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 """ inputs = data.pop("inputs", data) words = inputs["words"] imageData = inputs["image"] 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}