from typing import  Dict, List, Any
import torch
from transformers import AutoProcessor, Pix2StructVisionModel
from PIL import Image
import pdb
import requests

MODEL = "google/pix2struct-screen2words-large" 

class EndpointHandler():
    def __init__(self, path=""):
        #self.processor = AutoProcessor.from_pretrained("jasper-lu/pix2struct_embedding")
        #self.model = MarkupLMModel.from_pretrained("jasper-lu/pix2struct_embedding")
        self.processor = AutoProcessor.from_pretrained(MODEL)
        self.processor.image_processor.is_vqa = False
        self.model = Pix2StructVisionModel.from_pretrained(MODEL).cuda()

    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
        url = data.pop("inputs", data)
        device = "cuda"
        image = Image.open(requests.get(url, stream=True).raw)
        inputs = self.processor(images=image, return_tensors="pt").to(device)

        with torch.no_grad():
            outputs = self.model(**inputs)
        
        last_hidden_state = outputs['last_hidden_state']
        embedding = torch.mean(last_hidden_state, dim=1).flatten().tolist()
        return {"embedding": embedding}

"""
handler = EndpointHandler()
output = handler({"inputs": "https://figma-staging-api.s3.us-west-2.amazonaws.com/images/a8c6a0cc-c022-4f3a-9fc5-ac8582c964dd"})
print(output)
"""