Is there a reason for KL divergence to increase?

1

I am using Variational Autoencoder(VAE) and KL divergence with recon loss as criteria.
Interestingly after 100 epochs or so, KL divergence starts to increase exponentially.

image

If I understand correctly shouldn’t it decrease rather than an increase in such a manner?
The code I am using as below:

    def kl_divergence(z, mu, std):
        p = torch.distributions.Normal(torch.zeros_like(mu), torch.ones_like(std))
        q = torch.distributions.Normal(mu, std)
        log_qzx = q.log_prob(z)
        log_pz = p.log_prob(z)
        kl = (log_qzx - log_pz)
        kl = kl.sum(-1)
        return kl

where z is 1D spectrum data.

2

The same also observed using `torch.distributions.kl_divergence’

    def kl_divergence(self, z, mu, std):
        p = torch.distributions.Normal(torch.zeros_like(mu), torch.ones_like(std))
        q = torch.distributions.Normal(mu, std)
        kl_t = torch.distributions.kl_divergence(q, p).sum(-1)
        return kl_t

image