handler.py

from typing import Dict, List, Any
import numpy as np
from transformers import CLIPProcessor, CLIPModel
from PIL import Image
from io import BytesIO
import base64

class EndpointHandler():
    def __init__(self, path=""):
        # Preload all the elements you are going to need at inference.
        self.model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
        self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")


    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
       data args:
            inputs (:obj: `str` | `PIL.Image` | `np.array`)
            kwargs
      Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """

        words = data.pop("words", data)
        imageData = data.pop("image", data)
        image = Image.open(BytesIO(base64.b64decode(imageData)))
        inputs = self.processor(text=words, images=image, return_tensors="pt", padding=True)
        outputs = self.model(**inputs)
        embeddings = outputs.image_embeds.detach().numpy().flatten().tolist()
        return {"embeddings": embeddings}