bbbbbhtt555

inference.py

@@ -6,7 +6,11 @@ import torch
 import torch.nn as nn
 from PIL import Image
 from huggingface_hub import hf_hub_download
-import clip
+try:
+    import open_clip
+    CLIP_AVAILABLE = True
+except ImportError:
+    CLIP_AVAILABLE = False
 
 from utils.transforms import build_inference_transform
 from models.resnet_embedder import ResNetItemEmbedder
@@ -64,9 +68,16 @@ class InferenceService:
 
     def _load_clip(self) -> None:
         """Load CLIP model for category detection."""
+        if not CLIP_AVAILABLE:
+            print("⚠️ CLIP not available, using filename-based category detection")
+            self.clip_model, self.clip_preprocess = None, None
+            return
+            
         try:
             print("🔄 Loading CLIP model for category detection...")
-            self.clip_model, self.clip_preprocess = clip.load("ViT-B/32", device=self.device)
+            self.clip_model, _, self.clip_preprocess = open_clip.create_model_and_transforms(
+                'ViT-B-32', pretrained='laion2b_s34b_b79k', device=self.device
+            )
             print("✅ CLIP model loaded successfully")
         except Exception as e:
             print(f"❌ Failed to load CLIP model: {e}")
@@ -95,7 +106,7 @@ class InferenceService:
             
             # Prepare image and text
             image_input = self.clip_preprocess(image).unsqueeze(0).to(self.device)
-            text_inputs = clip.tokenize(categories).to(self.device)
+            text_inputs = open_clip.tokenize(categories).to(self.device)
             
             # Get predictions
             with torch.no_grad():
@@ -130,6 +141,32 @@ class InferenceService:
         except Exception as e:
             print(f"❌ CLIP category detection failed: {e}")
             return "other"
+    
+    def _detect_category_from_filename(self, filename: str) -> str:
+        """Fallback: Detect category from filename using keyword matching."""
+        if not filename:
+            return "other"
+            
+        filename_lower = filename.lower()
+        
+        # Upper body items
+        if any(kw in filename_lower for kw in ["shirt", "top", "blouse", "tank", "hoodie", "sweater", "jacket", "blazer", "coat"]):
+            return "shirt"
+        
+        # Bottom items  
+        if any(kw in filename_lower for kw in ["pant", "jean", "short", "skirt", "trouser", "legging", "jogger"]):
+            return "pants"
+        
+        # Shoes
+        if any(kw in filename_lower for kw in ["shoe", "boot", "sneaker", "sandal", "heel", "loafer", "oxford"]):
+            return "shoes"
+        
+        # Accessories
+        if any(kw in filename_lower for kw in ["watch", "ring", "necklace", "bracelet", "bag", "hat", "belt", "scarf"]):
+            return "accessory"
+        
+        # Default fallback
+        return "other"
 
     def _load_resnet(self) -> tuple[nn.Module, bool]:
         strategy = os.getenv("MODEL_LOAD_STRATEGY", "state_dict")
@@ -311,15 +348,18 @@ class InferenceService:
         for i, it in enumerate(items):
             print(f"🔍 DEBUG: Processing item {i}: id={it.get('id')}, has_image={it.get('image') is not None}, has_embedding={it.get('embedding') is not None}")
             
-            # Auto-detect category using CLIP if not provided or is None
+            # Auto-detect category if not provided or is None
             category = it.get("category")
             if not category or category == "None" or category == "":
-                if it.get("image") is not None:
+                if it.get("image") is not None and self.clip_model is not None:
                     print(f"🔍 DEBUG: Auto-detecting category for item {i} using CLIP...")
                     category = self._detect_category_with_clip(it["image"])
                 else:
-                    category = "other"
-                    print(f"🔍 DEBUG: No image available for item {i}, using 'other' category")
+                    # Fallback to filename-based detection
+                    filename = it.get("id", "")
+                    print(f"🔍 DEBUG: Auto-detecting category for item {i} using filename '{filename}'...")
+                    category = self._detect_category_from_filename(filename)
+                    print(f"🔍 DEBUG: Filename-based detection result: '{category}'")
             
             emb = it.get("embedding")
             if emb is None and it.get("image") is not None: