def _get_latent_loss(self, mu_lat, logs_lat, mu_next_obs, sigma_next_obs,
targ_next_obs):
log_prob = log_gaussian_prob(mu_next_obs, sigma_next_obs,
targ_next_obs) # P(s' | s, a, z)
kl = kl_regularization(mu_lat, logs_lat) # KL(q_phi || N(0,1))
return -log_prob + self._beta_kl * kl, log_prob, kl
def loss_fn(inputs: d, outputs: d, model_outputs: d) -> torch.Tensor:
return -log_gaussian_prob(model_outputs.next_obs,
model_outputs.next_obs_sigma,
outputs.next_obs.unsqueeze(1))