1、GoogleNet网络结构:(常用基础网络架构,经常拿来做主架构,在进行修改完成自己的需求)
蓝色块:卷积
红色块:池化
黄色:Softmax
绿色:拼接层
构造神经网络时候有些超参数比较难选,例如卷积核应该是3x3还是5x5…。GoogleNet就是把这些这些卷积都用,把结果放在一起,那个卷积核好用,相对于的权重就比较大,其他路线的权重不重要就很小。提供了几种候选,挑选一个最优的。
上图中画框中的是 Inception Module。除了图中的 Inception Module,还有其他种类。例如:
黄色块是 1x1的卷积核(卷积核的通道取决于输入通道(16)),可以增加或者降低通道数。
2、卷积神经网络的最大困境就是运算量太大,通过1x1的卷积核的作用直接改变通道数量(192—>16),虽然模型看上去复杂了,但是实际上运算量只有之前的十分之一。就可以继续尝试其他卷积的权重。(有时候称1x1的卷积是神经网络中的神经网络)。1x1的卷积核也可以改变通道数。
3、细化网络结构,最后将结构块拼接在一起,如下图:
4、上图网络结构中的各个支路的代码为:
import
# 分支1:卷积层构造(输入通道:in_channels,输出通道:24,卷积核大小:1)
self.branch_pool = nn.Conv2d(in_channels, 24, kernel_size=1)
# 分支1:均值池化
branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)
# 分支1:对均值池化层得到的层进行卷积
branch_pool = self.branch_pool(branch_pool)
# 分支2:卷积层构造(输入通道:in_channels,输出通道:16,卷积核大小:1)
self.branch_1x1 = nn.Conv2d(in_channels, 16, kernel_size=1)
branch_1x1 = self.branch_1x1(x)
# 分支3:卷积层构造(输入通道:in_channels,输出通道:16,卷积核大小:1)
self.branch_5x5_1 = nn.Conv2d(in_channels, 16, kernel_size=1)
# 分支3:卷积层构造(输入通道:16,输出通道:24,卷积核大小:5)
self.branch_5x5_2 = nn.Conv2d(16, 24, kernel_size=5, padding=2)
branch_5x5 = self.branch_5x5_1(x)
branch_5x5 = self.branch_5x5_2(branch_5x5)
# 分支4:卷积层构造(输入通道:in_channels,输出通道:16,卷积核大小:1)
self.branch_3x3_1 = nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch_3x3_2 = nn.Conv2d(16, 24, kernel_size=3, padding=1)
self.branch_3x3_3 = nn.Conv2d(24, 24, kernel_size=3, padding=1)
branch_3x3 = self.branch_3x3_1(x)
branch_3x3 = self.branch_3x3_2(branch_3x3)
branch_3x3 = self.branch_3x3_3(branch_3x3)
output = [branch_pool, branch_1x1, branch_5x5, branch_3x3]
return torch.cat(output, dim=1) # 沿着维度1(dim=1)将output拼接起来(维度:0,1,2,3)
5、将上面代码整合到上一篇博文(八)中,完整代码:
############## GoogleNet网络架构中的 inception module 模块为本次网络架构
import torch
from torchvision import transforms
from torchvision import datasets
from torch.utils.data import DataLoader
import torch.nn.functional as F
import torch.optim as optim
batch_size = 64
transform = transforms.Compose([
transforms.ToTensor(), # 将PIL格式图像转换成Tensor矩阵向量(维度28x28转换成1x28x28,1:为RGB通道)【 [0...255]--->[0,1] 】
transforms.Normalize((0.1307, ), (0.3081, )) # 均一化处理(均值、标准差)
])
# 训练集数据
train_dataset = datasets.MNIST(root='../dataset/mnist/',
train=True,
download=True,
transform=transform)
# 加载训练集数据
train_loader = DataLoader(train_dataset,
shuffle=True,
batch_size=batch_size)
# 测试集数据集
test_dataset = datasets.MNIST(root='../dataset/mnist/',
train=False,
download=True,
transform=transform)
# 加载测试集数据集
test_loader = DataLoader(test_dataset,
shuffle=False,
batch_size=batch_size)
class Inception_A(torch.nn.Module):
def __init__(self, in_channels):
super(Inception_A, self).__init__()
# 分支1:卷积层构造(输入通道:in_channels,输出通道:24,卷积核大小:1)
self.branch_pool = torch.nn.Conv2d(in_channels, 24, kernel_size=1)
# 分支2:卷积层构造(输入通道:in_channels,输出通道:16,卷积核大小:1)
self.branch_1x1 = torch.nn.Conv2d(in_channels, 16, kernel_size=1)
# 分支3:卷积层构造(输入通道:in_channels,输出通道:16,卷积核大小:1)
self.branch_5x5_1 = torch.nn.Conv2d(in_channels, 16, kernel_size=1)
# 分支3:卷积层构造(输入通道:16,输出通道:24,卷积核大小:5)
self.branch_5x5_2 = torch.nn.Conv2d(16, 24, kernel_size=5, padding=2)
# 分支4:卷积层构造(输入通道:in_channels,输出通道:16,卷积核大小:1)
self.branch_3x3_1 = torch.nn.Conv2d(in_channels, 16, kernel_size=1)
self.branch_3x3_2 = torch.nn.Conv2d(16, 24, kernel_size=3, padding=1)
self.branch_3x3_3 = torch.nn.Conv2d(24, 24, kernel_size=3, padding=1)
def forward(self, x):
# 分支1:均值池化
branch_pool = F.avg_pool2d(x, kernel_size=3, stride=1, padding=1)
# 分支1:对均值池化层得到的层进行卷积
branch_pool = self.branch_pool(branch_pool)
# 分支2:卷积
branch_1x1 = self.branch_1x1(x)
# 分支3:卷积
branch_5x5 = self.branch_5x5_1(x)
branch_5x5 = self.branch_5x5_2(branch_5x5)
# 分支4:卷积
branch_3x3 = self.branch_3x3_1(x)
branch_3x3 = self.branch_3x3_2(branch_3x3)
branch_3x3 = self.branch_3x3_3(branch_3x3)
output = [branch_pool, branch_1x1, branch_5x5, branch_3x3]
return torch.cat(output, dim=1) # 沿着维度1(dim=1)将output拼接起来(维度:0,1,2,3)
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 = Inception_A(in_channels=10)
self.incep2 = Inception_A(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()
# 将模型放到GPU上运行,需要加如下两行代码(训练集、测试集中的输入值、实际值也需要加载到GPU上)
# device = torch.device("cude:0" if torch.cuda.is_available() else "cpu")
# model.to(device)
###################3 构建损失函数、优化器###############################
criterion = torch.nn.CrossEntropyLoss() # 交叉熵损失
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5) # 参数优化
#####################4 循环训练 #########################
def train(epoch):
running_loss = 0.0
for batch_idx, data in enumerate(train_loader, 0):
# 准备数据(input:输入,target:实际值)
inputs, target = data
# 将输入、实际值加载到GPU
# inputs, target = inputs.to(device), target.to(device)
# 梯度清0
optimizer.zero_grad()
# 前向传播
outputs = model(inputs)
# 交叉熵损失函数计算
loss = criterion(outputs, target)
# 反向传播
loss.backward()
# 参数优化
optimizer.step()
# 累计loss
running_loss += loss.item()
# 数据集一共有batch_idx个数据,每隔300个打印一次平均损失函数值
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():
for data in test_loader:
images, labels = data
# 将输入、实际值加载到GPU中
# inputs, target = inputs.to(device), target.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, dim=1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy on test set: %d %%' % (100 * correct / total))
if __name__ == '__main__':
for epoch in range(10):
train(epoch)
test()
内容主要参考了b站up主:刘二大人。
|