Auto-deploy from GitHub: 9ca7e761f50169763dd91ccaf7c76c6bdcbe90b2

scripts/predict.py

@@ -31,13 +31,17 @@ def predict_pipeline(audio_file, lyrics):
     label : int
         A numerical representation of the prediction
     """
-
     # 1.) Instantiate LLM2Vec Model
     llm2vec_model = load_llm2vec_model()
 
     # 2.) Preprocess both audio and lyrics
     audio, lyrics = single_preprocessing(audio_file, lyrics)
 
+    # Truncate to 2 minutes to match explain pipeline
+    target_samples = int(2 * 60 * 22050)
+    if len(audio) > target_samples:
+        audio = audio[:target_samples]
+
     # 3.) Call the train method for both models
     audio_features = spectttra_predict(audio)
     lyrics_features = l2vec_single_train(llm2vec_model, lyrics)
@@ -59,7 +63,7 @@ def predict_pipeline(audio_file, lyrics):
     config = load_config("config/model_config.yml")
     classifier = build_mlp(input_dim=results.shape[1], config=config)
 
-    # 7.) Load trained weights (make sure this path matches where you saved your model)
+    # 7.) Load trained weights
     model_path = "models/mlp/mlp_best.pth"
     classifier.load_model(model_path)
     classifier.model.eval()

src/models/mlp.py

@@ -442,7 +442,9 @@ class MLPClassifier:
 
         return probabilities, predictions
 
-    def predict_single(self, features: np.ndarray, temperature: float = 2.5) -> Tuple[float, int, str]:
+    def predict_single(
+        self, features: np.ndarray, temperature: float = 2.5
+    ) -> Tuple[float, int, str]:
         """
         Predict whether a single song is AI-generated or human-composed.
 
@@ -487,15 +489,15 @@ class MLPClassifier:
         with torch.no_grad():
             features_tensor = torch.FloatTensor(features).to(self.device)
             outputs = self.model(features_tensor)
-            logit = torch.logit(outputs.clamp(1e-6, 1 - 1e-6))
-            probabilities = torch.sigmoid(logit / temperature).item()
-            probabilities = np.clip(probabilities, 0.01, 0.99)
+            probabilities = outputs.item()  # Just use raw output
 
         # Extract single results
         prediction = int(probabilities >= 0.5)
         label = "Human-Composed" if prediction == 1 else "AI-Generated"
-        probability = probabilities*100 if prediction == 1 else (1 - probabilities)*100
-        
+        probability = (
+            probabilities * 100 if prediction == 1 else (1 - probabilities) * 100
+        )
+
         return probability, prediction, label
 
     def predict_batch(self, features: np.ndarray, return_details: bool = False) -> Dict: