cGANs with Projection Discriminator

系列文章目录

一 Conditional Generative Adversarial Nets
二 cGANs with Projection Discriminator
三 Conditional Image Synthesis with Auxiliary Classifier GANs
四 InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

---

基于映射鉴别器的CGAN

模型中，判别器（Discriminator）不是通过将条件信息简单地与特征向量拼接（concatenate）来使用条件信息，而是采用一种基于投影的方式，这种方式更加尊重条件信息在底层概率模型中的作用。
判别器的构建是受到概率模型假设的启发，其中条件变量 y 给定 x 的分布是离散的或单峰连续分布。这种模型假设在许多实际应用中很常见，包括类条件图像生成和超分辨率。通过这种假设，可以形成一个需要在嵌入的条件向量 y 和特征向量之间进行内积的判别器结构。

代码实现

class DiscriminatorPCGAN(nn.Module):

    def __init__(self, x_dim, c_dim, dim=96, norm='none', weight_norm='spectral_norm'):
        super(DiscriminatorPCGAN, self).__init__()

        norm_fn = _get_norm_fn_2d(norm)
        weight_norm_fn = _get_weight_norm_fn(weight_norm)

        def conv_norm_lrelu(in_dim, out_dim, kernel_size=3, stride=1, padding=1):
            return nn.Sequential(
                weight_norm_fn(nn.Conv2d(in_dim, out_dim, kernel_size, stride, padding)),
                norm_fn(out_dim),
                nn.LeakyReLU(0.2)
            )

        self.ls = nn.Sequential(  # (N, x_dim, 32, 32)
            conv_norm_lrelu(x_dim, dim),
            conv_norm_lrelu(dim, dim),
            conv_norm_lrelu(dim, dim, stride=2),  # (N, dim , 16, 16)

            conv_norm_lrelu(dim, dim * 2),
            conv_norm_lrelu(dim * 2, dim * 2),
            conv_norm_lrelu(dim * 2, dim * 2, stride=2),  # (N, dim*2, 8, 8)

            conv_norm_lrelu(dim * 2, dim * 2, kernel_size=3, stride=1, padding=0),
            conv_norm_lrelu(dim * 2, dim * 2, kernel_size=1, stride=1, padding=0),
            conv_norm_lrelu(dim * 2, dim * 2, kernel_size=1, stride=1, padding=0),  # (N, dim*2, 6, 6)

            nn.AvgPool2d(kernel_size=6),  # (N, dim*2, 1, 1)
            torchlib.Reshape(-1, dim * 2),  # (N, dim*2)
        )

        self.l_logit = weight_norm_fn(nn.Linear(dim * 2, 1))  # (N, 1)
        self.l_projection = weight_norm_fn(nn.Linear(dim * 2, c_dim))  # (N, c_dim)

    def forward(self, x, c):
        # x: (N, x_dim, 32, 32), c: (N, c_dim)
        feat = self.ls(x)
        logit = self.l_logit(feat)
        # 做一个线性编码
        embed = (self.l_projection(feat) * c).mean(1, keepdim=True)
        logit += embed
        return logit

CGAN参考文章