=1, bias=False),nn.BatchNorm2d(out_channels))def forward(self, x):# 记录identityidentity = xout = self.conv1(x)out = self.bn1(out)out = self.relu(out)out = self.conv2(out)out = self.bn2(out)# 进行残差连接out += self.shortcut(identity)# 残差连接之后激活out = self.relu(out)return out# ResNet-18
class ResNet18(nn.Module):def __init__(self, num_classes=1000):super(ResNet18, self).__init__()# output_size = [(input_size - kernel_size + 2padding) / stride] + 1# 112 = [(224 - 7 + 2padding) / stride] + 1 -> padding = 3self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)self.bn1 = nn.BatchNorm2d(64)self.relu = nn.ReLU(inplace=True)# 56 = [(112 - 3 + 2padding) / 2] + 1 -> padding = 1self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)self.layer1 = self._make_layer(in_channels=64, out_channels=64, blocks=2, stride=1)self.layer2 = self._make_layer(in_channels=64, out_channels=128, blocks=2, stride=2)self.layer3 = self._make_layer(in_channels=128, out_channels=256, blocks=2, stride=2)self.layer4 = self._make_layer(in_channels=256, out_channels=512, blocks=2, stride=2)# 平均池化 参数就是out_sizeself.avgpool = nn.AdaptiveAvgPool2d((1, 1))self.fc = nn.Linear(512, num_classes)def _make_layer(self, in_channels, out_channels, blocks, stride=1):layer = []# 因为第一个残差块的输入通道可能与输出通道不同 所以单独拿出来赋值layer.append(BasicBlock(in_channels, out_channels, stride))for _ in range(1, blocks):layer.append(BasicBlock(out_channels, out_channels))return nn.Sequential(*layer)def forward(self, x):out = self.conv1(x)out = self.bn1(out)out = self.relu(out)out = self.maxpool(out)out = self.layer1(out)out = self.layer2(out)out = self.layer3(out)out = self.layer4(out)out = self.avgpool(out)out = torch.flatten(out, 1)out = self.fc(out)return outif __name__ == '__main__':DEVICE = "cuda" if torch.cuda.is_available() else "cpu"model = (ResNet18())model.to(DEVICE)# print(model)print(torch.cuda.is_available())torchsummary.summary(model, (3, 224, 224), 64)主要注意点是：
- BasicBlock（34层以下的残差块）中残差连接的方法，注意输入和输出的通道数，用1×1的卷积做好尺寸和维度对齐。
- 使用_make_layer()函数批量生成残差块。
在控制台输出网络结构：
ResNet18((conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)(layer1): Sequential((0): BasicBlock((conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential())(1): BasicBlock((conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential()))(layer2): Sequential((0): BasicBlock((conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential((0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)))(1): BasicBlock((conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential()))(layer3): Sequential((0): BasicBlock((conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential((0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)))(1): BasicBlock((conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential()))(layer4): Sequential((0): BasicBlock((conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential((0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)))(1): BasicBlock((conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(shortcut): Sequential()))(avgpool): AdaptiveAvgPool2d(output_size=(1, 1))(fc): Linear(in_features=512, out_features=1000, bias=True)
)
使用torchsummary来测试网络：
----------------------------------------------------------------Layer (type) Output Shape Param #
================================================================Conv2d-1 [64, 64, 112, 112] 9,408BatchNorm2d-2 [64, 64, 112, 112] 128ReLU-3 [64, 64, 112, 112] 0MaxPool2d-4 [64, 64, 56, 56] 0Conv2d-5 [64, 64, 56, 56] 36,864BatchNorm2d-6 [64, 64, 56, 56] 128ReLU-7 [64, 64, 56, 56] 0Conv2d-8 [64, 64, 56, 56] 36,864BatchNorm2d-9 [64, 64, 56, 56] 128ReLU-10 [64, 64, 56, 56] 0BasicBlock-11 [64, 64, 56, 56] 0Conv2d-12 [64, 64, 56, 56] 36,864BatchNorm2d-13 [64, 64, 56, 56] 128ReLU-14 [64, 64, 56, 56] 0Conv2d-15 [64, 64, 56, 56] 36,864BatchNorm2d-16 [64, 64, 56, 56] 128ReLU-17 [64, 64, 56, 56] 0BasicBlock-18 [64, 64, 56, 56] 0Conv2d-19 [64, 128, 28, 28] 73,728BatchNorm2d-20 [64, 128, 28, 28] 256ReLU-21 [64, 128, 28, 28] 0Conv2d-22 [64, 128, 28, 28] 147,456BatchNorm2d-23 [64, 128, 28, 28] 256Conv2d-24 [64, 128, 28, 28] 8,192BatchNorm2d-25 [64, 128, 28, 28] 256ReLU-26 [64, 128, 28, 28] 0BasicBlock-27 [64, 128, 28, 28] 0Conv2d-28 [64, 128, 28, 28] 147,456BatchNorm2d-29 [64, 128, 28, 28] 256ReLU-30 [64, 128, 28, 28] 0Conv2d-31 [64, 128, 28, 28] 147,456BatchNorm2d-32 [64, 128, 28, 28] 256ReLU-33 [64, 128, 28, 28] 0BasicBlock-34 [64, 128, 28, 28] 0Conv2d-35 [64, 256, 14, 14] 294,912BatchNorm2d-36 [64, 256, 14, 14] 512ReLU-37 [64, 256, 14, 14] 0Conv2d-38 [64, 256, 14, 14] 589,824BatchNorm2d-39 [64, 256, 14, 14] 512Conv2d-40 [64, 256, 14, 14] 32,768BatchNorm2d-41 [64, 256, 14, 14] 512ReLU-42 [64, 256, 14, 14] 0BasicBlock-43 [64, 256, 14, 14] 0Conv2d-44 [64, 256, 14, 14] 589,824BatchNorm2d-45 [64, 256, 14, 14] 512ReLU-46 [64, 256, 14, 14] 0Conv2d-47 [64, 256, 14, 14] 589,824BatchNorm2d-48 [64, 256, 14, 14] 512ReLU-49 [64, 256, 14, 14] 0BasicBlock-50 [64, 256, 14, 14] 0Conv2d-51 [64, 512, 7, 7] 1,179,648BatchNorm2d-52 [64, 512, 7, 7] 1,024ReLU-53 [64, 512, 7, 7] 0Conv2d-54 [64, 512, 7, 7] 2,359,296BatchNorm2d-55 [64, 512, 7, 7] 1,024Conv2d-56 [64, 512, 7, 7] 131,072BatchNorm2d-57 [64, 512, 7, 7] 1,024ReLU-58 [64, 512, 7, 7] 0BasicBlock-59 [64, 512, 7, 7] 0Conv2d-60 [64, 512, 7, 7] 2,359,296BatchNorm2d-61 [64, 512, 7, 7] 1,024ReLU-62 [64, 512, 7, 7] 0Conv2d-63 [64, 512, 7, 7] 2,359,296BatchNorm2d-64 [64, 512, 7, 7] 1,024ReLU-65 [64, 512, 7, 7] 0BasicBlock-66 [64, 512, 7, 7] 0
AdaptiveAvgPool2d-67 [64, 512, 1, 1] 0Linear-68 [64, 1000] 513,000
================================================================
Total params: 11,689,512
Trainable params: 11,689,512
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 36.75
Forward/backward pass size (MB): 4018.74
Params size (MB): 44.59
Estimated Total Size (MB): 4100.08
----------------------------------------------------------------
四、ResNet-50的Pytorch实现
import torch
import torch.nn as nn
import torchsummary# Bottleneck
class Bottleneck(nn.Module):# expansion = 4，因为Bottleneck中每个残差结构输出维度都是输入维度的4倍expansion = 4def __init__(self, in_channels, out_channels, stride=1):super(Bottleneck, self).__init__()# 注意这里1×1的卷积不用paddingself.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, bias=False)self.bn1 = nn.BatchNorm2d(out_channels)# 维持特征图尺寸 padding = 1self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)self.bn2 = nn.BatchNorm2d(out_channels)# 注意这里1×1的卷积不用paddingself.conv3 = nn.Conv2d(out_channels, out_channels * self.expansion, kernel_size=1, stride=1, bias=False)self.bn3 = nn.BatchNorm2d(out_channels * self.expansion)self.relu = nn.ReLU(inplace=True)self.shortcut = nn.Sequential()if stride != 1 or in_channels != out_channels * self.expansion:self.shortcut = nn.Sequential(nn.Conv2d(in_channels, out_channels * self.expansion, stride=stride, kernel_size=1, bias=False),nn.BatchNorm2d(out_channels * self.expansion))def forward(self, x):identity = xout = 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)# 残差连接out += self.shortcut(identity)out = self.relu(out)return out# ResNet50
class ResNet50(nn.Module):def __init__(self, num_classes=1000):super(ResNet50, self).__init__()# output_size = [(input_size - kernel_size + 2padding) / stride] + 1# 112 = [(224 - 7 + 2padding) / stride] + 1 -> padding = 3self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)self.bn1 = nn.BatchNorm2d(64)self.relu = nn.ReLU(inplace=True)# 56 = [(112 - 3 + 2padding) / 2] + 1 -> padding = 1self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)self.layer1 = self._make_layer(64, 64, blocks=3, stride=1)self.layer2 = self._make_layer(256, 128, blocks=4, stride=2)self.layer3 = self._make_layer(512, 256, blocks=6, stride=2)self.layer4 = self._make_layer(1024, 512, blocks=3, stride=2)# 平均池化 参数就是out_sizeself.avgpool = nn.AdaptiveAvgPool2d((1, 1))self.fc = nn.Linear(2048, num_classes)def _make_layer(self, in_channels, out_channels, blocks, stride=1):layer = []# 因为第一个残差块的输入通道可能与输出通道不同 所以单独拿出来赋值layer.append(Bottleneck(in_channels, out_channels, stride))for _ in range(1, blocks):# 接下来的每一个其输入通道都是输出的四倍layer.append(Bottleneck(out_channels * 4, out_channels))return nn.Sequential(*layer)def forward(self, x):out = self.conv1(x)out = self.bn1(out)out = self.relu(out)out = self.maxpool(out)out = self.layer1(out)out = self.layer2(out)out = self.layer3(out)out = self.layer4(out)out = self.avgpool(out)out = torch.flatten(out, 1)out = self.fc(out)return outif __name__ == '__main__':DEVICE = "cuda" if torch.cuda.is_available() else "cpu"model = ResNet50()model.to(DEVICE)# print(model)torchsummary.summary(model, (3, 224, 224), 64)
主要注意点是：
-
Bottleneck（50层以上残差块）中块与块直接连接的通道数问题。在BasicBlock中由于两个相同的块在连接的时候通道数是相同的，而Bottleneck中两个相同的块在连接的时候其通道数相差四倍。即下面这段代码：
for _ in range(1, blocks):# 接下来的每一个其输入通道都是输出的四倍layer.append(Bottleneck(out_channels * 4, out_channels)) -
注意最后全连接的通道数，ResNet50以上的最后全连接的in_features为2048。
-
注意在残差连接的时候，shortcut调整的维度一定要和输出维度对上，即下面这段代码：
self.shortcut = nn.Sequential() if stride != 1 or in_channels != out_channels * self.expansion:self.shortcut = nn.Sequential(nn.Conv2d(in_channels, out_channels * self.expansion, stride=stride, kernel_size=1, bias=False),nn.BatchNorm2d(out_channels * self.expansion)) -
写代码的时候，要注意只有每一堆残差块结构中的第一个残差块的第一个卷积操作的stride为2，其余都为1，（除了第一堆残差块要维持尺度不变），即：
self.layer1 = self._make_layer(64, 64, blocks=3, stride=1)
在控制台输出网络结构：
ResNet50((conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)(layer1): Sequential((0): Bottleneck((conv1): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential((0): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)))(1): Bottleneck((conv1): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(2): Bottleneck((conv1): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential()))(layer2): Sequential((0): Bottleneck((conv1): Conv2d(256, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential((0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)))(1): Bottleneck((conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(2): Bottleneck((conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(3): Bottleneck((conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential()))(layer3): Sequential((0): Bottleneck((conv1): Conv2d(512, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential((0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)))(1): Bottleneck((conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(2): Bottleneck((conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(3): Bottleneck((conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(4): Bottleneck((conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(5): Bottleneck((conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential()))(layer4): Sequential((0): Bottleneck((conv1): Conv2d(1024, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential((0): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False)(1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)))(1): Bottleneck((conv1): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential())(2): Bottleneck((conv1): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(relu): ReLU(inplace=True)(shortcut): Sequential()))(avgpool): AdaptiveAvgPool2d(output_size=(1, 1))(fc): Linear(in_features=2048, out_features=1000, bias=True)
)
使用torchsummary来测试网络：
----------------------------------------------------------------Layer (type) Output Shape Param #
================================================================Conv2d-1 [64, 64, 112, 112] 9,408BatchNorm2d-2 [64, 64, 112, 112] 128ReLU-3 [64, 64, 112, 112] 0MaxPool2d-4 [64, 64, 56, 56] 0Conv2d-5 [64, 64, 56, 56] 4,096BatchNorm2d-6 [64, 64, 56, 56] 128ReLU-7 [64, 64, 56, 56] 0Conv2d-8 [64, 64, 56, 56] 36,864BatchNorm2d-9 [64, 64, 56, 56] 128ReLU-10 [64, 64, 56, 56] 0Conv2d-11 [64, 256, 56, 56] 16,384BatchNorm2d-12 [64, 256, 56, 56] 512Conv2d-13 [64, 256, 56, 56] 16,384BatchNorm2d-14 [64, 256, 56, 56] 512ReLU-15 [64, 256, 56, 56] 0Bottleneck-16 [64, 256, 56, 56] 0Conv2d-17 [64, 64, 56, 56] 16,384BatchNorm2d-18 [64, 64, 56, 56] 128ReLU-19 [64, 64, 56, 56] 0Conv2d-20 [64, 64, 56, 56] 36,864BatchNorm2d-21 [64, 64, 56, 56] 128ReLU-22 [64, 64, 56, 56] 0Conv2d-23 [64, 256, 56, 56] 16,384BatchNorm2d-24 [64, 256, 56, 56] 512ReLU-25 [64, 256, 56, 56] 0Bottleneck-26 [64, 256, 56, 56] 0Conv2d-27 [64, 64, 56, 56] 16,384BatchNorm2d-28 [64, 64, 56, 56] 128ReLU-29 [64, 64, 56, 56] 0Conv2d-30 [64, 64, 56, 56] 36,864BatchNorm2d-31 [64, 64, 56, 56] 128ReLU-32 [64, 64, 56, 56] 0Conv2d-33 [64, 256, 56, 56] 16,384BatchNorm2d-34 [64, 256, 56, 56] 512ReLU-35 [64, 256, 56, 56] 0Bottleneck-36 [64, 256, 56, 56] 0Conv2d-37 [64, 128, 28, 28] 32,768BatchNorm2d-38 [64, 128, 28, 28] 256ReLU-39 [64, 128, 28, 28] 0Conv2d-40 [64, 128, 28, 28] 147,456BatchNorm2d-41 [64, 128, 28, 28] 256ReLU-42 [64, 128, 28, 28] 0Conv2d-43 [64, 512, 28, 28] 65,536BatchNorm2d-44 [64, 512, 28, 28] 1,024Conv2d-45 [64, 512, 28, 28] 131,072BatchNorm2d-46 [64, 512, 28, 28] 1,024ReLU-47 [64, 512, 28, 28] 0Bottleneck-48 [64, 512, 28, 28] 0Conv2d-49 [64, 128, 28, 28] 65,536BatchNorm2d-50 [64, 128, 28, 28] 256ReLU-51 [64, 128, 28, 28] 0Conv2d-52 [64, 128, 28, 28] 147,456BatchNorm2d-53 [64, 128, 28, 28] 256ReLU-54 [64, 128, 28, 28] 0Conv2d-55 [64, 512, 28, 28] 65,536BatchNorm2d-56 [64, 512, 28, 28] 1,024ReLU-57 [64, 512, 28, 28] 0Bottleneck-58 [64, 512, 28, 28] 0Conv2d-59 [64, 128, 28, 28] 65,536BatchNorm2d-60 [64, 128, 28, 28] 256ReLU-61 [64, 128, 28, 28] 0Conv2d-62 [64, 128, 28, 28] 147,456BatchNorm2d-63 [64, 128, 28, 28] 256ReLU-64 [64, 128, 28, 28] 0Conv2d-65 [64, 512, 28, 28] 65,536BatchNorm2d-66 [64, 512, 28, 28] 1,024ReLU-67 [64, 512, 28, 28] 0Bottleneck-68 [64, 512, 28, 28] 0Conv2d-69 [64, 128, 28, 28] 65,536BatchNorm2d-70 [64, 128, 28, 28] 256ReLU-71 [64, 128, 28, 28] 0Conv2d-72 [64, 128, 28, 28] 147,456BatchNorm2d-73 [64, 128, 28, 28] 256ReLU-74 [64, 128, 28, 28] 0Conv2d-75 [64, 512, 28, 28] 65,536BatchNorm2d-76 [64, 512, 28, 28] 1,024ReLU-77 [64, 512, 28, 28] 0Bottleneck-78 [64, 512, 28, 28] 0Conv2d-79 [64, 256, 14, 14] 131,072BatchNorm2d-80 [64, 256, 14, 14] 512ReLU-81 [64, 256, 14, 14] 0Conv2d-82 [64, 256, 14, 14] 589,824BatchNorm2d-83 [64, 256, 14, 14] 512ReLU-84 [64, 256, 14, 14] 0Conv2d-85 [64, 1024, 14, 14] 262,144BatchNorm2d-86 [64, 1024, 14, 14] 2,048Conv2d-87 [64, 1024, 14, 14] 524,288BatchNorm2d-88 [64, 1024, 14, 14] 2,048ReLU-89 [64, 1024, 14, 14] 0Bottleneck-90 [64, 1024, 14, 14] 0Conv2d-91 [64, 256, 14, 14] 262,144BatchNorm2d-92 [64, 256, 14, 14] 512ReLU-93 [64, 256, 14, 14] 0Conv2d-94 [64, 256, 14, 14] 589,824BatchNorm2d-95 [64, 256, 14, 14] 512ReLU-96 [64, 256, 14, 14] 0Conv2d-97 [64, 1024, 14, 14] 262,144BatchNorm2d-98 [64, 1024, 14, 14] 2,048ReLU-99 [64, 1024, 14, 14] 0Bottleneck-100 [64, 1024, 14, 14] 0Conv2d-101 [64, 256, 14, 14] 262,144BatchNorm2d-102 [64, 256, 14, 14] 512ReLU-103 [64, 256, 14, 14] 0Conv2d-104 [64, 256, 14, 14] 589,824BatchNorm2d-105 [64, 256, 14, 14] 512ReLU-106 [64, 256, 14, 14] 0Conv2d-107 [64, 1024, 14, 14] 262,144BatchNorm2d-108 [64, 1024, 14, 14] 2,048ReLU-109 [64, 1024, 14, 14] 0Bottleneck-110 [64, 1024, 14, 14] 0Conv2d-111 [64, 256, 14, 14] 262,144BatchNorm2d-112 [64, 256, 14, 14] 512ReLU-113 [64, 256, 14, 14] 0Conv2d-114 [64, 256, 14, 14] 589,824BatchNorm2d-115 [64, 256, 14, 14] 512ReLU-116 [64, 256, 14, 14] 0Conv2d-117 [64, 1024, 14, 14] 262,144BatchNorm2d-118 [64, 1024, 14, 14] 2,048ReLU-119 [64, 1024, 14, 14] 0Bottleneck-120 [64, 1024, 14, 14] 0Conv2d-121 [64, 256, 14, 14] 262,144BatchNorm2d-122 [64, 256, 14, 14] 512ReLU-123 [64, 256, 14, 14] 0Conv2d-124 [64, 256, 14, 14] 589,824BatchNorm2d-125 [64, 256, 14, 14] 512ReLU-126 [64, 256, 14, 14] 0Conv2d-127 [64, 1024, 14, 14] 262,144BatchNorm2d-128 [64, 1024, 14, 14] 2,048ReLU-129 [64, 1024, 14, 14] 0Bottleneck-130 [64, 1024, 14, 14] 0Conv2d-131 [64, 256, 14, 14] 262,144BatchNorm2d-132 [64, 256, 14, 14] 512ReLU-133 [64, 256, 14, 14] 0Conv2d-134 [64, 256, 14, 14] 589,824BatchNorm2d-135 [64, 256, 14, 14] 512ReLU-136 [64, 256, 14, 14] 0Conv2d-137 [64, 1024, 14, 14] 262,144BatchNorm2d-138 [64, 1024, 14, 14] 2,048ReLU-139 [64, 1024, 14, 14] 0Bottleneck-140 [64, 1024, 14, 14] 0Conv2d-141 [64, 512, 7, 7] 524,288BatchNorm2d-142 [64, 512, 7, 7] 1,024ReLU-143 [64, 512, 7, 7] 0Conv2d-144 [64, 512, 7, 7] 2,359,296BatchNorm2d-145 [64, 512, 7, 7] 1,024ReLU-146 [64, 512, 7, 7] 0Conv2d-147 [64, 2048, 7, 7] 1,048,576BatchNorm2d-148 [64, 2048, 7, 7] 4,096Conv2d-149 [64, 2048, 7, 7] 2,097,152BatchNorm2d-150 [64, 2048, 7, 7] 4,096ReLU-151 [64, 2048, 7, 7] 0Bottleneck-152 [64, 2048, 7, 7] 0Conv2d-153 [64, 512, 7, 7] 1,048,576BatchNorm2d-154 [64, 512, 7, 7] 1,024ReLU-155 [64, 512, 7, 7] 0Conv2d-156 [64, 512, 7, 7] 2,359,296BatchNorm2d-157 [64, 512, 7, 7] 1,024ReLU-158 [64, 512, 7, 7] 0Conv2d-159 [64, 2048, 7, 7] 1,048,576BatchNorm2d-160 [64, 2048, 7, 7] 4,096ReLU-161 [64, 2048, 7, 7] 0Bottleneck-162 [64, 2048, 7, 7] 0Conv2d-163 [64, 512, 7, 7] 1,048,576BatchNorm2d-164 [64, 512, 7, 7] 1,024ReLU-165 [64, 512, 7, 7] 0Conv2d-166 [64, 512, 7, 7] 2,359,296BatchNorm2d-167 [64, 512, 7, 7] 1,024ReLU-168 [64, 512, 7, 7] 0Conv2d-169 [64, 2048, 7, 7] 1,048,576BatchNorm2d-170 [64, 2048, 7, 7] 4,096ReLU-171 [64, 2048, 7, 7] 0Bottleneck-172 [64, 2048, 7, 7] 0
AdaptiveAvgPool2d-173 [64, 2048, 1, 1] 0Linear-174 [64, 1000] 2,049,000
================================================================
Total params: 25,557,032
Trainable params: 25,557,032
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 36.75
Forward/backward pass size (MB): 15200.01
Params size (MB): 97.49
Estimated Total Size (MB): 15334.25
----------------------------------------------------------------
参考链接：
-
https://link.zhihu.com/?target=https%3A//arxiv.org/pdf/1512.03385.pdf
-
https://blog.csdn.net/m0_64799972/article/details/132753608
-
https://blog.csdn.net/m0_50127633/article/details/117200212
-
https://www.bilibili.com/video/BV1Bo4y1T7Lc/?spm_id_from=333.999.0.0&vd_source=c7e390079ff3e10b79e23fb333bea49d
