fix

app.py

@@ -183,28 +183,26 @@ def load_model():
             return None, str(e)
 
 
-# ── Lazy model loading ─────────────────────────────────────────────────────
-# Loading the model at import time doubles peak memory (TF init + model
-# deserialization happen simultaneously). Deferring until the first request
-# keeps startup lightweight and avoids the OOM → SIGSEGV on free-tier HF.
-_model = None
-_model_error = ""
-
-def _get_model():
-    """Return the cached model, loading it on first call."""
-    global _model, _model_error
-    if _model is None and _model_error == "":
-        print("Lazy-loading model on first request…")
-        _model, _model_error = load_model()
-        if _model is None:
-            print(f"Warning: Could not load model. Error: {_model_error}")
-    return _model, _model_error
+# ── Model loading ──────────────────────────────────────────────────────────
+model, model_error = load_model()
+if model is None:
+    print(f"Warning: Could not load model. Error: {model_error}")
+else:
+    # Warm-up prediction: forces TensorFlow to trace and compile the graph
+    # on the main thread. If the first prediction happens in a Gradio 
+    # background thread, TF can deadlock and freeze the app.
+    print("Warming up model to prevent thread deadlock...")
+    dummy_spec = np.zeros((1, N_MELS, MAX_TIME_FRAMES, 1), dtype=np.float32)
+    dummy_stat = np.zeros((1, N_1D_FEATURES), dtype=np.float32)
+    try:
+        model.predict([dummy_spec, dummy_stat], verbose=0)
+        print("Model warm-up complete.")
+    except Exception as e:
+        print(f"Warm-up failed (this might be fine): {e}")
 
 preprocessor = MachineListenerPreprocessor()
 
-
 def predict(audio_filepath):
-    model, model_error = _get_model()
     if model is None:
         return f"Model not loaded properly. Error: {model_error}"
     if audio_filepath is None: