Update app.py

app.py

@@ -27,20 +27,25 @@ app.add_middleware(
 # --- CHARGE LE MODÈLE MARQO FASHIONCLIP ---
 print("⚠️ Démarrage du chargement du modèle Marqo fashionCLIP...")
 model = None
+tokenizer = None
 processor = None
 
 def load_marqo_model():
-    global model, processor
+    global model, tokenizer, processor
     try:
-        from transformers import CLIPProcessor, CLIPModel
+        from transformers import CLIPModel, CLIPTokenizer, CLIPImageProcessor
         
         model_name = "Marqo/marqo-fashionCLIP"
+        
+        # Charger les composants séparément
         model = CLIPModel.from_pretrained(
             model_name,
             cache_dir="/tmp/cache",
             torch_dtype=torch.float16
         )
-        processor = CLIPProcessor.from_pretrained(model_name)
+        tokenizer = CLIPTokenizer.from_pretrained(model_name)
+        processor = CLIPImageProcessor.from_pretrained(model_name)
+        
         print("✅ Modèle Marqo fashionCLIP chargé avec succès !")
     except Exception as e:
         print(f"❌ Erreur chargement modèle Marqo: {e}")
@@ -52,10 +57,10 @@ async def startup_event():
     thread.daemon = True
     thread.start()
 
-# Catégories fashion (textes plus courts et uniformes)
+# Catégories fashion (textes courts)
 categories = [
-    "t-shirt", "dress", "jeans", "shirt", "skirt", "sneakers", 
-    "handbag", "jacket", "shorts", "sweater", "coat", "heels"
+    "a t-shirt", "a dress", "jeans", "a shirt", "a skirt", 
+    "sneakers", "a handbag", "a jacket", "shorts", "a sweater"
 ]
 
 @app.get("/")
@@ -66,13 +71,14 @@ def read_root():
 def health_check():
     return {
         "model_loaded": model is not None,
+        "tokenizer_loaded": tokenizer is not None,
         "processor_loaded": processor is not None,
-        "status": "ready" if model and processor else "loading"
+        "status": "ready" if all([model, tokenizer, processor]) else "loading"
     }
 
 @app.post("/analyze")
 async def analyze_image(file: UploadFile = File(...)):
-    if model is None or processor is None:
+    if model is None or tokenizer is None or processor is None:
         return {"error": "Model not loaded yet. Please wait or check /health endpoint."}
     
     try:
@@ -83,32 +89,43 @@ async def analyze_image(file: UploadFile = File(...)):
         # Réduire la taille
         image.thumbnail((384, 384))
         
-        # --- SOLUTION DÉFINITIVE ---
-        # Traiter chaque catégorie SÉPARÉMENT pour éviter les problèmes de padding
-        similarities = []
+        # --- NOUVELLE APPROCHE SANS PROCESSOR BATCH ---
+        # 1. Préparer l'image
+        image_input = processor(images=image, return_tensors="pt")
+        
+        # 2. Préparer le texte - CHAQUE CATÉGORIE INDIVIDUELLEMENT
+        text_features_list = []
         
         for category in categories:
-            # Préparer les inputs pour UNE catégorie à la fois
-            inputs = processor(
-                text=[category],  # Une seule catégorie
-                images=image,
-                return_tensors="pt",
-                padding=True,    # Padding pour une seule phrase
-                truncation=True
+            # Tokenizer chaque catégorie séparément
+            text_inputs = tokenizer(
+                category, 
+                return_tensors="pt", 
+                padding=True, 
+                truncation=True,
+                max_length=77
             )
             
-            # Déplacer sur le device du modèle
-            device = next(model.parameters()).device
-            inputs = {k: v.to(device) for k, v in inputs.items()}
-            
             with torch.no_grad():
-                outputs = model(**inputs)
+                text_features = model.get_text_features(**text_inputs)
+                text_features_list.append(text_features)
+        
+        # 3. Get image features
+        with torch.no_grad():
+            image_features = model.get_image_features(**image_input)
+        
+        # 4. Calculer les similarités
+        similarities = []
+        for text_features in text_features_list:
+            # Normaliser les features
+            image_features_norm = image_features / image_features.norm(dim=-1, keepdim=True)
+            text_features_norm = text_features / text_features.norm(dim=-1, keepdim=True)
             
-            # Récupérer le score de similarité
-            similarity_score = outputs.logits_per_image.item()
-            similarities.append(similarity_score)
+            # Calculer la similarité cosinus
+            similarity = (image_features_norm @ text_features_norm.T).squeeze()
+            similarities.append(similarity.item())
         
-        # Convertir en tensor et calculer les probabilités
+        # 5. Convertir en probabilités
         similarities_tensor = torch.tensor(similarities)
         probs = torch.nn.functional.softmax(similarities_tensor, dim=0)
         
@@ -142,40 +159,25 @@ async def analyze_image(file: UploadFile = File(...)):
     except Exception as e:
         return {"error": f"Erreur lors de l'analyse: {str(e)}"}
 
-# Interface de test
+# Interface de test SIMPLIFIÉE
 @app.get("/test-ui", response_class=HTMLResponse)
 async def test_ui():
     return """
     <html>
         <head>
-            <title>Fashion Detection Test</title>
+            <title>Fashion Detection</title>
             <style>
-                body { font-family: Arial, sans-serif; margin: 40px; }
-                .container { max-width: 600px; margin: 0 auto; }
-                form { border: 2px dashed #ccc; padding: 30px; text-align: center; }
-                input[type="file"] { margin: 10px 0; }
-                input[type="submit"] { 
-                    background: #007bff; color: white; padding: 10px 20px; 
-                    border: none; cursor: pointer; border-radius: 5px;
-                }
-                .result { margin-top: 20px; padding: 20px; background: #f0f8ff; }
+                body { font-family: Arial, sans-serif; margin: 40px; text-align: center; }
+                form { border: 2px dashed #ccc; padding: 30px; display: inline-block; }
             </style>
         </head>
         <body>
-            <div class="container">
-                <h1>🎨 Fashion Detection AI</h1>
-                <form action="/analyze" method="post" enctype="multipart/form-data">
-                    <h3>Uploader une image de vêtement :</h3>
-                    <input type="file" name="file" accept="image/*" required>
-                    <br><br>
-                    <input type="submit" value="Analyser l'image 👗">
-                </form>
-                
-                <div class="result">
-                    <h3>📋 Résultat de l'analyse :</h3>
-                    <p>Attendez l'upload et le traitement de l'image...</p>
-                </div>
-            </div>
+            <h1>🎨 Fashion Detection</h1>
+            <form action="/analyze" method="post" enctype="multipart/form-data">
+                <input type="file" name="file" accept="image/*" required>
+                <br><br>
+                <input type="submit" value="Analyze">
+            </form>
         </body>
     </html>
     """