inference.py

from transformers import AutoImageProcessor, AutoModel
from typing import Dict

import numpy as np
from matplotlib import cm
from PIL import Image
from torch import Tensor

model = AutoModel.from_pretrained(
    "RGBD-SOD/bbsnet", trust_remote_code=True, cache_dir="model_cache"
)
image_processor = AutoImageProcessor.from_pretrained(
    "RGBD-SOD/bbsnet", trust_remote_code=True, cache_dir="image_processor_cache"
)


def inference(rgb: Image.Image, depth: Image.Image) -> Image.Image:
    rgb = rgb.convert(mode="RGB")
    depth = depth.convert(mode="L")

    preprocessed_sample: Dict[str, Tensor] = image_processor.preprocess(
        {
            "rgb": rgb,
            "depth": depth,
        }
    )

    output: Dict[str, Tensor] = model(
        preprocessed_sample["rgb"], preprocessed_sample["depth"]
    )
    postprocessed_sample: np.ndarray = image_processor.postprocess(
        output["logits"], [rgb.size[1], rgb.size[0]]
    )
    prediction = Image.fromarray(np.uint8(cm.gist_earth(postprocessed_sample) * 255))
    return prediction


if __name__ == "__main__":
    pass