Pytorch学习系列（6）：深度残差网络（Deep Residual Networks）

 

导包

# 包
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms

 

设置参数和图像处理模块

# 设备配置
torch.cuda.set_device(1) # 这句用来设置pytorch在哪块GPU上运行
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# 超参数设置
num_epochs = 80
learning_rate = 0.001

# 图像预处理模块
# 先padding ，再 翻转，然后 裁剪。数据增广的手段
transform = transforms.Compose([
    transforms.Pad(4),
    transforms.RandomHorizontalFlip(),
    transforms.RandomCrop(32),
    transforms.ToTensor()])

 

CIFAR-10 数据集

#  训练数据集
train_dataset = torchvision.datasets.CIFAR10(root='./data/cifar-10',
                                             train=True, 
                                             transform=transform,
                                             download=True)

# 测试数据集
test_dataset = torchvision.datasets.CIFAR10(root='./data/cifar-10',
                                            train=False, 
                                            transform=transforms.ToTensor())

# 数据加载器
# 训练数据加载器
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
                                           batch_size=100, 
                                           shuffle=True)
# 测试数据加载器
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
                                          batch_size=100, 
                                          shuffle=False)

深度残差网络模型设计

3×3 卷积层

# 3x3 convolution
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（残差单元）（Residual block）

# Residual block
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)
        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
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        if self.downsample:
            residual = self.downsample(x)
        out += residual
        out = self.relu(out)
        return out

残差网络搭建

# 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)
        self.relu = nn.ReLU(inplace=True)
        self.layer1 = self.make_layer(block, 16, layers[0])
        self.layer2 = self.make_layer(block, 32, layers[0], 2)
        self.layer3 = self.make_layer(block, 64, layers[1], 2)
        self.avg_pool = nn.AvgPool2d(8,ceil_mode=False) #  nn.AvgPool2d需要添加参数ceil_mode=False，否则该模块无法导出为onnx格式
        self.fc = nn.Linear(64, num_classes)

    def make_layer(self, block, out_channels, blocks, stride=1):
        downsample = None
        if (stride != 1) or (self.in_channels != out_channels):
            downsample = nn.Sequential(
                conv3x3(self.in_channels, out_channels, stride=stride),
                nn.BatchNorm2d(out_channels))
        layers = []
        layers.append(block(self.in_channels, out_channels, stride, downsample)) # 残差直接映射部分
        self.in_channels = out_channels
        for i in range(1, blocks):
            layers.append(block(out_channels, out_channels))
        return nn.Sequential(*layers)

    def forward(self, x):
        out = self.conv(x)
        out = self.bn(out)
        out = self.relu(out)
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.avg_pool(out)
        out = out.view(out.size(0), -1)
        out = self.fc(out)
        return out

实例化模型

# 实例化一个残差网络模型
model = ResNet(ResidualBlock, [2, 2, 2, 2]).to(device)

# 设置损失函数和优化器
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

# 用于更新参数组中的学习率
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)
curr_lr = learning_rate
for epoch in range(num_epochs):
    for i, (images, labels) in enumerate(train_loader):
        images = images.to(device)
        labels = labels.to(device)

        # 前向传播
        outputs = model(images)
        loss = criterion(outputs, labels)

        # 反向传播和优化
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        if ((i+1) % 100 == 0) and ((epoch+1) % 5 == 0):
            print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}"
                   .format(epoch+1, num_epochs, i+1, total_step, loss.item()))

    # 学习率衰减策略
    if (epoch+1) % 20 == 0:
        curr_lr /= 3
        update_lr(optimizer, curr_lr)

Epoch [5/80], Step [100/500] Loss: 0.6635
Epoch [5/80], Step [200/500] Loss: 0.6436
Epoch [5/80], Step [300/500] Loss: 0.8391
Epoch [5/80], Step [400/500] Loss: 0.5560
Epoch [5/80], Step [500/500] Loss: 0.5242
Epoch [10/80], Step [100/500] Loss: 0.3502
Epoch [10/80], Step [200/500] Loss: 0.5507
Epoch [10/80], Step [300/500] Loss: 0.5411
Epoch [10/80], Step [400/500] Loss: 0.4270
Epoch [10/80], Step [500/500] Loss: 0.5415
Epoch [15/80], Step [100/500] Loss: 0.4049
Epoch [15/80], Step [200/500] Loss: 0.4415
Epoch [15/80], Step [300/500] Loss: 0.4004
Epoch [15/80], Step [400/500] Loss: 0.4916
Epoch [15/80], Step [500/500] Loss: 0.4242
Epoch [20/80], Step [100/500] Loss: 0.3504
Epoch [20/80], Step [200/500] Loss: 0.3224
Epoch [20/80], Step [300/500] Loss: 0.2615
Epoch [20/80], Step [400/500] Loss: 0.3431
Epoch [20/80], Step [500/500] Loss: 0.4071
Epoch [25/80], Step [100/500] Loss: 0.3055
Epoch [25/80], Step [200/500] Loss: 0.3428
Epoch [25/80], Step [300/500] Loss: 0.3447
Epoch [25/80], Step [400/500] Loss: 0.1725
Epoch [25/80], Step [500/500] Loss: 0.2822
Epoch [30/80], Step [100/500] Loss: 0.2481
Epoch [30/80], Step [200/500] Loss: 0.3132
Epoch [30/80], Step [300/500] Loss: 0.1997
Epoch [30/80], Step [400/500] Loss: 0.1979
Epoch [30/80], Step [500/500] Loss: 0.2314
Epoch [35/80], Step [100/500] Loss: 0.1789
Epoch [35/80], Step [200/500] Loss: 0.2600
Epoch [35/80], Step [300/500] Loss: 0.2443
Epoch [35/80], Step [400/500] Loss: 0.2682
Epoch [35/80], Step [500/500] Loss: 0.3727
Epoch [40/80], Step [100/500] Loss: 0.2939
Epoch [40/80], Step [200/500] Loss: 0.1615
Epoch [40/80], Step [300/500] Loss: 0.1821
Epoch [40/80], Step [400/500] Loss: 0.2461
Epoch [40/80], Step [500/500] Loss: 0.2109
Epoch [45/80], Step [100/500] Loss: 0.1834
Epoch [45/80], Step [200/500] Loss: 0.1832
Epoch [45/80], Step [300/500] Loss: 0.2229
Epoch [45/80], Step [400/500] Loss: 0.1145
Epoch [45/80], Step [500/500] Loss: 0.1227
Epoch [50/80], Step [100/500] Loss: 0.1528
Epoch [50/80], Step [200/500] Loss: 0.1875
Epoch [50/80], Step [300/500] Loss: 0.1768
Epoch [50/80], Step [400/500] Loss: 0.1521
Epoch [50/80], Step [500/500] Loss: 0.2740
Epoch [55/80], Step [100/500] Loss: 0.1263
Epoch [55/80], Step [200/500] Loss: 0.0955
Epoch [55/80], Step [300/500] Loss: 0.1172
Epoch [55/80], Step [400/500] Loss: 0.0989
Epoch [55/80], Step [500/500] Loss: 0.3041
Epoch [60/80], Step [100/500] Loss: 0.1515
Epoch [60/80], Step [200/500] Loss: 0.1332
Epoch [60/80], Step [300/500] Loss: 0.2061
Epoch [60/80], Step [400/500] Loss: 0.1126
Epoch [60/80], Step [500/500] Loss: 0.2497
Epoch [65/80], Step [100/500] Loss: 0.1328
Epoch [65/80], Step [200/500] Loss: 0.0909
Epoch [65/80], Step [300/500] Loss: 0.2579
Epoch [65/80], Step [400/500] Loss: 0.0577
Epoch [65/80], Step [500/500] Loss: 0.2322
Epoch [70/80], Step [100/500] Loss: 0.1855
Epoch [70/80], Step [200/500] Loss: 0.1698
Epoch [70/80], Step [300/500] Loss: 0.1302
Epoch [70/80], Step [400/500] Loss: 0.1371
Epoch [70/80], Step [500/500] Loss: 0.1164
Epoch [75/80], Step [100/500] Loss: 0.1389
Epoch [75/80], Step [200/500] Loss: 0.1795
Epoch [75/80], Step [300/500] Loss: 0.1745
Epoch [75/80], Step [400/500] Loss: 0.1461
Epoch [75/80], Step [500/500] Loss: 0.1978
Epoch [80/80], Step [100/500] Loss: 0.1417
Epoch [80/80], Step [200/500] Loss: 0.1094
Epoch [80/80], Step [300/500] Loss: 0.0901
Epoch [80/80], Step [400/500] Loss: 0.0944
Epoch [80/80], Step [500/500] Loss: 0.1834

 

模型测试

# 设置为评估模式
model.eval() 
# 屏蔽梯度计算
with torch.no_grad():
    correct = 0
    total = 0
    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))

Accuracy of the model on the test images: 88.48 %

 

保存模型

# 保存模型
torch.save(model.state_dict(), 'resnet.ckpt')