dodox/metastable-baselines
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Why does KL double free?

Browse Source
This commit is contained in:
Dominik Moritz Roth 2022-08-07 18:04:40 +02:00
parent 75d73049b4
commit 12e422aec7
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
metastable_baselines/ppo/policies.py
View File

@ -307,7 +307,10 @@ class ActorCriticPolicy(BasePolicy):
return self.action_dist.proba_distribution(mean_actions, self.log_std, latent_pi) return self.action_dist.proba_distribution(mean_actions, self.log_std, latent_pi)
elif isinstance(self.action_dist, UniversalGaussianDistribution): elif isinstance(self.action_dist, UniversalGaussianDistribution):
if self.sqrt_induced_gaussian: if self.sqrt_induced_gaussian:
cov_sqrt = self.chol_net(latent_pi) chol_sqrt_cov = self.chol_net(latent_pi)
if len(chol_sqrt_cov.shape) == 2:
chol_sqrt_cov = th.diag_embed(chol_sqrt_cov)
cov_sqrt = th.bmm(chol_sqrt_cov.mT, chol_sqrt_cov)
dist = self.action_dist.proba_distribution_from_sqrt( dist = self.action_dist.proba_distribution_from_sqrt(
mean_actions, cov_sqrt, latent_pi) mean_actions, cov_sqrt, latent_pi)
mean, chol = get_mean_and_chol(dist, expand=False) mean, chol = get_mean_and_chol(dist, expand=False)