weight decay

Author: lucidrains

improving_transformers_world_model/agent.py

@@ -396,6 +396,8 @@ def __init__(
         optim_klass = AdoptAtan2,
         actor_lr = 1e-4,
         critic_lr = 1e-4,
+        actor_weight_decay = 1e-3,
+        critic_weight_decay = 1e-3,
         max_grad_norm = 0.5,
         actor_optim_kwargs: dict = dict(),
         critic_optim_kwargs: dict = dict(),
@@ -429,8 +431,8 @@ def __init__(
 
         self.max_grad_norm = max_grad_norm
 
-        self.actor_optim = optim_klass((*actor.parameters(), *impala.parameters()), lr = actor_lr, **actor_optim_kwargs)
-        self.critic_optim = optim_klass((*critic.parameters(), *impala.parameters()), lr = actor_lr, **actor_optim_kwargs)
+        self.actor_optim = optim_klass((*actor.parameters(), *impala.parameters()), lr = actor_lr, weight_decay = actor_weight_decay, **actor_optim_kwargs)
+        self.critic_optim = optim_klass((*critic.parameters(), *impala.parameters()), lr = critic_lr, weight_decay = critic_weight_decay, **critic_optim_kwargs)
 
         # use a batch norm for standardizing the target - section A.1.2 in paper
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "improving-transformers-world-model"
-version = "0.0.53"
+version = "0.0.54"
 description = "Improving Transformers World Model for RL"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }