hongshaorou/dreamer4
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

swap for discrete kl div, thanks to Dirk for pointing this out on the discord

Browse Source
This commit is contained in:
lucidrains 2025-10-29 11:54:18 -07:00
parent cb54121ace
commit c4a3cb09d5
2 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
dreamer4/dreamer4.py
View File

@ -2580,7 +2580,9 @@ class DynamicsWorldModel(Module):
if self.pmpo_kl_div_loss_weight > 0.:
new_unembedded_actions = self.action_embedder.unembed(policy_embed, pred_head_index = 0)
discrete_kl_div, continuous_kl_div = self.action_embedder.kl_div(new_unembedded_actions, old_action_unembeds)
# mentioned that the "reverse direction for the prior KL" was used
discrete_kl_div, continuous_kl_div = self.action_embedder.kl_div(old_action_unembeds, new_unembedded_actions)
# accumulate discrete and continuous kl div

2
pyproject.toml
View File

@ -1,6 +1,6 @@
[project]
name = "dreamer4"
version = "0.0.96"
version = "0.0.98"
description = "Dreamer 4"
authors = [
{ name = "Phil Wang", email = "lucidrains@gmail.com" }