Upload continuous_learning_session.py

continuous_learning_session.py

@@ -102,8 +102,8 @@ class ContinuousLearningSession:
              # If it's a python list
              adapter_params = [p for layer in self.model.flux_layers for p in layer.parameters()]
              
-        # Switch to SGD to save memory (Adam uses 2x states, causing OOM on T4)
-        self.optimizer = optim.SGD(controller_params + adapter_params, lr=1e-3, momentum=0.9)
+        # Switch back to Adam (Better convergence, relying on GC/Env for memory safety)
+        self.optimizer = optim.Adam(controller_params + adapter_params, lr=1e-4)
         
         self.model.train()  # Enable gradients for Controller/Adapters