handler.py

# handler.py  – place in repo root
import io, base64, torch
from PIL import Image
from transformers import CLIPModel, CLIPProcessor


class EndpointHandler:
    """
    Custom zero‑shot classifier replicating local OpenAI‑CLIP logic.

    Client JSON must look like:
      {
        "inputs": {
          "image": "<base64 PNG/JPEG>",
          "candidate_labels": ["car", "teddy bear", ...]
        }
      }
    """

    # -------- initialisation (runs once per container) --------
    def __init__(self, path: str = ""):
        # `path` points to ./ (repo root) where HF already downloaded weights
        self.model = CLIPModel.from_pretrained(path)
        self.processor = CLIPProcessor.from_pretrained(path)

        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model.to(self.device).eval()

        # cache: {prompt → 1×768 tensor on device}
        self.cache: dict[str, torch.Tensor] = {}

    # --------------------- inference --------------------------
    def __call__(self, data):
        payload = data.get("inputs", data)          # unwrap HF envelope

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

        # ---- prompt engineering identical to local code ----
        prompts = [f"a photo of a {p}" for p in names]

        # ---- text embeddings with cache --------------------
        missing = [p for p in prompts if p not in self.cache]
        if missing:
            txt_in = self.processor(text=missing, return_tensors="pt",
                                     padding=True).to(self.device)
            with torch.no_grad():
                txt_emb = self.model.get_text_features(**txt_in)
                txt_emb = txt_emb / txt_emb.norm(dim=-1, keepdim=True)
            for p, e in zip(missing, txt_emb):
                self.cache[p] = e
        txt_feat = torch.stack([self.cache[p] for p in prompts])

        # ---- image preprocessing ---------------------------
        img = Image.open(io.BytesIO(base64.b64decode(img_b64))).convert("RGB")
        img_in = self.processor(images=img, return_tensors="pt").to(self.device)

        with torch.no_grad(), torch.cuda.amp.autocast():
            img_feat = self.model.get_image_features(**img_in)
            img_feat = img_feat / img_feat.norm(dim=-1, keepdim=True)

        # ---- similarity & softmax (same as local) ----------
        probs = (100 * img_feat @ txt_feat.T).softmax(dim=-1)[0].tolist()

        # ---- return sorted list ----------------------------
        return [
            {"label": n, "score": float(p)}
            for n, p in sorted(zip(names, probs), key=lambda x: x[1], reverse=True)
        ]