上一篇博客学习了如何搭建Inception网络,这篇博客主要讲述如何利用pytorch搭建ResNets网络。
上一篇博客中遗留了一个问题,就是1*1卷积核的作用,第一个作用是减少参数,第二个作用是压缩通道数,减少计算量。
理论上,随着网络深度的加深,训练应该越来越好,但是,如果没有残差网络,深度越深意味着用优化算法越难计算,ResNets网络模型优点在于它能够训练深层次的网络模型,并且有助于解决梯度消失和梯度爆炸的问题,而且能保证良好的性能。
从上图中可以看出,Resnets网络在计算时,在执行最后一个步骤的激活时,加上了原先的x的值,这样的操作就是防止梯度消失。
import torch
from torchvision import datasets
from torchvision import transforms
from torch.utils.data import DataLoader
import torch.nn.functional as F
#数据增强
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,),(0.3081,))
])
#构造数据集
train_dataset = datasets.MNIST(
root='../dataset/mnist',
download=False,
train=True,
transform=transform
)
test_dataset = datasets.MNIST(
root='../dataset/mnist',
download=False,
train=False,
transform = transform
)
train_loader = DataLoader(
dataset=train_dataset,
batch_size=64,
shuffle=True
)
test_loader = DataLoader(
dataset=test_dataset,
batch_size=64,
shuffle=True
)
这些代码都是在这一系列实验中共有的部分,不在做过多的解释。
#构造残差模块
class ResidualBlock(torch.nn.Module):
def __init__(self,channels):
super(ResidualBlock,self).__init__()
self.channels = channels
#same卷积
self.conv1 = torch.nn.Conv2d(channels,channels,kernel_size=3,padding=1)
self.conv2 = torch.nn.Conv2d(channels,channels,kernel_size=3,padding=1)
def forward(self,x):
y = F.relu(self.conv1(x))
y = self.conv2(y)
return F.relu(x+y)
从这段代码中,首先对数据x进行卷积操作和激活操作得到y,然后对y进行卷积操作处理得到新的y,最后对原始和x加上y进行激活操作。
为了保证x可以和y相加,这个网络中采用的都是same卷积,这样会使图片数据的高度和宽度不变。并且这段代码的具有很高的重用性,在构造时,可以传入相应的通道数,这样他就可以作为一个单独的模块和其他网络一起构造。
#构造网络模型
class Net(torch.nn.Module):
def __init__(self):
super(Net,self).__init__()
self.conv1 = torch.nn.Conv2d(1,16,kernel_size=5)
self.conv2 = torch.nn.Conv2d(16,32,kernel_size=5)
#最大池化
self.mp = torch.nn.MaxPool2d(2)
self.rblock1 = ResidualBlock(channels=16)
self.rblock2 = ResidualBlock(channels=32)
self.fc = torch.nn.Linear(512,10)
def forward(self,x):
batch_size = x.size(0)
x = self.mp(F.relu(self.conv1(x)))
x = self.rblock1(x)
x = self.mp(F.relu(self.conv2(x)))
x = self.rblock2(x)
x = x.view(batch_size,-1)
x = self.fc(x)
return x
model = Net()
看代码不如看网络构件图直觉,所以我画了一个简单的图形。
上面那段代码就是根据这个网络构建图来写的,就不做过多的解释了。
model = Net()
#构造损失
criterion= torch.nn.CrossEntropyLoss()
#构造优化
optimizer = torch.optim.SGD(model.parameters(),lr=0.01,momentum=0.5)
#训练模型
def train(epoch):
running_loss = 0
for batchix,datas in enumerate(train_loader,0):
inputs,target = datas
optimizer.zero_grad()
label = model(inputs)
loss = criterion(label,target)
loss.backward()
optimizer.step()
running_loss+=loss.item()
if batchix%300==299:
print('[%d,%3d] 损失值:%.3f'%(epoch+1,batchix+1,running_loss/300))
running_loss = 0
#测试模型
def test():
total = 0
correct = 0
with torch.no_grad():
for data in (test_loader):
inputs,label = data
output = model(inputs)
_,pre = torch.max(output,dim=1)
total += label.size(0)
correct += (pre==label).sum().item()
print('准确率为%.3f'%(correct/total*100))
这些代码都是重用度非常高的代码,我每次学习一个新的网络结构,我都要在写一次,增加自己对网络结构的感觉并且练习一下这些常用的代码。
import torch
from torchvision import datasets
from torchvision import transforms
from torch.utils.data import DataLoader
import torch.nn.functional as F
#数据增强
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,),(0.3081,))
])
#构造数据集
train_dataset = datasets.MNIST(
root='../dataset/mnist',
download=False,
train=True,
transform=transform
)
test_dataset = datasets.MNIST(
root='../dataset/mnist',
download=False,
train=False,
transform = transform
)
train_loader = DataLoader(
dataset=train_dataset,
batch_size=64,
shuffle=True
)
test_loader = DataLoader(
dataset=test_dataset,
batch_size=64,
shuffle=True
)
#构造残差模块
class ResidualBlock(torch.nn.Module):
def __init__(self,channels):
super(ResidualBlock,self).__init__()
self.channels = channels
#same卷积
self.conv1 = torch.nn.Conv2d(channels,channels,kernel_size=3,padding=1)
self.conv2 = torch.nn.Conv2d(channels,channels,kernel_size=3,padding=1)
def forward(self,x):
y = F.relu(self.conv1(x))
y = self.conv2(y)
return F.relu(x+y)
#构造网络模型
class Net(torch.nn.Module):
def __init__(self):
super(Net,self).__init__()
self.conv1 = torch.nn.Conv2d(1,16,kernel_size=5)
self.conv2 = torch.nn.Conv2d(16,32,kernel_size=5)
#最大池化
self.mp = torch.nn.MaxPool2d(2)
self.rblock1 = ResidualBlock(channels=16)
self.rblock2 = ResidualBlock(channels=32)
self.fc = torch.nn.Linear(512,10)
def forward(self,x):
batch_size = x.size(0)
x = self.mp(F.relu(self.conv1(x)))
x = self.rblock1(x)
x = self.mp(F.relu(self.conv2(x)))
x = self.rblock2(x)
x = x.view(batch_size,-1)
x = self.fc(x)
return x
model = Net()
#构造损失
criterion= torch.nn.CrossEntropyLoss()
#构造优化
optimizer = torch.optim.SGD(model.parameters(),lr=0.01,momentum=0.5)
#训练模型
def train(epoch):
running_loss = 0
for batchix,datas in enumerate(train_loader,0):
inputs,target = datas
optimizer.zero_grad()
label = model(inputs)
loss = criterion(label,target)
loss.backward()
optimizer.step()
running_loss+=loss.item()
if batchix%300==299:
print('[%d,%3d] 损失值:%.3f'%(epoch+1,batchix+1,running_loss/300))
running_loss = 0
#测试模型
def test():
total = 0
correct = 0
with torch.no_grad():
for data in (test_loader):
inputs,label = data
output = model(inputs)
_,pre = torch.max(output,dim=1)
total += label.size(0)
correct += (pre==label).sum().item()
print('准确率为%.3f'%(correct/total*100))
if __name__=='__main__':
for epoch in range(3):
train(epoch)
test()
运行结果: