Update ml_engine/titan_engine.py

ml_engine/titan_engine.py

@@ -1,10 +1,9 @@
 # ==============================================================================
-# 🛡️ ml_engine/titan_engine.py (V3.0 - PyTorch ResNet Production Edition)
+# 🛡️ ml_engine/titan_engine.py (V3.1 - Fix List vs DataFrame Issue)
 # ==============================================================================
 # GEM-Architect Approved
-# - Matches Titan V2 Training Pipeline exactly.
-# - Implements cached Scaler loading.
-# - Uses PyTorch for inference.
+# - Fixes AttributeError: 'list' object has no attribute 'empty'
+# - Auto-converts raw list data to Pandas DataFrame with correct columns.
 # ==============================================================================
 
 import os
@@ -156,7 +155,10 @@ class TitanEngine:
         Assuming 'df' has at least 100 rows.
         """
         try:
-            df = df.copy().sort_values('timestamp')
+            df = df.copy()
+            # Ensure sorting if timestamp exists, else assume ordered list
+            if 'timestamp' in df.columns:
+                df = df.sort_values('timestamp')
             
             # Basic conversions
             close = df['close'].values.astype(float)
@@ -229,44 +231,45 @@ class TitanEngine:
     def predict(self, ohlcv_data: dict) -> dict:
         """
         Main Interface used by Processor.
-        Expected input: {'5m': dataframe, '15m': dataframe ...}
-        We primarily use '15m' as per the training configuration.
+        Handles both List and DataFrame inputs robustly.
         """
         if not self.initialized: return {'score': 0.0, 'error': 'Not Initialized'}
         
         try:
-            # We use 15m data as the main driver (matching training)
+            # We use 15m data as the main driver
             target_tf = '15m'
-            df = ohlcv_data.get(target_tf)
+            raw_data = ohlcv_data.get(target_tf)
             
-            if df is None or df.empty:
-                # Fallback to 5m resampled if needed, but for now error out
+            if raw_data is None:
                 return {'score': 0.0, 'error': 'No 15m Data'}
+
+            # ✅ FIX: Auto-detect List vs DataFrame
+            if isinstance(raw_data, list):
+                # Standard CCXT OHLCV structure: [timestamp, open, high, low, close, volume]
+                df = pd.DataFrame(raw_data, columns=['timestamp', 'open', 'high', 'low', 'close', 'volume'])
+            elif isinstance(raw_data, pd.DataFrame):
+                df = raw_data
+            else:
+                return {'score': 0.0, 'error': f'Invalid Data Type: {type(raw_data)}'}
+            
+            if df.empty:
+                return {'score': 0.0, 'error': 'Empty Data'}
             
             # Preprocess
             X_raw = self.preprocess_live_data(df)
             if X_raw is None:
                 return {'score': 0.0, 'error': 'Not enough data for window'}
             
-            # Scale (CRITICAL: Transform, do not fit)
-            # Scaler expects (N_samples, N_features). X_raw is (64, 10).
-            # We need to flatten? No, scaler works on features.
-            # In training: scaler.transform(chunk_df[FEATURES])
-            # So we scale the whole window row by row.
+            # Scale
             X_scaled = self.scaler.transform(X_raw)
             
-            # Prepare Tensor: (Batch, Channels, Length) -> (1, 10, 64)
-            # Transpose: (64, 10) -> (10, 64)
+            # Prepare Tensor
             X_tensor = torch.tensor(X_scaled.T).unsqueeze(0).to(self.device)
             
             # Inference
             with torch.no_grad():
                 logits = self.model(X_tensor)
                 probs = torch.softmax(logits, dim=1).cpu().numpy()[0]
-                
-            # probs = [Prob(Neutral), Prob(Loss), Prob(Win)]
-            # Titan Score = Prob(Win)
-            # We also return raw probs for Oracle
             
             return {
                 'score': float(probs[2]), # Win Probability