censor.py

from functools import lru_cache
from typing import List, Tuple

from huggingface_hub import hf_hub_download
from imgutils.data import ImageTyping, load_image, rgb_encode

from onnx_ import _open_onnx_model
from plot import detection_visualize
from yolo_ import _image_preprocess, _data_postprocess

_CENSOR_MODELS = [
    'censor_detect_v0.9_s',
    'censor_detect_v0.8_s',
    'censor_detect_v0.7_s',
]
_DEFAULT_CENSOR_MODEL = _CENSOR_MODELS[0]


@lru_cache()
def _open_censor_detect_model(model_name):
    return _open_onnx_model(hf_hub_download(
        f'deepghs/anime_censor_detection',
        f'{model_name}/model.onnx'
    ))


_LABELS = ['nipple_f', 'penis', 'pussy']


def detect_censors(image: ImageTyping, model_name: str, max_infer_size=640,
                   conf_threshold: float = 0.25, iou_threshold: float = 0.5) \
        -> List[Tuple[Tuple[int, int, int, int], str, float]]:
    image = load_image(image, mode='RGB')
    new_image, old_size, new_size = _image_preprocess(image, max_infer_size)

    data = rgb_encode(new_image)[None, ...]
    output, = _open_censor_detect_model(model_name).run(['output0'], {'images': data})
    return _data_postprocess(output[0], conf_threshold, iou_threshold, old_size, new_size, _LABELS)


def _gr_detect_censors(image: ImageTyping, model_name: str, max_infer_size=640,
                       conf_threshold: float = 0.25, iou_threshold: float = 0.5):
    ret = detect_censors(image, model_name, max_infer_size, conf_threshold, iou_threshold)
    return detection_visualize(image, ret, _LABELS)