rxnn 0.2.43__tar.gz → 0.2.45__tar.gz

{rxnn-0.2.43 → rxnn-0.2.45}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.2.43
+Version: 0.2.45
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.2.43 → rxnn-0.2.45}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rxnn"
-version = "0.2.43"
+version = "0.2.45"
 description = "RxNN: Reactive Neural Networks Platform"
 
 license = "Apache-2.0"

{rxnn-0.2.43 → rxnn-0.2.45}/src/rxnn/memory/stm.py

@@ -62,6 +62,9 @@ class ShortTermMemory(nn.Module):
     def reset(self, init_type: str = None):
         self.memory = self._init_tensor(init_type).to(self.memory.device)
 
+    def clone_detach_reset(self):
+        self.memory = self.memory.detach().clone()
+
     def resize(self, new_stm_size: int, init_type: str = None):
         self.stm_size = new_stm_size
         delattr(self, 'memory')

{rxnn-0.2.43 → rxnn-0.2.45}/src/rxnn/rxt/models.py

@@ -301,6 +301,9 @@ class RxTAlphaMemoryAttention(nn.Module, PyTorchModelHubMixin, license="apache-2
     def reset_memory(self, init_type: str = None):
         self.model.stm.reset(init_type)
 
+    def clone_reset_memory(self):
+        self.model.stm.clone_detach_reset()
+
     def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None) -> torch.Tensor:
         return self.model(x, attention_mask=attention_mask)
 

{rxnn-0.2.43 → rxnn-0.2.45}/src/rxnn/training/models.py

@@ -146,6 +146,9 @@ class MrlActorModel(nn.Module):
     def reset_memory(self):
         self.memory_attention.reset_memory()
 
+    def clone_reset_memory(self):
+        self.memory_attention.clone_reset_memory()
+
     def memory_parameters(self) -> list[nn.Parameter]:
         return list(set(
             self.encoder.memory_parameters() +

{rxnn-0.2.43 → rxnn-0.2.45}/src/rxnn/training/mrl.py

@@ -91,6 +91,7 @@ class MrlTrajectoryEpisode(TypedDict):
     reset_stm: bool
     steps: list[MrlTrajectoryStep]
 
+OptimField: TypeAlias = Literal['lr', 'critic_lr', 'weight_decay', 'critic_weight_decay', 'separate_memory_lr', 'memory_lr']
 
 class MRLTrainer:
     def __init__(
@@ -646,6 +647,8 @@ class MRLTrainer:
                 step_critic_values = episode_critic_values[step_idx]
                 step_advantages = episode_advantages[step_idx]
 
+                self.actor.clone_reset_memory()
+
                 # 7. In memory aware critic version, encode and update STM before critic update, to include its gradients in critic loss too
                 if self.memory_aware_critic:
                     self.encode_and_update_stm(query, answer)
@@ -979,8 +982,19 @@ class MRLTrainer:
         self.reward = config.get('reward_model', self.shared_reward_model)  # MRL Reward Model for curriculum stage
         self.update_epochs = config.get('update_epochs', self.shared_update_epochs)  # Internal update epochs
         self.freeze_embeddings = config.get('freeze_embeddings', self.shared_freeze_embeddings)
-        if config['lr'] is not None or config['critic_lr'] is not None or config['weight_decay'] is not None or config[
-            'critic_weight_decay'] is not None or (config['separate_memory_lr'] and config['memory_lr'] is not None):
+
+
+
+        def has_param(field: OptimField) -> bool:
+            return field in config and config[field] is not None
+
+        optim_params: list[OptimField] = ['lr', 'critic_lr', 'weight_decay', 'critic_weight_decay']
+
+        has_any_optim_param = any(
+            has_param(field) for field in optim_params
+        ) or (has_param('separate_memory_lr') and config['separate_memory_lr'] and has_param('memory_lr'))
+
+        if has_any_optim_param:
             if config.get('separate_memory_lr', False):
                 self.optim_config = {
                     'lr': config.get('lr', self.base_optim_config['lr']),