evolutionary-policy-optimization 0.0.44__py3-none-any.whl → 0.0.45__py3-none-any.whl

evolutionary_policy_optimization/epo.py

@@ -628,6 +628,7 @@ class Agent(Module):
         latent_lr = 1e-5,
         use_critic_ema = True,
         critic_ema_beta = 0.99,
+        max_grad_norm = 0.5,
         batch_size = 16,
         calc_gae_kwargs: dict = dict(
             use_accelerated = False,
@@ -662,7 +663,11 @@ class Agent(Module):
         self.actor_loss = partial(actor_loss, **actor_loss_kwargs)
         self.calc_gae = partial(calc_generalized_advantage_estimate, **calc_gae_kwargs)
 
+        # learning hparams
+
         self.batch_size = batch_size
+        self.max_grad_norm = max_grad_norm
+        self.has_grad_clip = exists(max_grad_norm)
 
         # optimizers
 
@@ -821,6 +826,9 @@ class Agent(Module):
 
                 actor_loss.backward()
 
+                if exists(self.has_grad_clip):
+                    nn.utils.clip_grad_norm_(self.actor.parameters(), self.max_grad_norm)
+
                 self.actor_optim.step()
                 self.actor_optim.zero_grad()
 
@@ -834,6 +842,9 @@ class Agent(Module):
 
                 critic_loss.backward()
 
+                if exists(self.has_grad_clip):
+                    nn.utils.clip_grad_norm_(self.critic.parameters(), self.max_grad_norm)
+
                 self.critic_optim.step()
                 self.critic_optim.zero_grad()
 

{evolutionary_policy_optimization-0.0.44.dist-info → evolutionary_policy_optimization-0.0.45.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.44
+Version: 0.0.45
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization

{evolutionary_policy_optimization-0.0.44.dist-info → evolutionary_policy_optimization-0.0.45.dist-info}/RECORD

@@ -1,8 +1,8 @@
 evolutionary_policy_optimization/__init__.py,sha256=0q0aBuFgWi06MLMD8FiHzBYQ3_W4LYWrwmCtF3u5H2A,201
-evolutionary_policy_optimization/epo.py,sha256=8ZyW21nbEokTii7alduJVDfPLhRLsLCjttWni6-BNFE,31072
+evolutionary_policy_optimization/epo.py,sha256=NA-d7pWDJyQYULDIVB25lnbpTbwMyxc1U8RU8XGTNts,31500
 evolutionary_policy_optimization/experimental.py,sha256=9FrJGviLESlYysHI3i83efT9g2ZB9ha4u3K9HXN98_w,1100
 evolutionary_policy_optimization/mock_env.py,sha256=6AIc4mwL_C6JkAxwESJgCLxXHMzCAu2FcffVg3HkSm0,920
-evolutionary_policy_optimization-0.0.44.dist-info/METADATA,sha256=8tJisD2xu5Q7udBrmFnfy6oxl_RQrCF9DYrE6yHkP2M,6213
-evolutionary_policy_optimization-0.0.44.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.0.44.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.0.44.dist-info/RECORD,,
+evolutionary_policy_optimization-0.0.45.dist-info/METADATA,sha256=3jXsZBoltrWQJk2Yd6zu1KmCcl9AEuhxES_mX8E1lAk,6213
+evolutionary_policy_optimization-0.0.45.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.0.45.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.0.45.dist-info/RECORD,,