Update semantic_breed_recommender.py

semantic_breed_recommender.py

@@ -37,6 +37,7 @@ class SemanticBreedRecommender:
         """Initialize the semantic recommender"""
         self.model_name = 'all-MiniLM-L6-v2'  # Efficient SBERT model
         self.sbert_model = None
+        self._sbert_loading_attempted = False
         self.breed_vectors = {}
         self.breed_list = self._get_breed_list()
         self.comparative_keywords = {
@@ -44,13 +45,9 @@ class SemanticBreedRecommender:
             'then': 0.7, 'second': 0.7, 'followed': 0.6,
             'third': 0.5, 'least': 0.3, 'dislike': 0.2
         }
-        # self.query_engine = QueryUnderstandingEngine()
-        # self.constraint_manager = ConstraintManager()
-        # self.multi_head_scorer = None  # Will be initialized with SBERT model
-        # self.score_calibrator = ScoreCalibrator()
-        # self.config_manager = get_config_manager()
-        self._initialize_model()
-        self._build_breed_vectors()
+        # Defer SBERT model loading until needed in GPU context
+        # This prevents CUDA initialization issues in ZeroGPU environment
+        print("SemanticBreedRecommender initialized (SBERT loading deferred)")
 
         # Initialize multi-head scorer with SBERT model if enhanced mode is enabled
         # if self.sbert_model:
@@ -74,18 +71,24 @@ class SemanticBreedRecommender:
                    'Bulldog', 'Poodle', 'Beagle', 'Rottweiler', 'Yorkshire_Terrier']
 
     def _initialize_model(self):
-        """Initialize SBERT model with fallback"""
+        """Initialize SBERT model with fallback - designed for ZeroGPU compatibility"""
+        if self.sbert_model is not None or self._sbert_loading_attempted:
+            return self.sbert_model
+            
         try:
-            print("Loading SBERT model...")
+            print("Loading SBERT model in GPU context...")
             # Try different model names if the primary one fails
             model_options = ['all-MiniLM-L6-v2', 'all-mpnet-base-v2', 'all-MiniLM-L12-v2']
 
             for model_name in model_options:
                 try:
-                    self.sbert_model = SentenceTransformer(model_name)
+                    # Specify device explicitly to handle ZeroGPU environment
+                    import torch
+                    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+                    self.sbert_model = SentenceTransformer(model_name, device=device)
                     self.model_name = model_name
-                    print(f"SBERT model {model_name} loaded successfully")
-                    return
+                    print(f"SBERT model {model_name} loaded successfully on {device}")
+                    return self.sbert_model
                 except Exception as model_e:
                     print(f"Failed to load {model_name}: {str(model_e)}")
                     continue
@@ -93,12 +96,16 @@ class SemanticBreedRecommender:
             # If all models fail
             print("All SBERT models failed to load. Using basic text matching fallback.")
             self.sbert_model = None
+            return None
 
         except Exception as e:
             print(f"Failed to initialize any SBERT model: {str(e)}")
             print(traceback.format_exc())
             print("Will provide basic text-based recommendations without embeddings")
             self.sbert_model = None
+            return None
+        finally:
+            self._sbert_loading_attempted = True
 
     def _create_breed_description(self, breed: str) -> str:
         """Create comprehensive natural language description for breed with all key characteristics"""
@@ -321,10 +328,14 @@ class SemanticBreedRecommender:
             return f"{breed.replace('_', ' ')} is a dog breed with unique characteristics."
 
     def _build_breed_vectors(self):
-        """Build vector representations for all breeds"""
+        """Build vector representations for all breeds - called lazily when needed"""
         try:
             print("Building breed vector database...")
 
+            # Initialize model if not already done
+            if self.sbert_model is None:
+                self._initialize_model()
+                
             # Skip if model is not available
             if self.sbert_model is None:
                 print("SBERT model not available, skipping vector building")
@@ -959,12 +970,20 @@ class SemanticBreedRecommender:
         try:
             print(f"Processing user input: {user_input}")
 
+            # 嘗試載入SBERT模型（如果尚未載入）
+            if self.sbert_model is None:
+                self._initialize_model()
+                
             # Check if model is available - if not, raise error
             if self.sbert_model is None:
                 error_msg = "SBERT model not available. This could be due to:\n• Model download failed\n• Insufficient memory\n• Network connectivity issues\n\nPlease check your environment and try again."
                 print(f"ERROR: {error_msg}")
                 raise RuntimeError(error_msg)
 
+            # 確保breed vectors已建構
+            if not self.breed_vectors:
+                self._build_breed_vectors()
+
             # Generate user input embedding
             user_embedding = self.sbert_model.encode(user_input, convert_to_tensor=False)
 
@@ -1584,6 +1603,10 @@ def get_breed_recommendations_by_description(user_description: str,
     try:
         print("Initializing Enhanced SemanticBreedRecommender...")
         recommender = SemanticBreedRecommender()
+        
+        # 嘗試載入SBERT模型（如果尚未載入）
+        if not recommender.sbert_model:
+            recommender._initialize_model()
 
         # 優先使用整合統一評分系統的增強推薦
         print("Using enhanced recommendation system with unified scoring")
@@ -1628,11 +1651,19 @@ def get_enhanced_recommendations_with_unified_scoring(user_description: str, top
         # 創建基本推薦器實例
         recommender = SemanticBreedRecommender()
 
+        # 嘗試載入SBERT模型（如果尚未載入）
+        if not recommender.sbert_model:
+            recommender._initialize_model()
+            
         if not recommender.sbert_model:
             print("SBERT model not available, using basic text matching...")
             # 使用基本文字匹配邏輯
             return _get_basic_text_matching_recommendations(user_description, top_k)
 
+        # 確保breed vectors已建構
+        if not recommender.breed_vectors:
+            recommender._build_breed_vectors()
+
         # 使用語意相似度推薦
         recommendations = []
         user_embedding = recommender.sbert_model.encode(user_description)
@@ -2212,4 +2243,4 @@ def _get_basic_text_matching_recommendations(user_description: str, top_k: int =
     except Exception as e:
         error_msg = f"Error in basic text matching: {str(e)}"
         print(f"ERROR: {error_msg}")
-        raise RuntimeError(error_msg) from e
+        raise RuntimeError(error_msg) from e