原文地址: https://arxiv.org/pdf/1512.03385.pdf
论文就不解读了,大部分解读都是翻译,看的似懂非懂,自己搞懂就行了。
最近想着实现一下经典的网络结构,看了原文之后,根据原文代码结构开始实现。
起初去搜了下各种版本的实现,发现很多博客都是错误百出,有些博文都发布几年了,错误还是没人发现,评论区几十号人不知道是真懂还是装懂,颇有些无奈啊。
因此打算自己手动实现网络结构,锻炼下自己的代码能力,也加深对网络结构的理解。
写完之后也很欣慰,毕竟一直认为自己是个菜鸡,最近竟然接连不断的发现很多博文的错误之处,而且很多人看后都没发现的,想想自己似乎还有点小水平。
最后在一套代码里,实现了各版本ResNet,为了方便。
其实最后还是觉得应该每个网络分开写比较好。因为不同版本的网络内部操作是有很大差异的,本文下面的代码是将ResidualBlock和 BottleNeckBlock分开写的,但是在维度的变换上差异还是很复杂,一方面想提高代码的复用性,另一方面也受制于复杂度。所以最后写出的算不上高复用性的精简代码。勉强能用。关于ResNet的结构,除各版本分开写之外,重复的block其实也可以分开写,因为BottleNeckBlock的维度变换太复杂,参数变换多,能分开就分开,复杂度小的地方可以复用。
以下是网络结构和实现代码,检验后都是对的;水平有限,如发现有错误,欢迎评论告知!
import torch.nn as nn
from torch.nn import functional as F
class ResNetModel(nn.Module):
""" 实现通用的ResNet模块,可根据需要定义 """
def __init__(self, num_classes=1000, layer_num=[],bottleneck = False):
super(ResNetModel, self).__init__()
#conv1
self.pre = nn.Sequential(
#in 224*224*3
nn.Conv2d(3,64,7,2,3,bias=False), #输入通道3,输出通道64,卷积核7*7*64,步长2,根据以上计算出padding=3
#out 112*112*64
nn.BatchNorm2d(64), #输入通道C = 64
nn.ReLU(inplace=True), #inplace=True, 进行覆盖操作
# out 112*112*64
nn.MaxPool2d(3,2,1), #池化核3*3,步长2,计算得出padding=1;
# out 56*56*64
)
if bottleneck: #resnet50以上使用BottleNeckBlock
self.residualBlocks1 = self.add_layers(64, 256, layer_num[0], 64, bottleneck=bottleneck)
self.residualBlocks2 = self.add_layers(128, 512, layer_num[1], 256, 2,bottleneck)
self.residualBlocks3 = self.add_layers(256, 1024, layer_num[2], 512, 2,bottleneck)
self.residualBlocks4 = self.add_layers(512, 2048, layer_num[3], 1024, 2,bottleneck)
self.fc = nn.Linear(2048, num_classes)
else: #resnet34使用普通ResidualBlock
self.residualBlocks1 = self.add_layers(64,64,layer_num[0])
self.residualBlocks2 = self.add_layers(64,128,layer_num[1])
self.residualBlocks3 = self.add_layers(128,256,layer_num[2])
self.residualBlocks4 = self.add_layers(256,512,layer_num[3])
self.fc = nn.Linear(512, num_classes)
def add_layers(self, inchannel, outchannel, nums, pre_channel=64, stride=1, bottleneck=False):
layers = []
if bottleneck is False:
#添加大模块首层, 首层需要判断inchannel == outchannel ?
#跨维度需要stride=2,shortcut也需要1*1卷积扩维
layers.append(ResidualBlock(inchannel,outchannel))
#添加剩余nums-1层
for i in range(1,nums):
layers.append(ResidualBlock(outchannel,outchannel))
return nn.Sequential(*layers)
else: #resnet50使用bottleneck
#传递每个block的shortcut,shortcut可以根据是否传递pre_channel进行推断
#添加首层,首层需要传递上一批blocks的channel
layers.append(BottleNeckBlock(inchannel,outchannel,pre_channel,stride))
for i in range(1,nums): #添加n-1个剩余blocks,正常通道转换,不传递pre_channel
layers.append(BottleNeckBlock(inchannel,outchannel))
return nn.Sequential(*layers)
def forward(self, x):
x = self.pre(x)
x = self.residualBlocks1(x)
x = self.residualBlocks2(x)
x = self.residualBlocks3(x)
x = self.residualBlocks4(x)
x = F.avg_pool2d(x, 7)
x = x.view(x.size(0), -1)
return self.fc(x)
class ResidualBlock(nn.Module):
''' 定义普通残差模块 resnet34为普通残差块,resnet50为瓶颈结构 '''
def __init__(self, inchannel, outchannel, stride=1, padding=1, shortcut=None):
super(ResidualBlock, self).__init__()
#resblock的首层,首层如果跨维度,卷积stride=2,shortcut需要1*1卷积扩维
if inchannel != outchannel:
stride= 2
shortcut=nn.Sequential(
nn.Conv2d(inchannel,outchannel,1,stride,bias=False),
nn.BatchNorm2d(outchannel)
)
# 定义残差块的左部分
self.left = nn.Sequential(
nn.Conv2d(inchannel, outchannel, 3, stride, padding, bias=False),
nn.BatchNorm2d(outchannel),
nn.ReLU(inplace=True),
nn.Conv2d(outchannel, outchannel, 3, 1, padding, bias=False),
nn.BatchNorm2d(outchannel),
)
#定义右部分
self.right = shortcut
def forward(self, x):
out = self.left(x)
residual = x if self.right is None else self.right(x)
out = out + residual
return F.relu(out)
class BottleNeckBlock(nn.Module):
''' 定义resnet50的瓶颈结构 '''
def __init__(self,inchannel,outchannel, pre_channel=None, stride=1,shortcut=None):
super(BottleNeckBlock, self).__init__()
#首个bottleneck需要承接上一批blocks的输出channel
if pre_channel is None: #为空则表示不是首个bottleneck,
pre_channel = outchannel #正常通道转换
else: # 传递了pre_channel,表示为首个block,需要shortcut
shortcut = nn.Sequential(
nn.Conv2d(pre_channel,outchannel,1,stride,0,bias=False),
nn.BatchNorm2d(outchannel)
)
self.left = nn.Sequential(
#1*1,inchannel
nn.Conv2d(pre_channel, inchannel, 1, stride, 0, bias=False),
nn.BatchNorm2d(inchannel),
nn.ReLU(inplace=True),
#3*3,inchannel
nn.Conv2d(inchannel,inchannel,3,1,1,bias=False),
nn.BatchNorm2d(inchannel),
nn.ReLU(inplace=True),
#1*1,outchannel
nn.Conv2d(inchannel,outchannel,1,1,0,bias=False),
nn.BatchNorm2d(outchannel),
nn.ReLU(inplace=True),
)
self.right = shortcut
def forward(self,x):
out = self.left(x)
residual = x if self.right is None else self.right(x)
return F.relu(out+residual)
if __name__ == '__main__':
# channel_nums = [64,128,256,512,1024,2048]
num_classes = 6
#layers = 18, 34, 50, 101, 152
layer_nums = [[2,2,2,2],[3,4,6,3],[3,4,6,3],[3,4,23,3],[3,8,36,3]]
#选择resnet版本,
# resnet18 ——0;resnet34——1,resnet-50——2,resnet-101——3,resnet-152——4
i = 3;
bottleneck = i >= 2 #i<2, false,使用普通的ResidualBlock; i>=2,true,使用BottleNeckBlock
model = ResNetModel(num_classes,layer_nums[i],bottleneck)
print(model)
版权声明:本文内容由互联网用户自发贡献,该文观点仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 举报,一经查实,本站将立刻删除。
发布者:全栈程序员栈长,转载请注明出处:https://javaforall.cn/185104.html原文链接:https://javaforall.cn