Update handler.py

handler.py

@@ -1,16 +1,17 @@
 # handler.py  (repo root)
 
+# handler.py  (repo root)
+
 import io, base64, torch
 from PIL import Image
 import open_clip
-from open_clip import fuse_conv_bn_sequential
 
 
 class EndpointHandler:
     """
-    Zero‑shot classifier for MobileCLIP‑B (OpenCLIP).
+    Zero‑shot classifier for MobileCLIP‑B (OpenCLIP) with a text‑embedding cache.
 
-    Client JSON format:
+    Client JSON:
     {
       "inputs": {
         "image": "<base64 PNG/JPEG>",
@@ -19,68 +20,57 @@ class EndpointHandler:
     }
     """
 
-    # ----------------------------------------------------- #
-    #               INITIALISATION (once)                  #
-    # ----------------------------------------------------- #
+    # ------------------------------------------------- #
+    #                 INITIALISATION                    #
+    # ------------------------------------------------- #
     def __init__(self, path: str = ""):
         weights = f"{path}/mobileclip_b.pt"
 
-        # Load model + transforms
         self.model, _, self.preprocess = open_clip.create_model_and_transforms(
             "MobileCLIP-B", pretrained=weights
         )
+        self.model.eval()
 
-        # Fuse Conv+BN for faster inference
-        self.model = fuse_conv_bn_sequential(self.model).eval()
-
-        # Tokeniser
         self.tokenizer = open_clip.get_tokenizer("MobileCLIP-B")
-
-        # Device
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         self.model.to(self.device)
 
-        # -------- text‑embedding cache --------
-        # key: prompt string  •  value: torch.Tensor [512] on correct device
+        # cache: {prompt -> 1×512 tensor on device}
         self.label_cache: dict[str, torch.Tensor] = {}
 
-    # ----------------------------------------------------- #
-    #              INFERENCE  (per request)                #
-    # ----------------------------------------------------- #
+    # ------------------------------------------------- #
+    #                    INFERENCE                      #
+    # ------------------------------------------------- #
     def __call__(self, data):
-        # 1. Unwrap the HF "inputs" envelope
         payload = data.get("inputs", data)
 
-        img_b64  = payload["image"]
-        labels   = payload.get("candidate_labels", [])
+        img_b64 = payload["image"]
+        labels  = payload.get("candidate_labels", [])
         if not labels:
             return {"error": "candidate_labels list is empty"}
 
-        # 2. Decode & preprocess image
+        # --- image ----
         image = Image.open(io.BytesIO(base64.b64decode(img_b64))).convert("RGB")
         img_tensor = self.preprocess(image).unsqueeze(0).to(self.device)
 
-        # 3. Text embeddings with cache
+        # --- text (with cache) ----
         missing = [l for l in labels if l not in self.label_cache]
         if missing:
             tokens = self.tokenizer(missing).to(self.device)
             with torch.no_grad():
                 emb = self.model.encode_text(tokens)
                 emb = emb / emb.norm(dim=-1, keepdim=True)
-            for lbl, vec in zip(missing, emb):
-                self.label_cache[lbl] = vec  # store on device
-
+            for l, e in zip(missing, emb):
+                self.label_cache[l] = e
         txt_feat = torch.stack([self.label_cache[l] for l in labels])
 
-        # 4. Forward pass for image
+        # --- forward & softmax ----
         with torch.no_grad(), torch.cuda.amp.autocast():
             img_feat = self.model.encode_image(img_tensor)
             img_feat = img_feat / img_feat.norm(dim=-1, keepdim=True)
+            probs = (100 * img_feat @ txt_feat.T).softmax(dim=-1)[0].tolist()
 
-        # 5. Similarity & softmax
-        probs = (100 * img_feat @ txt_feat.T).softmax(dim=-1)[0].tolist()
-
-        # 6. Return sorted list
+        # --- sorted output ----
         return [
             {"label": l, "score": float(p)}
             for l, p in sorted(zip(labels, probs), key=lambda x: x[1], reverse=True)