### 如何使用 PyTorch 实现 CGAN 条件生成对抗网络 (CGAN) 是一种扩展版的 GAN，在其中输入到生成器和判别器中的噪声向量被附加了一个条件变量。这使得模型能够根据给定条件生成特定类型的样本。 #### 构建 CGAN 的基本框架 为了构建基于 PyTorch 的 CGAN，主要组件包括定义生成器和判别器结构以及它们各自的损失函数。下面是一个简单的实现例子： ```python import torch from torch import nn, optim class Generator(nn.Module): def __init__(self, input_dim=10, output_dim=784, condition_dim=10): super(Generator, self).__init__() self.model = nn.Sequential( nn.Linear(input_dim + condition_dim, 256), nn.ReLU(), nn.Linear(256, output_dim), nn.Tanh() ) def forward(self, z, y): inputs = torch.cat([z, y], dim=-1) return self.model(inputs) class Discriminator(nn.Module): def __init__(self, input_dim=784, condition_dim=10): super(Discriminator, self).__init__() self.model = nn.Sequential( nn.Linear(input_dim + condition_dim, 256), nn.LeakyReLU(0.2), nn.Linear(256, 1), nn.Sigmoid() ) def forward(self, x, y): inputs = torch.cat([x.view(x.size(0), -1), y], dim=-1) return self.model(inputs).squeeze() def train_cgan(generator, discriminator, dataloader, num_epochs=200, device='cuda'): generator.to(device) discriminator.to(device) optimizer_g = optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999)) optimizer_d = optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999)) criterion = nn.BCELoss().to(device) real_label = 1. fake_label = 0. for epoch in range(num_epochs): for i, data in enumerate(dataloader): # 更新 D 网络: maximize log(D(x)) + log(1 - D(G(z))) ## 训练真实数据 discriminator.zero_grad() images, conditions = data[0].to(device), data[1].to(device) batch_size = images.size(0) label = torch.full((batch_size,), real_label, dtype=torch.float, device=device) outputs = discriminator(images, conditions) errD_real = criterion(outputs, label) errD_real.backward() D_x = outputs.mean().item() ## 训练伪造数据 noise = torch.randn(batch_size, nz, device=device) fake_images = generator(noise, conditions) label.fill_(fake_label) outputs = discriminator(fake_images.detach(), conditions) errD_fake = criterion(outputs, label) errD_fake.backward() D_G_z1 = outputs.mean().item() errD = errD_real + errD_fake optimizer_d.step() # 更新 G 网络: maximize log(D(G(z))) generator.zero_grad() label.fill_(real_label) # 假标签作为真标签来更新G outputs = discriminator(fake_images, conditions) errG = criterion(outputs, label) errG.backward() D_G_z2 = outputs.mean().item() optimizer_g.step() if i % 100 == 0: print(f'[{epoch}/{num_epochs}][{i}/{len(dataloader)}] ' f'Loss_D: {errD.item():.4f} Loss_G: {errG.item():.4f}' f'D(x): {D_x:.4f} D(G(z)): {D_G_z1:.4f} / {D_G_z2:.4f}') ``` 此代码片段展示了如何创建一个基础版本的 CGAN 并对其进行训练[^3]。需要注意的是，实际应用中可能还需要调整超参数设置、优化算法以及其他细节部分以获得更好的性能表现。