Update handler.py

handler.py

@@ -2,33 +2,31 @@
 
 import io, base64, torch, open_clip
 from PIL import Image
+from mobileclip.modules.common.mobileone import reparameterize_model             # optional
 
 class EndpointHandler:
     """
-    MobileCLIP‑B (pretrained='datacompdr') zero‑shot classifier.
-    Expects JSON:
-      {
-        "inputs": {
-          "image": "<base64 PNG/JPEG>",
-          "candidate_labels": ["a photo of a cat", ...]
-        }
+    MobileCLIP‑B ('datacompdr') · text‑embedding cache.
+    Expects: {
+      "inputs": {
+        "image": "<base64>",
+        "candidate_labels": ["a photo of a cat", ...]
       }
+    }
     """
 
     def __init__(self, path=""):
-        # Model + transforms
+        # -- Load MobileCLIP‑B checkpoint identical to local run -------------
         self.model, _, self.preprocess = open_clip.create_model_and_transforms(
             "mobileclip_b", pretrained="datacompdr"
         )
-        self.model.eval()
+        self.model = reparameterize_model(self.model).eval()   # matches local pipeline
 
-        # Tokeniser & device
         self.tokenizer = open_clip.get_tokenizer("mobileclip_b")
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         self.model.to(self.device)
 
-        # Cache {prompt → 1×512 tensor}
-        self.cache: dict[str, torch.Tensor] = {}
+        self.cache: dict[str, torch.Tensor] = {}               # label → embedding
 
     def __call__(self, data):
         payload = data.get("inputs", data)
@@ -37,22 +35,22 @@ class EndpointHandler:
         if not labels:
             return {"error": "candidate_labels list is empty"}
 
-        # Image → tensor
+        # -------- image preprocessing --------------------------------------
         img = Image.open(io.BytesIO(base64.b64decode(img_b64))).convert("RGB")
         img_t = self.preprocess(img).unsqueeze(0).to(self.device)
 
-        # Text embeddings with cache
-        new_labels = [l for l in labels if l not in self.cache]
-        if new_labels:
-            tok = self.tokenizer(new_labels).to(self.device)
+        # -------- text embeddings with cache -------------------------------
+        new = [l for l in labels if l not in self.cache]
+        if new:
+            tok = self.tokenizer(new).to(self.device)
             with torch.no_grad():
                 emb = self.model.encode_text(tok)
                 emb = emb / emb.norm(dim=-1, keepdim=True)
-            for l, e in zip(new_labels, emb):
+            for l, e in zip(new, emb):
                 self.cache[l] = e
         txt_t = torch.stack([self.cache[l] for l in labels])
 
-        # Forward
+        # -------- forward & softmax ----------------------------------------
         with torch.no_grad(), torch.cuda.amp.autocast():
             img_f = self.model.encode_image(img_t)
             img_f = img_f / img_f.norm(dim=-1, keepdim=True)