from typing import Any, Dict, List, AnyStr
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="") -> None:
        "Preload all the elements we need at inference."
        self.model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
        self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
        self.path = path
    
    def __call__(self, data: Dict[str, AnyStr]) -> List[Dict[str, AnyStr]]:
        "Run the inference."
        inputs = data.get('inputs')
        text = inputs.get('text')
        imageData = inputs.get('image')
        image = Image.open(BytesIO(base64.b64decode(imageData)))
        inputs = self.processor(text=text, images=image, return_tensors="pt", padding=True)
        outputs = self.model(**inputs)
        image_embeds = outputs.image_embeds.detach().numpy().flatten().tolist()
        text_embeds = outputs.text_embeds.detach().numpy().flatten().tolist()
        logits_per_image = outputs.logits_per_image.detach().numpy().flatten().tolist()
        return {'image_embeddings': image_embeds, 'text_embeddings': text_embeds, 'logits_per_image': logits_per_image}