feat: enhance model compatibility and input handling for deep learning wrappers

local_demo.py

@@ -79,6 +79,7 @@ def _safe_model_name(name: str) -> str:
         .replace("(", "")
         .replace(")", "")
         .replace("/", "_")
+        .replace("-", "_")
     )
 
 
@@ -132,10 +133,16 @@ def _load_dataset_summary() -> dict[str, Any]:
 
 
 def _find_matching_name(raw_name: str, training_results: dict[str, Any]) -> str:
+    # DL model files are stored as "model_dl_<name>" — strip the "dl_" infix too
+    candidates = [raw_name]
+    if raw_name.startswith("dl_"):
+        candidates.append(raw_name[3:])  # "dl_deep_mlp_..." -> "deep_mlp_..."
+
     for name in training_results:
         if name.startswith("_"):
             continue
-        if _safe_model_name(name) == raw_name:
+        safe = _safe_model_name(name)
+        if safe in candidates:
             return name
     return raw_name.replace("_", " ").title()
 
@@ -157,6 +164,15 @@ def _load_artifacts() -> DemoArtifacts:
     scaler = _load_pickle(scaler_path)
     feature_cols = _load_json(columns_path)
     training_results = _load_json(results_path)
+
+    # Merge DL metrics into the unified training_results dict
+    dl_results_path = MODELS_DIR / "deep_learning_results.json"
+    if dl_results_path.exists():
+        dl_results = _load_json(dl_results_path)
+        for name, metrics in dl_results.items():
+            if name not in training_results:
+                training_results[name] = metrics
+
     feature_importance = training_results.get("_feature_importance", {})
     best_model_name = training_results.get("_best_model", "Gradient Boosting")
 
@@ -327,6 +343,11 @@ def _build_signal_html(features: dict[str, float]) -> str:
     return "".join(parts)
 
 
+def _is_dl_wrapper(model: Any) -> bool:
+    """True for TorchSklearnWrapper — it has its own internal scaler."""
+    return type(model).__name__ == "TorchSklearnWrapper"
+
+
 def _build_model_table_html(
     selected_model_name: str,
     feature_vector: np.ndarray,
@@ -338,7 +359,9 @@ def _build_model_table_html(
         try:
             with warnings.catch_warnings():
                 warnings.simplefilter("ignore", category=UserWarning)
-                probability = float(model.predict_proba(scaled)[0][1])
+                # DL wrappers scale internally — pass raw vector to avoid double-scaling
+                input_vector = feature_vector if _is_dl_wrapper(model) else scaled
+                probability = float(model.predict_proba(input_vector)[0][1])
         except Exception:  # noqa: BLE001
             continue
         scored_rows.append((name, probability))
@@ -461,11 +484,15 @@ def analyze_audio(audio_file: Any, selected_model_label: str) -> tuple[str, str,
         return error_html, "", "", ""
 
     feature_vector = _build_feature_vector(features)
-    scaled = ARTIFACTS.scaler.transform(feature_vector)
     model = ARTIFACTS.loaded_models[selected_model_name]
+    # DL wrappers scale internally; ML models expect pre-scaled input
+    if _is_dl_wrapper(model):
+        input_vector = feature_vector
+    else:
+        input_vector = ARTIFACTS.scaler.transform(feature_vector)
     with warnings.catch_warnings():
         warnings.simplefilter("ignore", category=UserWarning)
-        ai_prob = float(model.predict_proba(scaled)[0][1])
+        ai_prob = float(model.predict_proba(input_vector)[0][1])
     elapsed = time.time() - start_time
 
     result_html = _build_result_html(ai_prob, duration, elapsed, selected_model_name)
@@ -485,8 +512,8 @@ def build_models_md() -> str:
         f"- Ozellik sayisi: **{training_results.get('_n_features', len(ARTIFACTS.feature_cols))}**",
         f"- CV kat sayisi: **{training_results.get('_n_folds', '?')}**",
         "",
-        "| Model | CV AUC | Holdout AUC | Acc | F1 |",
-        "|------|--------|-------------|-----|----|",
+        "| Model | Tip | CV AUC | Holdout AUC | Acc | F1 |",
+        "|------|-----|--------|-------------|-----|----|",
     ]
 
     model_names = [
@@ -499,8 +526,9 @@ def build_models_md() -> str:
     for name in model_names:
         result = training_results[name]
         display = f"**{name}**" if name == ARTIFACTS.best_model_name else name
+        model_type = "DL" if result.get("type") == "deep_learning" else "ML"
         lines.append(
-            f"| {display} | {result.get('roc_auc', 0.0):.4f} | "
+            f"| {display} | {model_type} | {result.get('roc_auc', 0.0):.4f} | "
             f"{result.get('validation_auc', 0.0):.4f} | "
             f"{result.get('accuracy', 0.0):.4f} | {result.get('f1', 0.0):.4f} |"
         )