v0.3.4 kl weight
This version adds a new feature to the CLI, the --kl-weight
parameter. When --kl-weight
is specified, the log likelihood and kl divergence terms in the loss function are reduced by taking the mean rather than the sum. The ELBO is then computed as
ELBO = mean(log_likelihood) + kl_weight * mean(kl_div)
This feature will be useful in making comparisons across isomorphous data sets of different sizes.