我們都知道神經網路深度不斷增加，會出現兩個問題：
1.梯度彌散、梯度爆炸
2.退化問題，訓練集的準確率下降
基于以上問題何凱明于2015年提出了殘差神經網路（ResNet），此網路在深度夠深的前提下訓練集的準確率也不會下降太多，而且能一直保持最優狀態，那它具體是怎么構建的呢?
以Resnet18為例，它是由殘差塊堆疊而成的網路：
1個卷積層+8個殘差塊（每個殘差塊有2個卷積層）+1個全連接層
如下圖：
我個人認為ResNet是最適合深度學習小白研究的網路之一，知曉了原理，看著圖就能搭建好網路，代碼風格清晰明朗，網上有博主形容其“簡單而實用”，我看確實如此，在一開始學習ResNet時，我試著成功搭建10層的Resnet用于訓練我的垃圾分類資料集，今晚將其改成18層，
代碼如下：
import torch
from torch import nn
from torch.nn import functional as F
class ResBlk(nn.Module):
"""
resnet block
"""
def __init__(self, ch_in, ch_out, stride=1):
"""
:param ch_in:
:param ch_out:
"""
super(ResBlk, self).__init__()
self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=stride, padding=1)
self.bn1 = nn.BatchNorm2d(ch_out)
self.conv2 = nn.Conv2d(ch_out, ch_out, kernel_size=3, stride=1, padding=1)
self.bn2 = nn.BatchNorm2d(ch_out)
self.extra = nn.Sequential()
if ch_out != ch_in:
# [b, ch_in, h, w] => [b, ch_out, h, w]
self.extra = nn.Sequential(
nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=stride),
nn.BatchNorm2d(ch_out)
)
def forward(self, x):
"""
:param x: [b, ch, h, w]
:return:
"""
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out = self.extra(x) + out
out = F.relu(out)
return out
class ResNet18(nn.Module):
def __init__(self, num_class):
super(ResNet18, self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
)
# conv1:[2,3,224,224]=>[2,64,56,56]
self.blk1 = ResBlk(64, 64)
# [2, 64, 56, 56] => [2, 64, 56, 56]
self.blk2 = ResBlk(64, 128, stride=2)
# [2, 64, 56, 56]=>[2, 128, 28, 28]
self.blk2_1 = ResBlk(128, 128)
# [2, 128, 28, 28] =>[2, 128, 28, 28]
self.blk3 = ResBlk(128, 256, stride=2)
# [2, 128, 28, 28]=>[2, 256, 14, 14]
self.blk3_1 = ResBlk(256, 256)
# [2, 256, 14, 14]=>[2, 256, 14, 14]
self.blk4 = ResBlk(256, 512, stride=2)
# [2, 256, 14, 14]=>[2, 512, 7, 7]
self.blk4_1 = ResBlk(512, 512)
# [2, 512, 7, 7]=>[2, 512, 7, 7]
self.pool2 = nn.AvgPool2d(kernel_size=7, stride=1, padding=0)
self.outlayer = nn.Linear(512, num_class)
def forward(self, x):
"""
:param x:
:return:
"""
x = F.relu(self.conv1(x)) # conv1:[b,3,224,224]=>[b,64,56,56]
# print(x.shape)
x = self.blk1(x) # [b, 64, 56, 56]=>[2, 64, 56, 56]
x = self.blk1(x) # [b, 64, 56, 56]=>[2, 64, 56, 56]
x = self.blk2(x) # [2, 64, 56, 56]=>[2, 128, 28, 28]
x = self.blk2_1(x) # [2, 128, 28, 28] =>[2, 128, 28, 28]
x = self.blk3(x) # [2, 128, 28, 28]=>[2, 256, 14, 14]
x = self.blk3_1(x) # [2, 256, 14, 14]=>[2, 256, 14, 14]
x = self.blk4(x) # [2, 256, 14, 14]=>[2, 512, 7, 7]
x = self.blk4_1(x) # [2, 512, 7, 7]=>[2, 512, 7, 7]
x = self.pool2(x) # [2, 512, 7, 7]=>[2,512,1,1]
x = x.view(x.size(0), -1) # flatten
x = self.outlayer(x)
return x
def main():
model = ResNet18(7)
tmp = torch.randn(1, 3, 224, 224)
out = model(tmp)
print('resnet:', out.shape)
if __name__ == '__main__':
main()
參考：
（圖出自鏈接1，0基礎ResNet建議看第2篇）
https://blog.csdn.net/weixin_44331304/article/details/106127552?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522161779576016780271556440%2522%252C%2522scm%2522%253A%252220140713.130102334…%2522%257D&request_id=161779576016780271556440&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2alltop_positive~default-1-106127552.first_rank_v2_pc_rank_v29&utm_term=resnet18
https://blog.csdn.net/weixin_44331304/article/details/106127552?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522161779576016780271556440%2522%252C%2522scm%2522%253A%252220140713.130102334…%2522%257D&request_id=161779576016780271556440&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2alltop_positive~default-1-106127552.first_rank_v2_pc_rank_v29&utm_term=resnet18
https://www.bilibili.com/video/BV1j64y1D748?p=102&spm_id_from=pageDriver
轉載請註明出處，本文鏈接：https://www.uj5u.com/qita/273646.html
標籤：AI