CNN 进阶
GoogleNet
可以看到，网络中又有网络（Inception Module）
Inception Model
为什么里面需要一个1×1的卷积呢？
通过对三个通道的一次1×1卷积运算，相当于把三个通道的信息融合在了一个通道中，类似高中时期，最后的排名是将所有成绩相加再来进行比较的
- 如果直接对192@28×28进行一个5×5卷积，最终的计算量为： 卷积核的5²×原尺寸28²×原通道数192×现通道数32=120422400
- 如果先进行一次1×1卷积，压缩通道数量，再来进行5×5卷积，则最终的计算量为： 卷积核的1²×原尺寸28²×原通道数192×现通道数16+卷积核的5²×原尺寸28²×原通道数16×现通道数32=12433648
- （12433648/120422400）*100%=10.325% 相当于计算量减少为为原来的1/10，也就是说运算时间也减少1/10，原来跑10小时的，现在只需要跑1小时，大大缩短时间成本
相应代码
- self.xxx是在__init__()里面的代码，剩下的是在forward()里面的代码
- 通过设置Padding来避免w、h的减小
相应的拼接代码：
outputs=[branch1x1,branch5x5,branch3x3,branch_pool1]
return torch.cat(outputs,dim=1)
其中的dim=1，指的是从通道维度进行拼接操作，4个维度分别为：p:patch、c:通道、w:宽度、h:高度 分布对应dim=0、1、2、3
代码
import matplotlib.pyplot as plt
import torch
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision import transforms #将w（宽）×h（高）×c（channel）转换成c×w×h，即把通道提到最前面
batch_size = 64
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,)) #mnist数据集的均值，前人已经算好的，直接用这两个数就行
])
train_dataset = datasets.MNIST(root='./dataset/minist/',
train=True,
download=True,
transform=transform)
train_loader = DataLoader(train_dataset,
shuffle=True,
batch_size=batch_size)
test_dataset = datasets.MNIST(root='./dataset/mnist/',
train=False,
download=True,
transform=transform)
test_loader = DataLoader(test_dataset,
shuffle=False,
batch_size=batch_size)
#---------------------------------------------------以下为CNN-------------------------------------------------------------------------#
class InceptionA(torch.nn.Module):
def __init__(self, in_channels):
super(InceptionA, self).__init__()
self.branch1x1 = torch.nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch5x5_1 = torch.nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch5x5_2 = torch.nn.Conv2d(16, 24, kernel_size=5, padding=2)
self.branch3x3_1 = torch.nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch3x3_2 = torch.nn.Conv2d(16, 24, kernel_size=3, padding=1)
self.branch3x3_3 = torch.nn.Conv2d(24, 24, kernel_size=3, padding=1)
self.branch_pool = torch.nn.Conv2d(in_channels, 24, kernel_size=1)
def forward(self, x):
branch1x1 = self.branch1x1(x) #1×1
branch5x5 = self.branch5x5_1(x) #5×5
branch5x5 = self.branch5x5_2(branch5x5)
branch3x3 = self.branch3x3_1(x)
branch3x3 = self.branch3x3_2(branch3x3)
branch3x3 = self.branch3x3_3(branch3x3)
branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1) #先平均池化
branch_pool = self.branch_pool(branch_pool) #再1×1卷积
#将4个branch的结果按通道拼接
outputs = [branch1x1, branch5x5, branch3x3, branch_pool]
return torch.cat(outputs, dim=1)
class Net(torch.nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = torch.nn.Conv2d(1, 10, kernel_size=5)
self.conv2 = torch.nn.Conv2d(88, 20, kernel_size=5)
self.incep1 = InceptionA(in_channels=10)
self.incep2 = InceptionA(in_channels=20)
self.mp = torch.nn.MaxPool2d(2)
self.fc = torch.nn.Linear(1408, 10)
def forward(self, x):
in_size = x.size(0)
x = F.relu(self.mp(self.conv1(x)))
x = self.incep1(x)
x = F.relu(self.mp(self.conv2(x)))
x = self.incep2(x)
x = x.view(in_size, -1)
x = self.fc(x)
return x
model = Net()
#**********************以下为调用GPU改进处*******************************#
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
#**********************以上为调用GPU改进处*******************************#
print(device)
#---------------------------------------------------以上为CNN-------------------------------------------------------------------------#
criterion = torch.nn.CrossEntropyLoss() #交叉熵
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5) #momentum：动量，有助于更快收敛，也有助于跳出局部最优
def train(epoch):
running_loss = 0.0
for batch_idx, data in enumerate(train_loader, 0):
inputs, target = data
#**********************以下为调用GPU改进处*******************************#
inputs, target = inputs.to(device), target.to(device)
#**********************以上为调用GPU改进处*******************************#
optimizer.zero_grad()
#前馈+反馈+更新
outputs = model(inputs)
loss = criterion(outputs, target)
loss.backward()
optimizer.step()
running_loss += loss.item()
if batch_idx % 300 == 299:
print('[%d, %5d] loss: %.3f' % (epoch + 1, batch_idx + 1, running_loss / 300))
running_loss = 0.0
def test():
correct = 0
total = 0
with torch.no_grad(): #test不需要算梯度
for data in test_loader:
images, labels = data
#**********************以下为调用GPU改进处*******************************#
images, labels = images.to(device), labels.to(device)
#**********************以上为调用GPU改进处*******************************#
outputs = model(images)
_, predicted = torch.max(outputs.data, dim=1) # 用_表示一个不重要的值，后面也没用到，就只占个位置，dim=1表示横向求max
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy on test set: %d %% [%d/%d]' % (100 * correct / total, correct, total))
return correct / total
epoch_list = []
accu_list = []
if __name__ == '__main__':
for epoch in range(10):
train(epoch)
accu = test()
epoch_list.append(epoch)
accu_list.append(accu)
plt.plot(epoch_list, accu_list, 'o-')
plt.xticks(range(10))
plt.xlabel('epoch')
plt.ylabel('accuracy')
plt.grid(alpha=0.4)
plt.show()
cuda:0
[1, 300] loss: 0.966
[1, 600] loss: 0.217
[1, 900] loss: 0.157
Accuracy on test set: 96 % [9669/10000]
[2, 300] loss: 0.125
[2, 600] loss: 0.107
[2, 900] loss: 0.095
Accuracy on test set: 97 % [9758/10000]
[3, 300] loss: 0.087
[3, 600] loss: 0.083
[3, 900] loss: 0.077
Accuracy on test set: 98 % [9802/10000]
[4, 300] loss: 0.069
[4, 600] loss: 0.067
[4, 900] loss: 0.070
Accuracy on test set: 98 % [9846/10000]
[5, 300] loss: 0.061
[5, 600] loss: 0.061
[5, 900] loss: 0.060
Accuracy on test set: 98 % [9830/10000]
[6, 300] loss: 0.055
[6, 600] loss: 0.055
[6, 900] loss: 0.052
Accuracy on test set: 98 % [9845/10000]
[7, 300] loss: 0.050
[7, 600] loss: 0.045
[7, 900] loss: 0.050
Accuracy on test set: 98 % [9835/10000]
[8, 300] loss: 0.043
[8, 600] loss: 0.047
[8, 900] loss: 0.046
Accuracy on test set: 98 % [9864/10000]
[9, 300] loss: 0.042
[9, 600] loss: 0.041
[9, 900] loss: 0.042
Accuracy on test set: 98 % [9867/10000]
[10, 300] loss: 0.039
[10, 600] loss: 0.038
[10, 900] loss: 0.042
Accuracy on test set: 98 % [9872/10000]
Residual-Net网络
- 通过再原来的F(x)基础上添加一个x，能够有效避免
梯度消失问题
ResidualNet代码实现
import matplotlib.pyplot as plt
import torch
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision import transforms #将w（宽）×h（高）×c（channel）转换成c×w×h，即把通道提到最前面
batch_size = 64
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,)) #mnist数据集的均值，前人已经算好的，直接用这两个数就行
])
train_dataset = datasets.MNIST(root='./dataset/minist/',
train=True,
download=True,
transform=transform)
train_loader = DataLoader(train_dataset,
shuffle=True,
batch_size=batch_size)
test_dataset = datasets.MNIST(root='./dataset/mnist/',
train=False,
download=True,
transform=transform)
test_loader = DataLoader(test_dataset,
shuffle=False,
batch_size=batch_size)
#---------------------------------------------------以下为CNN-------------------------------------------------------------------------#
class ResidualBlock(torch.nn.Module):
def __init__(self, channels):
super(ResidualBlock, self).__init__()
self.channels = channels
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)
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)
self.mp = torch.nn.MaxPool2d(2)
self.rblock1=ResidualBlock(16)
self.rblock2=ResidualBlock(32)
self.fc = torch.nn.Linear(512, 10)
def forward(self, x):
in_size = x.size(0)
x = F.relu(self.mp(self.conv1(x)))
x = self.rblock1(x)
x = F.relu(self.mp(self.conv2(x)))
x = self.rblock2(x)
x = x.view(in_size, -1)
x = self.fc(x)
return x
model = Net()
#**********************以下为调用GPU改进处*******************************#
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
#**********************以上为调用GPU改进处*******************************#
print(device)
#---------------------------------------------------以上为CNN-------------------------------------------------------------------------#
criterion = torch.nn.CrossEntropyLoss() #交叉熵
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5) #momentum：动量，有助于更快收敛，也有助于跳出局部最优
def train(epoch):
running_loss = 0.0
for batch_idx, data in enumerate(train_loader, 0):
inputs, target = data
#**********************以下为调用GPU改进处*******************************#
inputs, target = inputs.to(device), target.to(device)
#**********************以上为调用GPU改进处*******************************#
optimizer.zero_grad()
#前馈+反馈+更新
outputs = model(inputs)
loss = criterion(outputs, target)
loss.backward()
optimizer.step()
running_loss += loss.item()
if batch_idx % 300 == 299:
print('[%d, %5d] loss: %.3f' % (epoch + 1, batch_idx + 1, running_loss / 300))
running_loss = 0.0
def test():
correct = 0
total = 0
with torch.no_grad(): #test不需要算梯度
for data in test_loader:
images, labels = data
#**********************以下为调用GPU改进处*******************************#
images, labels = images.to(device), labels.to(device)
#**********************以上为调用GPU改进处*******************************#
outputs = model(images)
_, predicted = torch.max(outputs.data, dim=1) # 用_表示一个不重要的值，后面也没用到，就只占个位置，dim=1表示横向求max
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy on test set: %d %% [%d/%d]' % (100 * correct / total, correct, total))
return correct / total
epoch_list = []
accu_list = []
if __name__ == '__main__':
for epoch in range(10):
train(epoch)
accu = test()
epoch_list.append(epoch)
accu_list.append(accu)
plt.plot(epoch_list, accu_list, 'o-')
plt.xticks(range(10))
plt.xlabel('epoch')
plt.ylabel('accuracy')
plt.grid(alpha=0.4)
plt.show()
cuda:0
[1, 300] loss: 0.529
[1, 600] loss: 0.155
[1, 900] loss: 0.114
Accuracy on test set: 96 % [9677/10000]
[2, 300] loss: 0.091
[2, 600] loss: 0.072
[2, 900] loss: 0.070
Accuracy on test set: 98 % [9841/10000]
[3, 300] loss: 0.059
[3, 600] loss: 0.062
[3, 900] loss: 0.049
Accuracy on test set: 98 % [9870/10000]
[4, 300] loss: 0.045
[4, 600] loss: 0.044
[4, 900] loss: 0.049
Accuracy on test set: 98 % [9847/10000]
[5, 300] loss: 0.043
[5, 600] loss: 0.039
[5, 900] loss: 0.038
Accuracy on test set: 98 % [9886/10000]
[6, 300] loss: 0.034
[6, 600] loss: 0.035
[6, 900] loss: 0.032
Accuracy on test set: 98 % [9886/10000]
[7, 300] loss: 0.030
[7, 600] loss: 0.031
[7, 900] loss: 0.029
Accuracy on test set: 98 % [9870/10000]
[8, 300] loss: 0.026
[8, 600] loss: 0.025
[8, 900] loss: 0.029
Accuracy on test set: 98 % [9877/10000]
[9, 300] loss: 0.023
[9, 600] loss: 0.025
[9, 900] loss: 0.023
Accuracy on test set: 98 % [9865/10000]
[10, 300] loss: 0.018
[10, 600] loss: 0.023
[10, 900] loss: 0.023
Accuracy on test set: 98 % [9873/10000]
其他类型的网络结构
接下来的学习方向
- 理论《深度学习》（花书）
- 阅读Pytorch文档（至少通读一遍）：看看支持的运算、有哪些层
- 复现经典工作：读代码–→训练、测试、数据读取架构、损失函数构建–→根据论文讲的东西自己去完成，读代码和写代码相互迭代
- 选择特定领域，大量阅读论文，看看大家都用了什么技巧，积累一些模块，扩充视野，看多了就可以想创新点
完结谢辞
整个深度学习基础课程到这里结束，由于目前做的是CV方向，后面的循环神经网络暂时用不上，就先不看了。
在此感谢B站刘二大人的课程给予我指导,也给我后续的进一步学习指明了方向
577
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
