残差网络(Residual Net, ResNet)自从2015年面世以来,凭借其优异的性能在ILSVRC中以绝对优势获得第一名,并成功应用于许多领域。

1. 传统深度网络的问题

在深度学习或者神经领域的研究中,一般认为,网络越深(层数越多),网络的性能应该会更好,因为更深的网络倾向于有更强大的表示能力,即网络容量越大。

但是在实际过程中,我们发现过深的网络反而会导致性能的下降。在网络结构的设计中似乎存在一种“阈值”,在到达一定的层数之后,训练误差和测试误差都会加大。下图为一个24层网络和一个56层网络在CIFAR10数据集的训练表现。
在这里插入图片描述
显然,这种性能的下降并不是因为过拟合引起的。因为过拟合意味着训练误差正常减小,而测试误差显著增大。

对这种现象的一种解释是,在网络深度过深的时候,低层参数的细微变动都会引起高层参数的剧烈变化,优化算法没有能力去得到一个最优解。

做这样一个假设,假设有一个50层的网络,但在其优化过程中,最容易优化出最佳解的层数是25,那么这个网络的后25层应当作为一个恒等映射
x 25 = f 1 ( x ) \mathbf{x}_{25}=f_{1}(\mathbf{x}) x25=f1(x) o u t = f 2 ( x 25 ) \mathbf{out}=f_{2}(\mathbf{x}_{25}) out=f2(x25)

由于神经网络由非线性层组合而成,学习一个恒等映射是比较困难的。优化算法的局限性使得“冗余”的网络层学习到了不适合恒等映射的参数。

2. 残差结构和残差网络

2.1 残差是什么

残差的统计学定义:实际观测值和估计值(拟合值)之间的差值。

如果存在某个k层的网络 F F F是当前最优的网络,那么可以构造一个更深的网络,其最后几层仅是网络f第k层输出的恒等映射,就可以取得与 F F F一致的结果

如果k还不是最佳层数,那么更深的网络就可以取得更好的结果。所以,如果深层网络的效果不如浅层网络,那么则说明新加入层不好学

如果不好学,则可以使用类似“分治法”,分开求解恒等映射和非恒等映射。
x代表之前浅层网络已经学到的东西
F(x)代表已经学到的东西和要学的东西的之间的残差

现在只学F(x)就能与x组合起来。

H ( x ) = F ( x ) + x F ( x ) = H ( x ) − x H(\mathbf{x})=F(\mathbf{x})+\mathbf{x} \\ F(\mathbf{x})=H(\mathbf{x})-\mathbf{x} H(x)=F(x)+xF(x)=H(x)x

x x x成为恒等映射,那么只需要学习残差 F ( x ) F(x) F(x)作为非恒等映射。
残差在这里,指的是直接的映射H(x)与快捷连接x的差值,也就是 F ( x ) F(\mathbf{x}) F(x)

2.2 残差模块 Residual Block

据此,我们设计一个残差模块(Residual Block)的结构如下:
在这里插入图片描述

y l = h ( x l ) + F ( x l , W l ) \mathbf{y}_{l}=h(\mathbf{x}_{l})+\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l}) yl=h(xl)+F(xl,Wl) x l + 1 = f ( y l ) \mathbf{x}_{l+1}=f(\mathbf{y}_{l}) xl+1=f(yl)
在网络实现中:
h ( x l ) = x l f = R e L U h(\mathbf{x}_{l})=\mathbf{x}_{l}\qquad f=\mathrm{ReLU} h(xl)=xlf=ReLU x l + 1 ≡ y l \mathbf{x}_{l+1}\equiv\mathbf{y}_{l} xl+1yl
最后得到的残差模块表达式如下:
x l + 1 = x l + F ( x l , W l ) \mathbf{x}_{l+1}=\mathbf{x}_{l}+\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l}) xl+1=xl+F(xl,Wl)

2.3 基本模块BasicBlock和BottleNeck

在残差网络中,基本的残差模块由两个3×3的卷积层和ReLU激活函数、BatchNorm层组成。其结构如下(以64个channel的输入为例):

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)
)

在网络层数过深的时候,考虑到训练成本,作者提出了一种新的结构设计BottleNeck。将原来的两个3×3的卷积层变为两个1×1的卷积层和一个3×3的卷积层。其中,两个1×1的卷积层负责降低/恢复通道维度,3×3的卷积层负责“真正”的卷积运算。其结构如下图右图所示,BottleNeck的运算具有更小的时间复杂度。
在这里插入图片描述

2.4 残差网络ResNet设计

在文中,作者给出了ResNet18、ResNet34、ResNet50、ResNet101、ResNet152等网络设计,分别对应不同层数的卷积运算层。在50层及以上的网络中,都使用BottleNeck结构进行网络构建。
在这里插入图片描述

2.4.1 恒等映射与残差的连接

在网络设计中,特别需要注意的是在不同layer之间,例如conv2_x和conv3_x间,输出和输入的尺寸是不一样的,如下图虚线所示。

对于残差运算,这可以很简单的通过卷积进行尺寸的变换,对于恒等映射,作者考虑了如下几个方法进行变换:

  1. 给恒等映射 x \mathbf{x} x添加0,扩充其维度
  2. 用一个1×1的卷积进行下采样
    在这里插入图片描述

在代码实现中,采用下采样的方法对恒等映射进行变换。

3. Forward/Backward Propagation

3.1 Forward propogation

传统网络

F ( x , w ) = x w F(x,w)=xw F(x,w)=xw
x L = F ( x L − 1 , w L − 1 ) = F ( F ( x L − 2 , w L − 2 ) , w L − 1 ) ⋯ = ∏ i = 1 L − 1 x i w i x_{L}=F(x_{L-1},w_{L-1})=F(F(x_{L-2},w_{L-2}),w_{L-1})\cdots=\prod_{i=1}^{L-1}{x_{i}w_{i}} xL=F(xL1,wL1)=F(F(xL2,wL2),wL1)=i=1L1xiwi

残差网络
x 2 = x 1 + F ( x 1 , w 1 ) x_{2}=x_{1}+F(x_1,w_{1}) x2=x1+F(x1,w1)
x 3 = x 2 + F ( x 2 , w 2 ) = x 1 + F ( x 1 , w 1 ) + F ( x 2 , w 2 ) x_{3}=x_{2}+F(x_2,w_{2})=x_{1}+F(x_{1},w_{1})+F(x_{2},w_{2}) x3=x2+F(x2,w2)=x1+F(x1,w1)+F(x2,w2)
⋯ \cdots
x L = x 1 + ∑ i = 1 L − 1 F ( x i , w i ) x_{L}=x_{1}+\sum_{i=1}^{L-1}{F(x_{i},w_{i})} xL=x1+i=1L1F(xi,wi)

3.2 Back Propogation

传统网络
浅层网络是 g ( x ) g(x) g(x),加入层以后变成 f ( g ( x ) ) f(g(x)) f(g(x))
∂ f ( g ( x ) ) ∂ x = ∂ f ( g ( x ) ) ∂ g ( x ) ∂ g ( x ) ∂ x \frac{\partial{f(g(x))}}{\partial{x}}=\frac{\partial{f(g(x))}}{\partial{{g(x)}}}\frac{\partial{g(x)}}{\partial{x}} xf(g(x))=g(x)f(g(x))xg(x)
残差网络
∂ ( f ( g ( x ) ) + g ( x ) ) ∂ x = ∂ f ( g ( x ) ) ∂ g ( x ) ∂ g ( x ) ∂ x + ∂ g ( x ) ∂ x \frac{\partial{(f(g(x))+g(x))}}{\partial{x}}=\frac{\partial{f(g(x))}}{\partial{{g(x)}}}\frac{\partial{g(x)}}{\partial{x}}+\frac{\partial{g(x)}}{\partial{x}} x(f(g(x))+g(x))=g(x)f(g(x))xg(x)+xg(x)

可以看到,在求梯度的过程中,残差网络相比传统网络多加了一项,这有利于解决梯度消失,是网络训练的更快

损失函数对网络的第 l l l求梯度:
传统网络
∂ L o s s ∂ x l = ∂ L o s s ∂ x L ∂ x L ∂ x l = ∂ L o s s ∂ x L ∂ ∏ i = 1 L − 1 x i w i ∂ x l \frac{\partial{Loss}}{\partial{x_{l}}}=\frac{\partial{Loss}}{\partial{x_{L}}}\frac{\partial{x_{L}}}{\partial{x_{l}}}=\frac{\partial{Loss}}{\partial{x_{L}}}\frac{\partial{\prod_{i=1}^{L-1}{x_{i}w_{i}}}}{\partial{x_{l}}} xlLoss=xLLossxlxL=xLLossxli=1L1xiwi
残差网络
∂ L o s s ∂ x l = ∂ L o s s ∂ x L ∂ x L ∂ x l = ∂ L o s s ∂ x L ( 1 + ∂ ∑ i = l L − 1 F ( x i , w i ) ∂ x l ) \frac{\partial{Loss}}{\partial{x_{l}}}=\frac{\partial{Loss}}{\partial{x_{L}}}\frac{\partial{x_{L}}}{\partial{x_{l}}}=\frac{\partial{Loss}}{\partial{x_{L}}}(1+\frac{\partial{\sum_{i=l}^{L-1}{F(x_{i},w_{i})}}}{\partial{x_{l}}}) xlLoss=xLLossxlxL=xLLoss(1+xli=lL1F(xi,wi))

可以看到,在残差网络中,梯度由乘法变加法,这可以有效缓解梯度消失和梯度爆炸。

4. 代码分析

PyTorch现已将ResNet整合为python库,可以直接调用。源码的地址如下:
https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py
基本的卷积层

def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=dilation, groups=groups, bias=False, dilation=dilation)


def conv1x1(in_planes, out_planes, stride=1):
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)

其中,对输入进行下采样在卷积层中有两种实现方式:

Basic Block
其中,expension代表了经过一个Block之后,channel数量的变化。这里输出channel维度与预设一样,expension为1。

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
                 base_width=64, dilation=1, norm_layer=None):
        super(BasicBlock, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        if groups != 1 or base_width != 64:
            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
        if dilation > 1:
            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = norm_layer(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = norm_layer(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

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

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

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

        out += identity
        out = self.relu(out)

        return out

BottleNeck
BottleNeck输出的channel输出是对应BasicBlock的4倍,所以expension=4

class Bottleneck(nn.Module):
    # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
    # while original implementation places the stride at the first 1x1 convolution(self.conv1)
    # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
    # This variant is also known as ResNet V1.5 and improves accuracy according to
    # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.

    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
                 base_width=64, dilation=1, norm_layer=None):
        super(Bottleneck, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        width = int(planes * (base_width / 64.)) * groups
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv1x1(inplanes, width)
        self.bn1 = norm_layer(width)
        self.conv2 = conv3x3(width, width, stride, groups, dilation)
        self.bn2 = norm_layer(width)
        self.conv3 = conv1x1(width, planes * self.expansion)
        self.bn3 = norm_layer(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = 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:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

ResNet
在构造每个层的时候,要注意在输出通道和输入通道数量不一致的时候,要添加一个下采样层对恒等映射进行下采样

class ResNet(nn.Module):

    def __init__(self, block, layers, num_classes=1000, zero_init_residual=False,
                 groups=1, width_per_group=64, replace_stride_with_dilation=None,
                 norm_layer=None):
        super(ResNet, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.inplanes = 64
        self.dilation = 1
        if replace_stride_with_dilation is None:
            # each element in the tuple indicates if we should replace
            # the 2x2 stride with a dilated convolution instead
            replace_stride_with_dilation = [False, False, False]
        if len(replace_stride_with_dilation) != 3:
            raise ValueError("replace_stride_with_dilation should be None "
                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
        self.groups = groups
        self.base_width = width_per_group
        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = norm_layer(self.inplanes)
        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=2,
                                       dilate=replace_stride_with_dilation[0])
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
                                       dilate=replace_stride_with_dilation[1])
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
                                       dilate=replace_stride_with_dilation[2])
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:
            self.dilation *= stride
            stride = 1
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes, groups=self.groups,
                                base_width=self.base_width, dilation=self.dilation,
                                norm_layer=norm_layer))

        return nn.Sequential(*layers)

    def _forward_impl(self, x):
        # See note [TorchScript super()]
        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.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)

        return x

    def forward(self, x):
        return self._forward_impl(x)

5. 恒等映射

对残差做一个简单的改进: x l + 1 = λ l x l + F ( x l , W l ) \mathbf{x}_{l+1}=\lambda_{l}\mathbf{x}_{l}+\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l}) xl+1=λlxl+F(xl,Wl)
则有:
x L = ( ∏ i = l L − 1 λ i ) x l + ∑ i = l L − 1 F ( x l , W l ) \mathbf{x}_{L}=(\prod_{i=l}^{L-1}\lambda_{i})\mathbf{x}_{l}+\sum_{i=l}^{L-1}{\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l})} xL=(i=lL1λi)xl+i=lL1F(xl,Wl)
求梯度,有:
∂ L o s s ∂ x l = ∂ L o s s ∂ x L ∂ x L ∂ x l = ∂ L o s s ∂ x L ( ∏ i = l L − 1 λ i + ∂ ∑ i = l L − 1 F ( x l , W l ) ∂ x l ) \frac{\partial{Loss}}{\partial{\mathbf{x}_{l}}}=\frac{\partial{Loss}}{\partial{\mathbf{x}_{L}}}\frac{\partial{\mathbf{x}_{L}}}{\partial{\mathbf{x}_{l}}}=\frac{\partial{Loss}}{\partial{\mathbf{x}_{L}}}(\prod_{i=l}^{L-1}{\lambda_{i}}+\frac{\partial{\sum_{i=l}^{L-1}{\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l})}}}{\partial{\mathbf{x}_{l}}}) xlLoss=xLLossxlxL=xLLoss(i=lL1λi+xli=lL1F(xl,Wl))

可以看到, λ \lambda λ大于1的时候,累乘会造成梯度爆炸;在小于1的时候,累乘会造成梯度消失。

6. 分析残差连接

作者给出了一下几种残差连接的变体:
Original
x l + 1 = x l + F ( x l , W l ) \mathbf{x}_{l+1}=\mathbf{x}_{l}+\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l}) xl+1=xl+F(xl,Wl)constant
x l + 1 = λ 1 x l + λ 2 F ( x l , W l ) \mathbf{x}_{l+1}=\lambda_{1}\mathbf{x}_{l}+\lambda_{2}\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l}) xl+1=λ1xl+λ2F(xl,Wl)exclusive gating
x l + 1 = ( 1 − g ( x l ) ) x l + g ( x l ) F ( x l , W l ) \mathbf{x}_{l+1}=(1-g(\mathbf{x}_{l}))\mathbf{x}_{l}+g(\mathbf{x}_{l})\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l}) xl+1=(1g(xl))xl+g(xl)F(xl,Wl)shortcut-only gating
x l + 1 = ( 1 − g ( x l ) ) x l + F ( x l , W l ) \mathbf{x}_{l+1}=(1-g(\mathbf{x}_{l}))\mathbf{x}_{l}+\mathcal{F}(\mathbf{x}_{l},\mathcal{W}_{l}) xl+1=(1g(xl))xl+F(xl,Wl)
其余还包括1×1 conv shortcutdropout shortcut

这几种残差连接的示意图如下所示:
在这里插入图片描述
作者给出的实验结果如下:
在这里插入图片描述
可以看出,原版的残差连接时效果最好的。使用exclusive gate的效果则强烈依赖于偏差的设定。

7. 不同结构的残差模块

作者接下来分析了不同残差模块的设计带来的影响。
在这里插入图片描述

  • 在(b)中,由于BN层的存在,使得 x l + 1 = f ( y l ) \mathbf{x}_{l+1}=f(\mathbf{y}_{l}) xl+1=f(yl)不再是一个线性映射,这会影响残差网络的性能
  • 在(c)中,残差目标最后的ReLU激活层使得残差的输出范围是非负的。然而,无论是数学定义上还是经验上,残差的范围应该是 ( − ∞ , + ∞ ) (-\infty,+\infty) (,+),非负的残差影响模型性能
  • zai(d)和(e)中,作者采用了一种pre-activation的想法。
    在原版设计中, f f f会影响到残差模块的两个部分:
    y l + 1 = f ( y l ) + F ( f ( y l ) , W l + 1 ) \mathbf{y}_{l+1}=f(\mathbf{y}_{l})+\mathcal{F}(f(\mathbf{y}_{l}),\mathcal{W}_{l+1}) yl+1=f(yl)+F(f(yl),Wl+1)
    pre-activation使得 f f f只影响残差部分,不影响恒等映射
    y l + 1 = y l + F ^ ( f ( y l ) , W l + 1 ) \mathbf{y}_{l+1}=\mathbf{y}_{l}+\hat{\mathcal{F}}(f(\mathbf{y}_{l}),\mathcal{W}_{l+1}) yl+1=yl+F^(f(yl),Wl+1)
    在网络设计中,结构如下:
    在这里插入图片描述
    实际上,当我们将BN层一同放在卷积层的前面的时候,网络性能会进一步提升,这可以看做BN层起到了正则化的作用。
Logo

开放原子开发者工作坊旨在鼓励更多人参与开源活动,与志同道合的开发者们相互交流开发经验、分享开发心得、获取前沿技术趋势。工作坊有多种形式的开发者活动,如meetup、训练营等,主打技术交流,干货满满,真诚地邀请各位开发者共同参与!

更多推荐