第六周：生成式对抗网络

# 开始训练，一共训练total_epochs
for epoch in range(total_epochs):

    # torch.nn.Module.train() 指的是模型启用 BatchNormalization 和 Dropout
    # torch.nn.Module.eval() 指的是模型不启用 BatchNormalization 和 Dropout
    # 因此，train()一般在训练时用到， eval() 一般在测试时用到
    generator = generator.train()

    # 训练一个epoch
    for i, data in enumerate(dataloader):

        # 加载真实数据
        real_images, real_labels = data
        real_images = real_images.to(device)
        # 把对应的标签转化成 one-hot 类型
        tmp = torch.FloatTensor(real_labels.size(0), 10).zero_()
        real_labels = tmp.scatter_(dim=1, index=torch.LongTensor(real_labels.view(-1, 1)), value=1)
        real_labels = real_labels.to(device)

        # 生成数据
        # 用正态分布中采样batch_size个随机噪声
        z = torch.randn([batch_size, z_dim]).to(device)
        # 生成 batch_size 个 ont-hot 标签
        c = torch.FloatTensor(batch_size, 10).zero_()
        c = c.scatter_(dim=1, index=torch.LongTensor(np.random.choice(10, batch_size).reshape([batch_size, 1])), value=1)
        c = c.to(device)
        # 生成数据
        fake_images = generator(z,c)

        # 计算判别器损失，并优化判别器
        real_loss = bce(discriminator(real_images, real_labels), ones)
        fake_loss = bce(discriminator(fake_images.detach(), c), zeros)
        d_loss = real_loss + fake_loss

        d_optimizer.zero_grad()
        d_loss.backward()
        d_optimizer.step()

        # 计算生成器损失，并优化生成器
        g_loss = bce(discriminator(fake_images, c), ones)

        g_optimizer.zero_grad()
        g_loss.backward()
        g_optimizer.step()

    # 输出损失
    print("[Epoch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, total_epochs, d_loss.item(), g_loss.item()))

#用于生成效果图
# 生成100个随机噪声向量
fixed_z = torch.randn([100, z_dim]).to(device)
# 生成100个one_hot向量，每类10个
fixed_c = torch.FloatTensor(100, 10).zero_()
fixed_c = fixed_c.scatter_(dim=1, index=torch.LongTensor(np.array(np.arange(0, 10).tolist()*10).reshape([100, 1])), value=1)
fixed_c = fixed_c.to(device)

generator = generator.eval()
fixed_fake_images = generator(fixed_z, fixed_c)

plt.figure(figsize=(8, 8))
for j in range(10):
    for i in range(10):
        img = fixed_fake_images[j*10+i, 0, :, :].detach().cpu().numpy()
        img = img.reshape([28, 28])
        plt.subplot(10, 10, j*10+i+1)
        plt.imshow(img, ‘gray‘)