为什么VAE损失没有收敛到零？

Question

我正在使用变分自动编码器，这是我对损失函数的实现：

class VariationalAutoencoder(nn.Module):
    # ...some functions...

    def gaussian_likelihood(self, x_hat, logscale, x):
        scale = torch.exp(logscale)
        mean = x_hat
        dist = torch.distributions.Normal(mean, scale)
        # measure prob of seeing image under p(x|z)
        log_pxz = dist.log_prob(x)
        return log_pxz.sum(dim=(1, 2, 3))

    def forward(self, input):
        mu, logvar = self.encode(input)
        z = self.reparameterise(mu, logvar)
        return self.decoder(z), mu, logvar, z

    def loss_function(self, x_hat, x, mu, logvar, β=1):
        std = torch.exp(logvar / 2)
        q = torch.distributions.Normal(mu, std)
        z = q.rsample()

        # reconstruction loss
        recon_loss = self.gaussian_likelihood(x_hat, self.log_scale, x)

        # kl
        kl = self.kl_divergence(z, mu, std)

        # elbo
        elbo = (kl - recon_loss)
        elbo = elbo.mean()
        return elbo

    def kl_divergence(self, z, mu, std):
        # --------------------------
        # Monte carlo KL divergence
        # --------------------------
        # 1. define the first two probabilities (in this case Normal for both)
        p = torch.distributions.Normal(torch.zeros_like(mu), torch.ones_like(std))
        q = torch.distributions.Normal(mu, std)

        # 2. get the probabilities from the equation
        log_qzx = q.log_prob(z)
        log_pz = p.log_prob(z)

        # kl
        kl = (log_qzx - log_pz)
        kl = kl.sum(-1)

        return kl

退出解码器时我使用 Sigmoid() 函数。我这样训练模型：

for epoch in range(0, epochs + 1):
        if epoch > 0:  # test untrained net first
            model.train()
            train_loss = 0
            loop = tqdm(train_loader)
            optimizer = model.setOptimizer(model)
            for x in loop:
                x = x.to(device)
                x_hat, mu, logvar, features = model(x)
                loss = model.loss_function(x_hat, x, mu, logvar)
                train_loss += loss.item()
                optimizer.zero_grad()
                loss.backward()
                optimizer.step()
                loop.set_postfix(loss=loss)
            train_loss = train_loss /= len(train_loader.dataset)
            print(f'====> Epoch: {epoch} Average loss: {train_loss:.4f}')

损失不会稳定为零，而是变为负值（大约 -2）。如果我删除

train_loss = train_loss /= len(train_loader.dataset)

，它就完全分歧了。

怎样才能让损失收敛到零？

Answer 1

VAE 的损失可能是负的。它具有对数似然 - 可能为负。这并没有什么问题。