handler.py

from typing import Dict, List, Any
from PIL import Image
from io import BytesIO
from transformers import pipeline
import base64

class EndpointHandler():
    def __init__(self, path=""):
        self.pipeline = pipeline(task="zero-shot-object-detection",model=path, device = 0 ) #device = 0 to use GPU rather than -1 which would be CPU

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
       data args:
            images (:obj:`string`)
            candiates (:obj:`list`)
      Return:
            A :obj:`list`:. The list contains items that are dicts should be liked {"label": "XXX", "score": 0.82}
        """
        inputs = data.pop("inputs", data)

        # decode base64 image to PIL
        image = Image.open(BytesIO(base64.b64decode(inputs['image'])))

        # run prediction one image wit provided candiates
        detector = self.pipeline(images=[image], candidate_labels=inputs["candidate_labels"])
        return detector