handler.py

from typing import Dict, Any
from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
from PIL import Image
import requests
from io import BytesIO
import torch

class EndpointHandler():
    def __init__(self, path=""):
        # Ładowanie modelu i procesora
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model = AutoModelForZeroShotObjectDetection.from_pretrained(path).to(self.device)
        self.processor = AutoProcessor.from_pretrained(path)

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        # Sprawdzamy, czy dane wejściowe zawierają wymagane pola
        if "inputs" not in data:
            return {"error": "Payload must contain 'inputs' key with 'image' and 'text'."}
        
        inputs = data["inputs"]
        if "image" not in inputs or "text" not in inputs:
            return {"error": "Payload must contain 'image' (base64 or URL) and 'text' (queries)."}

        # Pobieramy obraz (URL lub Base64)
        image_data = inputs["image"]
        if image_data.startswith("http"):  # URL
            response = requests.get(image_data)
            image = Image.open(BytesIO(response.content))
        else:
            return {"error": "Handler currently supports only URL-based images."}

        # Pobieramy tekst zapytań
        text_queries = inputs["text"]
        if isinstance(text_queries, list):
            text_queries = ". ".join([t.lower().strip() + "." for t in text_queries])

        # Przygotowujemy dane wejściowe
        processed_inputs = self.processor(images=image, text=text_queries, return_tensors="pt").to(self.device)
        
        # Przeprowadzamy inferencję
        with torch.no_grad():
            outputs = self.model(**processed_inputs)

        # Post-process wyników
        results = self.processor.post_process_grounded_object_detection(
            outputs,
            processed_inputs.input_ids,
            box_threshold=0.4,
            text_threshold=0.3,
            target_sizes=[image.size[::-1]]
        )

        # Zwracamy wyniki
        return {"detections": results}