As per Understanding Posterior Collapse in Generative Latent Variable Modelsdiagonal_normal PPCA flavor (linear VAE) is identical to the global maximum for the marginal log-likelihood of pPCA. That means we can remove multivariate_normal PPCA flavor and leave only diagonal_normal (optionally rename it to linear_vae).
ordabayevy
commented
1 year ago
As per Understanding Posterior Collapse in Generative Latent Variable Models
diagonal_normalPPCA flavor (linear VAE) is identical to the global maximum for the marginal log-likelihood of pPCA. That means we can removemultivariate_normalPPCA flavor and leave onlydiagonal_normal(optionally rename it tolinear_vae).