ELBO loss and exact gradients for fixed reparameterisation noise

ELBO loss and exact gradients for fixed reparameterisation noise

Usage

morie_geron_vae_loss_and_grads(X, params, eps, beta = 1)

Arguments

X

Data (n, d).

params

list(Wmu, bmu, Wlv, blv, Wd, bd).

eps

Matrix of standard-normal noise (n, latent_dim).

beta

KL weight.

Value

list(loss, recon, kl, grads) with grads a same-shaped list as params.

Examples

set.seed(5)
X <- matrix(rnorm(20), 5, 4)
params <- list(matrix(rnorm(8, 0, 0.3), 4, 2), rep(0, 2),
               matrix(rnorm(8, 0, 0.3), 4, 2), rep(0, 2),
               matrix(rnorm(8, 0, 0.3), 2, 4), rep(0, 4))
eps <- matrix(rnorm(10), 5, 2)
r <- morie_geron_vae_loss_and_grads(X, params, eps)
c(r$loss, r$recon, r$kl)
#> [1] 2.1232572 1.3996456 0.7236116