handler.py

import base64
from io import BytesIO
import os
from typing import Dict, List, Any
import cv2
import groundingdino
from groundingdino.util.inference import load_model, load_image, predict, annotate
import tempfile

# /app
HOME = os.getcwd()

# /opt/conda/lib/python3.9/site-packages/groundingdino
PACKAGE_HOME = os.path.dirname(groundingdino.__file__)
CONFIG_PATH = os.path.join(PACKAGE_HOME, "config", "GroundingDINO_SwinT_OGC.py")

class EndpointHandler():
    def __init__(self, path):
        # Preload all the elements you are going to need at inference.

        self.model = load_model(CONFIG_PATH, os.path.join(path, "weights", "groundingdino_swint_ogc.pth"))

        self.box_threshold = 0.35
        self.text_threshold = 0.25

    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")
        image_base64 = inputs.pop("image")
        prompt = inputs.pop("prompt")
        
        image_data = base64.b64decode(image_base64)

        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=True) as f:
            f.write(image_data)
            image_source, image = load_image(f.name)
            boxes, logits, phrases = predict(
                model=self.model,
                image=image,
                caption=prompt,
                box_threshold=self.box_threshold,
                text_threshold=self.text_threshold
            )
            annotated_frame = annotate(image_source=image_source, boxes=boxes, logits=logits, phrases=phrases)
            _, annotated_image = cv2.imencode(".jpg", annotated_frame)
            annotated_image_b64 = base64.b64encode(annotated_image).decode("utf-8")
            num_found = boxes.size[0]

            return [{
                "image": annotated_image_b64,
                "prompt": prompt,
                "num_found": num_found,
            }]