handler.py

# 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).

    Client JSON format:
    {
      "inputs": {
        "image": "<base64 PNG/JPEG>",
        "candidate_labels": ["cat", "dog", ...]
      }
    }
    """

    # ----------------------------------------------------- #
    #               INITIALISATION (once)                  #
    # ----------------------------------------------------- #
    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
        )

        # 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
        self.label_cache: dict[str, torch.Tensor] = {}

    # ----------------------------------------------------- #
    #              INFERENCE  (per request)                #
    # ----------------------------------------------------- #
    def __call__(self, data):
        # 1. Unwrap the HF "inputs" envelope
        payload = data.get("inputs", data)

        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.open(io.BytesIO(base64.b64decode(img_b64))).convert("RGB")
        img_tensor = self.preprocess(image).unsqueeze(0).to(self.device)

        # 3. Text embeddings 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

        txt_feat = torch.stack([self.label_cache[l] for l in labels])

        # 4. Forward pass for image
        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)

        # 5. Similarity & softmax
        probs = (100 * img_feat @ txt_feat.T).softmax(dim=-1)[0].tolist()

        # 6. Return sorted list
        return [
            {"label": l, "score": float(p)}
            for l, p in sorted(zip(labels, probs), key=lambda x: x[1], reverse=True)
        ]

# # handler.py  (repo root)
# import io, base64, torch
# from PIL import Image
# import open_clip

# class EndpointHandler:
#     """
#     Zero‑shot classifier for MobileCLIP‑B (OpenCLIP).

#     Expected client JSON *to the endpoint*:
#     {
#       "inputs": {
#         "image": "<base64 PNG/JPEG>",
#         "candidate_labels": ["cat", "dog", ...]
#       }
#     }
#     """

#     def __init__(self, path: str = ""):
#         weights = f"{path}/mobileclip_b.pt"
#         self.model, _, self.preprocess = open_clip.create_model_and_transforms(
#             "MobileCLIP-B", pretrained=weights
#         )
#         self.model.eval()

#         self.tokenizer = open_clip.get_tokenizer("MobileCLIP-B")
#         self.device = "cuda" if torch.cuda.is_available() else "cpu"
#         self.model.to(self.device)

#     def __call__(self, data):
#         # ── unwrap Hugging Face's `inputs` envelope ───────────
#         payload = data.get("inputs", data)

#         img_b64 = payload["image"]
#         labels  = payload.get("candidate_labels", [])
#         if not labels:
#             return {"error": "candidate_labels list is empty"}

#         # Decode & preprocess image
#         image = Image.open(io.BytesIO(base64.b64decode(img_b64))).convert("RGB")
#         img_tensor = self.preprocess(image).unsqueeze(0).to(self.device)

#         # Tokenise labels
#         text_tokens = self.tokenizer(labels).to(self.device)

#         # Forward pass
#         with torch.no_grad(), torch.cuda.amp.autocast():
#             img_feat = self.model.encode_image(img_tensor)
#             txt_feat = self.model.encode_text(text_tokens)
#             img_feat = img_feat / img_feat.norm(dim=-1, keepdim=True)
#             txt_feat = txt_feat / txt_feat.norm(dim=-1, keepdim=True)
#             probs = (100 * img_feat @ txt_feat.T).softmax(dim=-1)[0].tolist()

#         # Sorted output
#         return [
#             {"label": l, "score": float(p)}
#             for l, p in sorted(zip(labels, probs), key=lambda x: x[1], reverse=True)
#         ]