使用Pytorch训练two-head网络的操作

在PyTorch中，two-head网络是一种常用的网络结构，用于处理多任务学习问题。本文将提供一个完整的攻略，介绍如何使用PyTorch训练two-head网络。我们将提供两个示例，分别是使用nn.ModuleList和使用nn.Sequential。

示例1：使用nn.ModuleList

以下是一个示例，展示如何使用nn.ModuleList训练two-head网络。

1. 导入库

import torch
import torch.nn as nn
import torch.optim as optim

2. 定义模型

class TwoHeadNet(nn.Module):
    def __init__(self):
        super(TwoHeadNet, self).__init__()
        self.shared_layers = nn.Sequential(
            nn.Conv2d(3, 6, 5),
            nn.ReLU(),
            nn.MaxPool2d(2, 2),
            nn.Conv2d(6, 16, 5),
            nn.ReLU(),
            nn.MaxPool2d(2, 2),
            nn.Flatten()
        )
        self.head1_layers = nn.ModuleList([
            nn.Linear(16 * 5 * 5, 120),
            nn.ReLU(),
            nn.Linear(120, 84),
            nn.ReLU(),
            nn.Linear(84, 10)
        ])
        self.head2_layers = nn.ModuleList([
            nn.Linear(16 * 5 * 5, 120),
            nn.ReLU(),
            nn.Linear(120, 84),
            nn.ReLU(),
            nn.Linear(84, 2)
        ])

    def forward(self, x):
        x = self.shared_layers(x)
        out1 = x
        out2 = x
        for layer in self.head1_layers:
            out1 = layer(out1)
        for layer in self.head2_layers:
            out2 = layer(out2)
        return out1, out2

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = TwoHeadNet().to(device)

3. 定义损失函数和优化器

criterion1 = nn.CrossEntropyLoss()
criterion2 = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

4. 训练模型

num_epochs = 10
for epoch in range(num_epochs):
    running_loss1 = 0.0
    running_loss2 = 0.0
    for i, (inputs, labels1, labels2) in enumerate(trainloader, 0):
        inputs = inputs.to(device)
        labels1 = labels1.to(device)
        labels2 = labels2.to(device)

        optimizer.zero_grad()

        outputs1, outputs2 = model(inputs)
        loss1 = criterion1(outputs1, labels1)
        loss2 = criterion2(outputs2, labels2)
        loss = loss1 + loss2
        loss.backward()
        optimizer.step()

        running_loss1 += loss1.item()
        running_loss2 += loss2.item()
        if i % 2000 == 1999:    # 每2000个小批量数据打印一次损失值
            print('[%d, %5d] loss1: %.3f, loss2: %.3f' %
                  (epoch + 1, i + 1, running_loss1 / 2000, running_loss2 / 2000))
            running_loss1 = 0.0
            running_loss2 = 0.0

print('Finished Training')

示例2：使用nn.Sequential

以下是一个示例，展示如何使用nn.Sequential训练two-head网络。

1. 导入库

import torch
import torch.nn as nn
import torch.optim as optim

2. 定义模型

class TwoHeadNet(nn.Module):
    def __init__(self):
        super(TwoHeadNet, self).__init__()
        self.shared_layers = nn.Sequential(
            nn.Conv2d(3, 6, 5),
            nn.ReLU(),
            nn.MaxPool2d(2, 2),
            nn.Conv2d(6, 16, 5),
            nn.ReLU(),
            nn.MaxPool2d(2, 2),
            nn.Flatten()
        )
        self.head1_layers = nn.Sequential(
            nn.Linear(16 * 5 * 5, 120),
            nn.ReLU(),
            nn.Linear(120, 84),
            nn.ReLU(),
            nn.Linear(84, 10)
        )
        self.head2_layers = nn.Sequential(
            nn.Linear(16 * 5 * 5, 120),
            nn.ReLU(),
            nn.Linear(120, 84),
            nn.ReLU(),
            nn.Linear(84, 2)
        )

    def forward(self, x):
        x = self.shared_layers(x)
        out1 = self.head1_layers(x)
        out2 = self.head2_layers(x)
        return out1, out2

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = TwoHeadNet().to(device)

3. 定义损失函数和优化器

criterion1 = nn.CrossEntropyLoss()
criterion2 = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

4. 训练模型

num_epochs = 10
for epoch in range(num_epochs):
    running_loss1 = 0.0
    running_loss2 = 0.0
    for i, (inputs, labels1, labels2) in enumerate(trainloader, 0):
        inputs = inputs.to(device)
        labels1 = labels1.to(device)
        labels2 = labels2.to(device)

        optimizer.zero_grad()

        outputs1, outputs2 = model(inputs)
        loss1 = criterion1(outputs1, labels1)
        loss2 = criterion2(outputs2, labels2)
        loss = loss1 + loss2
        loss.backward()
        optimizer.step()

        running_loss1 += loss1.item()
        running_loss2 += loss2.item()
        if i % 2000 == 1999:    # 每2000个小批量数据打印一次损失值
            print('[%d, %5d] loss1: %.3f, loss2: %.3f' %
                  (epoch + 1, i + 1, running_loss1 / 2000, running_loss2 / 2000))
            running_loss1 = 0.0
            running_loss2 = 0.0

print('Finished Training')

总结

本文提供了一个完整的攻略，介绍了如何使用PyTorch训练two-head网络。我们提供了两个示例，分别是使用nn.ModuleList和使用nn.Sequential。在实现过程中，我们使用了PyTorch的nn.ModuleList和nn.Sequential，并使用了CrossEntropyLoss损失函数和SGD优化器。