training WIP

ANNOTATE/core/real_training_adapter.py

@@ -4090,20 +4090,37 @@ class RealTrainingAdapter:
                 return None
             
             # Override the future candle target with actual candle data
-            actual = prediction_sample['actual_candle']  # [O, H, L, C]
+            actual = prediction_sample['actual_candle']  # [O, H, L, C, V] or [O, H, L, C]
             
             # Create target tensor for the specific timeframe
             import torch
-            device = batch['prices_1m'].device if 'prices_1m' in batch else torch.device('cpu')
+            # Get device from any available tensor in batch
+            device = torch.device('cpu')
+            for key in ['price_data_1m', 'price_data_1h', 'price_data_1d', 'prices_1m']:
+                if key in batch and batch[key] is not None:
+                    device = batch[key].device
+                    break
             
-            # Target candle: [O, H, L, C, V] - we don't have actual volume, use predicted
-            target_candle = [
-                actual[0],  # Open
-                actual[1],  # High
-                actual[2],  # Low
-                actual[3],  # Close
-                prediction_sample['predicted_candle'][4]  # Volume (from prediction)
-            ]
+            # Target candle: [O, H, L, C, V]
+            # Use actual volume if available, otherwise use predicted volume
+            if len(actual) >= 5:
+                target_candle = [
+                    float(actual[0]),  # Open
+                    float(actual[1]),  # High
+                    float(actual[2]),  # Low
+                    float(actual[3]),  # Close
+                    float(actual[4])   # Volume (from actual)
+                ]
+            else:
+                # Fallback: use predicted volume if actual doesn't have it
+                predicted = prediction_sample.get('predicted_candle', [0, 0, 0, 0, 0])
+                target_candle = [
+                    float(actual[0]),  # Open
+                    float(actual[1]),  # High
+                    float(actual[2]),  # Low
+                    float(actual[3]),  # Close
+                    float(predicted[4] if len(predicted) > 4 else 0.0)  # Volume (from prediction)
+                ]
             
             # Add to batch based on timeframe
             if timeframe == '1s':