fixes

enhanced_realtime_training.py

@@ -1454,9 +1454,10 @@ class EnhancedRealtimeTrainingSystem:
             model.train()
             optimizer.zero_grad()
 
-            # Convert numpy arrays to PyTorch tensors
-            features_tensor = torch.from_numpy(features).float()
-            targets_tensor = torch.from_numpy(targets).long()
+            # Convert numpy arrays to PyTorch tensors and move to device
+            device = next(model.parameters()).device
+            features_tensor = torch.from_numpy(features).float().to(device)
+            targets_tensor = torch.from_numpy(targets).long().to(device)
 
             # Ensure features_tensor has the correct shape for CNN (batch_size, channels, height, width)
             # Assuming features are flattened (batch_size, 15*20) and need to be reshaped to (batch_size, 1, 15, 20)
@@ -1471,7 +1472,21 @@ class EnhancedRealtimeTrainingSystem:
             # If the CNN expects (batch_size, channels, sequence_length)
             # features_tensor = features_tensor.view(features_tensor.shape[0], 1, 15 * 20) # Example for 1D CNN
             
-            # Let's assume the CNN expects 2D input (batch_size, flattened_features)
+            # Ensure proper shape for CNN input
+            if len(features_tensor.shape) == 2:
+                # If it's (batch_size, features), keep as is for 1D CNN
+                pass
+            elif len(features_tensor.shape) == 1:
+                # If it's (features), add batch dimension
+                features_tensor = features_tensor.unsqueeze(0)
+            else:
+                # Reshape to (batch_size, features) if needed
+                features_tensor = features_tensor.view(features_tensor.shape[0], -1)
+            
+            # Limit input size to prevent shape mismatches
+            if features_tensor.shape[1] > 1000:  # Limit to 1000 features
+                features_tensor = features_tensor[:, :1000]
+            
             outputs = model(features_tensor)
             
             loss = criterion(outputs, targets_tensor)
@@ -1857,12 +1872,17 @@ class EnhancedRealtimeTrainingSystem:
                 and self.orchestrator.rl_agent):
                 
                 # Get Q-values from model
-                q_values = self.orchestrator.rl_agent.act(current_state, return_q_values=True)
-                if isinstance(q_values, tuple):
-                    action, q_vals = q_values
-                    q_values = q_vals.tolist() if hasattr(q_vals, 'tolist') else [0, 0, 0]
+                action = self.orchestrator.rl_agent.act(current_state, explore=False)
+                # Get Q-values separately if available
+                if hasattr(self.orchestrator.rl_agent, 'policy_net'):
+                    with torch.no_grad():
+                        state_tensor = torch.FloatTensor(current_state).unsqueeze(0).to(self.orchestrator.rl_agent.device)
+                        q_values_tensor = self.orchestrator.rl_agent.policy_net(state_tensor)
+                        if isinstance(q_values_tensor, tuple):
+                            q_values = q_values_tensor[0].cpu().numpy()[0].tolist()
+                        else:
+                            q_values = q_values_tensor.cpu().numpy()[0].tolist()
                 else:
-                    action = q_values
                     q_values = [0.33, 0.33, 0.34]  # Default uniform distribution
                 
                 confidence = max(q_values) / sum(q_values) if sum(q_values) > 0 else 0.33