import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
设备配置
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
Hyper parameters
num_epochs = 5
num_classes = 10
batch_size = 100
learning_rate = 0.001
卷积神经网络
CIFAR10 dataset
train_dataset = torchvision.datasets.CIFAR10(root='./datasets',
train=True,
transform=transforms.ToTensor(),
download=True)
test_dataset = torchvision.datasets.CIFAR10(root='./datasets',
train=False,
transform=transforms.ToTensor())
Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
img,label=next(iter(train_dataset))
aa=nn.Conv2d(1, 16, kernel_size=5, stride=1, padding=2)(img)
aa.shape
卷积神经网络 (两个卷积层)
class ConvNet(nn.Module):
def init(self, num_classes=10):
super(ConvNet, self).init()
特征提取层1
self.layer1 = nn.Sequential(
因为核大小是5,不填充的话少两行两列,所以填充2,表示上下左右都填充两行(列)
以保持尺寸不变
nn.Conv2d(3, 16, kernel_size=5, stride=1, padding=2),
nn.BatchNorm2d(16),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2))
特征提取层2
self.layer2 = nn.Sequential(
nn.Conv2d(16, 32, kernel_size=5, stride=1, padding=2),
nn.BatchNorm2d(32),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2))
输出层
self.fc = nn.Linear(8*8*32, num_classes)
def forward(self, x):
out = self.layer1(x) # (16,16,16)
out = self.layer2(out) # (32,8,8)
out = out.reshape(out.size(0), -1)
out = self.fc(out)
return out
model = ConvNet(num_classes).to(device)
Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
Train the model
total_step = len(train_loader) # 总批次数
for epoch in range(num_epochs):
遍历数据加载器中的每个批次的数据
for i, (images, labels) in enumerate(train_loader):
images = images.to(device)
labels = labels.to(device)
前向传播,计算损失
outputs = model(images)
loss = criterion(outputs, labels)
清理之前的梯度,反向传播,根据梯度更新参数
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i+1) % 100 == 0: # 打印日志
print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'
.format(epoch+1, num_epochs, i+1, total_step, loss.item()))
评估模型
当模型被设置为评估模式时，其内部的Batch Normalization层会利用之前训练过程中积累下来的统计信息
（移动平均值和方差）来进行归一化操作，而不是依据当前输入的具体小批量数据。这种做法有利于提高模型在
新数据上的泛化能力。
model.eval()
with torch.no_grad():
correct = 0 # 用来统计模型预测正确的样本数
total = 0 # 用来统计总样本数
for images, labels in test_loader:
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Test Accuracy of the model on the 10000 test images: {} %'.format(100 * correct / total))
Save the model checkpoint
torch.save(model.state_dict(), 'model.ckpt')
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
Hyper-parameters
num_epochs = 60
batch_size = 100
learning_rate = 0.001
图像预处理模块(训练集上的增强)
transform = transforms.Compose([
transforms.Pad(4),
transforms.RandomHorizontalFlip(), # 随机水平翻转
transforms.RandomCrop(32), # 随机裁剪
transforms.ToTensor()]) # 0-1的归一化
CIFAR-10 dataset
train_dataset = torchvision.datasets.CIFAR10(root='./datasets',
train=True,
transform=transform,
download=True)
test_dataset = torchvision.datasets.CIFAR10(root='datasets',
train=False,
transform=transforms.ToTensor())
Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
3x3 卷积,因为3x3的卷积,所以上下左右都设置为填充1,bias=False不计算截距
def conv3x3(in_channels, out_channels, stride=1):
return nn.Conv2d(in_channels, out_channels, kernel_size=3,
stride=stride, padding=1, bias=False)
残差块
class ResidualBlock(nn.Module):
def init(self, in_channels, out_channels, stride=1, downsample=None):
super(ResidualBlock, self).init()
self.conv1 = conv3x3(in_channels, out_channels, stride) # 3x3卷积
self.bn1 = nn.BatchNorm2d(out_channels)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(out_channels, out_channels)
self.bn2 = nn.BatchNorm2d(out_channels)
self.downsample = downsample
def forward(self, x):
residual = x # 残差前段
out = self.conv1(x) # 标准卷积
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
如果设置了downsample的话,就做下采样
if self.downsample:
residual = self.downsample(x)
残差的两部分不会经过激活函数处理,这是为了维持卷积的线性
out += residual # 残差连接
out = self.relu(out) # 残差后经过激活函数处理
return out
m = nn.AvgPool2d((3, 2), stride=(2, 1))
input = torch.randn(20, 16, 50, 32)
output = m(input)
a=torch.randn(1,64,8,8)
b=nn.AvgPool2d(8)(a)
b.shape
ResNet
class ResNet(nn.Module):
def init(self, block, layers, num_classes=10):
super(ResNet, self).init()
self.in_channels = 16
self.conv = conv3x3(3, 16)
self.bn = nn.BatchNorm2d(16)
inplace=True：这个参数指定了是否进行就地操作。在就地操作模式下，ReLU直接修改输入张量的内容，
而不是创建一个新的输出张量。这可以节省内存，因为不需要额外的空间来存储新的输出张量。但是，需要
注意的是，就地操作可能会影响计算图的构建，特别是当你需要保留原始输入数据以供后续使用时。
self.relu = nn.ReLU(inplace=True)
self.layer1 = self.make_layer(block, 16, layers[0])
self.layer2 = self.make_layer(block, 32, layers[1], 2)
self.layer3 = self.make_layer(block, 64, layers[2], 2)
self.avg_pool = nn.AvgPool2d(8)
self.fc = nn.Linear(64, num_classes)
def make_layer(self, block, out_channels, blocks, stride=1):
downsample = None
如果步长为2(内部要下采样)或者输入数据的通道数不等于最后的输出通道数
if (stride != 1) or (self.in_channels != out_channels):
设置残差前段转换块
downsample = nn.Sequential(
conv3x3(self.in_channels, out_channels, stride=stride),
nn.BatchNorm2d(out_channels))
layers = []
添加第一个残差提取块
layers.append(block(self.in_channels, out_channels, stride, downsample))
self.in_channels = out_channels
之后的提取块(不需要传入downsample),因为这时候输入输出通道相同,步长为1
for i in range(1, blocks):
layers.append(block(out_channels, out_channels))
对列表进行拆包,构建多次提取块
return nn.Sequential(*layers)
整个前向传播可以看成嵌套复合函数
def forward(self, x):
out = self.conv(x) # 3-->16
out = self.bn(out)
out = self.relu(out)
第一个特征提取块(不会改变特征图尺寸和通道)
out = self.layer1(out)
第二个提取块(改变特征图尺寸和通道)
out = self.layer2(out) # 16-->32 (16,16)
第三个提取块(改变特征图尺寸和通道)
out = self.layer3(out) # 32-->64 (8,8)
平均池化
out = self.avg_pool(out)
变形,把样本表示成向量形式
out = out.view(out.size(0), -1)
out = self.fc(out) # 输出层
return out
第二个列表是残差块的提取深度列表
model = ResNet(ResidualBlock, [2, 2, 2]).to(device)
Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
用于更新学习率
def update_lr(optimizer, lr):
for param_group in optimizer.param_groups:
param_group['lr'] = lr
训练模型
total_step = len(train_loader) #一个轮次的训练批次
curr_lr = learning_rate # 初始学习率
for epoch in range(num_epochs):
遍历每个批次
for i, (images, labels) in enumerate(train_loader):
images = images.to(device)
labels = labels.to(device)
获取logits,计算预测和真实之间的误差
outputs = model(images)
loss = criterion(outputs, labels)
清理之前梯度,反向传播,根据梯度更新模块参数
optimizer.zero_grad()
loss.backward()
optimizer.step()
每割100次打印日志
if (i+1) % 100 == 0:
print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}"
.format(epoch+1, num_epochs, i+1, total_step, loss.item()))
根据轮次改变学习率(每20个轮次)
if (epoch+1) % 20 == 0:
curr_lr /= 3
update_lr(optimizer, curr_lr)
评估模式
model.eval()
禁用梯度是为了减少内存消耗
with torch.no_grad():
correct = 0 # 统计预测正确的样本数
total = 0 # 统计总样本数
for images, labels in test_loader:
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy of the model on the test images: {} %'.format(100 * correct / total))
Save the model checkpoint
torch.save(model.state_dict(), 'resnet.ckpt')