CGAN：条件生成的开山之作

import torchimport torch.nn as nnimport torch.optim as optim torch.manual_seed(42)  def build_labeled_data(num_samples=1600, num_classes=10):    """构造带类别标签的简单图案数据集：每张图8x8，标签k的图    在两个固定位置点亮像素，方便观察CGAN能否按类别生成对应图案。"""    data, labels = [], []    for i in range(num_samples):        label = i % num_classes        image = torch.zeros(8 * 8)        image[label] = 1.0        image[(label + 3) % num_classes] = 1.0        data.append(image)        labels.append(label)    data = torch.stack(data).float()    labels = torch.tensor(labels, dtype=torch.long)    return data, labels  class Generator(nn.Module):    """生成器：输入(噪声z, 标签y)，输出8x8图像"""     def __init__(self, noise_dim, num_classes):        super().__init__()        self.net = nn.Sequential(            nn.Linear(noise_dim + num_classes, 64),            nn.ReLU(),            nn.Linear(64, 64),            nn.ReLU(),            nn.Linear(64, 64),            nn.ReLU(),            nn.Linear(64, 8 * 8),            nn.Sigmoid(),        )     def forward(self, z, y):        return self.net(torch.cat([z, y], dim=1))  class Discriminator(nn.Module):    """判别器：输入(图像x, 标签y)，输出真假分数"""     def __init__(self, num_classes):        super().__init__()        self.net = nn.Sequential(            nn.Linear(8 * 8 + num_classes, 64),            nn.ReLU(),            nn.Linear(64, 32),            nn.ReLU(),            nn.Linear(32, 1),        )     def forward(self, x, y):        return self.net(torch.cat([x, y], dim=1))  def train_cgan(epochs=6, batch_size=64, noise_dim=16):    data, labels = build_labeled_data()    num_classes = labels.max().item() + 1    G = Generator(noise_dim, num_classes)    D = Discriminator(num_classes)    g_opt = optim.Adam(G.parameters(), lr=1e-3)    d_opt = optim.Adam(D.parameters(), lr=1e-3)    loss_fn = nn.BCEWithLogitsLoss()    n = len(data)     for epoch in range(epochs):        perm = torch.randperm(n)        epoch_d_loss, epoch_g_loss = 0.0, 0.0        for i in range(0, n, batch_size):            idx = perm[i:i + batch_size]            x_real = data[idx]            y_labels = labels[idx]            y = torch.nn.functional.one_hot(y_labels, num_classes).float()             z = torch.randn(len(idx), noise_dim)            x_fake = G(z, y)             d_opt.zero_grad()            real_logits = D(x_real, y)            fake_logits = D(x_fake.detach(), y)            d_loss = loss_fn(real_logits, torch.ones_like(real_logits)) \                   + loss_fn(fake_logits, torch.zeros_like(fake_logits))            d_loss.backward()            d_opt.step()             g_opt.zero_grad()            fake_logits = D(x_fake, y)            g_loss = loss_fn(fake_logits, torch.ones_like(fake_logits))            g_loss.backward()            g_opt.step()             epoch_d_loss += d_loss.item() * len(idx)            epoch_g_loss += g_loss.item() * len(idx)         print(f"epoch {epoch+1}/{epochs}  d_loss={epoch_d_loss/n:.4f}  "              f"g_loss={epoch_g_loss/n:.4f}")     return G  def demo_generate(G, noise_dim=16, num_classes=10):    """按指定标签生成图案并打印"""    for label in [0, 3, 7, 9]:        y = torch.nn.functional.one_hot(            torch.tensor([label]), num_classes).float()        z = torch.randn(1, noise_dim)        with torch.no_grad():            img = G(z, y).view(8, 8)        print(f"标签 {label} 生成的图案：")        print((img > 0.5).int().numpy())  if __name__ == "__main__":    G = train_cgan(epochs=6)    demo_generate(G)