残差⽹络学习及实现
残差⽹络ResNet由⽬前已经出到V2版本了，那么V2和V1由何不同呢，以下是阅读论⽂后的个⼈总结：
⾸先直接上图：
左边是v1，右边是v2，可以看出其主要差别是在BN和ReLU相对weight和addition的操作顺序上，即论⽂中提到的post-activation（后激活）还是pre-activation（先激活）的区别。
结合计算公式，可以看出对于V1来说f为ReLU函数，⽽V2为直接的identity mapping（恒等映射）。这样做可以带来两个好处：1.对于深层的⽹络（论⽂实验中使⽤1001层的ResNet）优化来说更简单，因为ReLU的特性所致，信号为负时，shortcut不起作⽤，所以初期训练会⼗分慢。⽽采⽤identity mapping使得⽆论训练初期还是末期，信号都能顺利在各层传输，也使得⽹络更容易训练。当然这种现象在层数不⾼时并不严重（实验使⽤低于164层）2.减少过拟合现象，在V1中虽然对输⼊进⾏了BN正则化，但是加上shortcut后，传到下个层的weight层的信号未正则化。⽽V2中对于传⼊每个层的信号都正则化可以帮助减少过拟合现象。
童话镇第二季
总结来说，对于层数较少的⽹络可以继续使⽤V1 block，对于层数较多使⽤V2 block。
最后，使⽤pytorch实现resnet：
# ResNet34 u v1 and v2
from torch import nn
import functional as F
class ResidualBlockV1(nn.Module):
# Residual Block v1
少儿读物排行榜def__init__(lf, inchannel, outchannel, stride=1, shortcut=None):
super(ResidualBlockV1, lf).__init__()
lf.basic = nn.Sequential(
nn.Conv2d(inchannel, outchannel,3, stride,1, bias=Fal),
nb是什么意思
nn.BatchNorm2d(outchannel),
nn.ReLU(inplace=True),
nn.Conv2d(outchannel, outchannel,3,1,1, bias=Fal),
三校生高复nn.BatchNorm2d(outchannel))
lf.shortcut = shortcut #shortcut function is customizable
def forward(lf, x):
什么是商务英语out = lf.basic(x)
residual = x if lf.shortcut is None el lf.shortcut(x)
out += residual
lu(out)
class ResidualBlockV2(nn.Module):
# Residual Block v2
# Residual Block v2
def__init__(lf, inchannel, outchannel, stride=1, shortcut=None):
super(ResidualBlockV2, lf).__init__()
lf.basic = nn.Sequential(
nn.BatchNorm2d(inchannel),
nn.ReLU(inplace=True),
nn.Conv2d(inchannel, outchannel,3, stride,1, bias=Fal),
nn.BatchNorm2d(outchannel),
nn.ReLU(inplace=True),
nn.Conv2d(outchannel, outchannel,3,1,1, bias=Fal))
lf.shortcut = shortcut #shortcut function is customizable
def forward(lf, x):
out = lf.basic(x)realized
residual = x if lf.shortcut is None el lf.shortcut(x)
out += residual
return out
class make_layer(nn.Module):
# Reusable layer
def__init__(lf,
inchannel,
accidentlyoutchannel,
block_num,
stride=1,
u_v1_block=True):
super(make_layer, lf).__init__()
layers =[]
shortcut = nn.Sequential(
nn.Conv2d(inchannel, outchannel,1, stride, bias=Fal),
nn.BatchNorm2d(outchannel))
if u_v1_block:
layers.append(ResidualBlockV1(inchannel, outchannel, stride, shortcut))
for i in range(1, block_num):
layers.append(ResidualBlockV1(outchannel, outchannel)) el:
layers.append(ResidualBlockV2(inchannel, outchannel, stride, shortcut))
for i in range(1, block_num):
layers.append(ResidualBlockV2(outchannel, outchannel))
lf.layer = nn.Sequential(*layers)
def forward(lf, x):
out = lf.layer(x)
return out
class ResNet34V1(nn.Module):
def__init__(lf):
super(ResNet34V1, lf).__init__()
lf.pre = nn.Sequential(
nn.Conv2d(3,64,7,2,3, bias=Fal),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.MaxPool2d(3,2,1))
# build each layer
lf.layer1 = make_layer(64,64,3)
lf.layer2 = make_layer(64,128,4, stride=2)
lf.layer3 = make_layer(128,256,6, stride=2)
lf.layer4 = make_layer(256,512,3, stride=2)
lf.fc = nn.Linear(512,1000)
def forward(lf, x):
def forward(lf, x):
x = lf.pre(x)
x = lf.layer1(x)
x = lf.layer2(x)
x = lf.layer3(x)
x = lf.layer4(x)
x = F.avg_pool2d(x,7)
x = x.view(x.size(0),-1)
esp是什么return lf.fc(x)
class ResNet34V2(nn.Module):
def__init__(lf):
super(ResNet34V2, lf).__init__()
lf.pre = nn.Sequential(
nn.Conv2d(3,64,7,2,3, bias=Fal),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
nn.MaxPool2d(3,2,1))
# build each layer野心英文
lf.layer1 = make_layer(64,64,3, u_v1_block=Fal)
lf.layer2 = make_layer(64,128,4, stride=2, u_v1_block=Fal) lf.layer3 = make_layer(128,256,6, stride=2, u_v1_block=Fal) lf.layer4 = make_layer(256,512,3, stride=2, u_v1_block=Fal) lf.fc = nn.Linear(512,1000)
def forward(lf, x):
x = lf.pre(x)英豪教育
x = lf.layer1(x)
x = lf.layer2(x)
x = lf.layer3(x)
x = lf.layer4(x)
x = F.avg_pool2d(x,7)
x = x.view(x.size(0),-1)
return lf.fc(x)
print('This is ResNet34 v1:')
print(ResNet34V1())
print('This is ResNet34 v2:')
print(ResNet34V2())
参考：
1.
2.
