face.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

_FACE_MODELS = [
    'face_detect_best_n.onnx',
    'face_detect_best_s.onnx',
]
_DEFAULT_FACE_MODEL = _FACE_MODELS[0]


@lru_cache()
def _open_face_detect_model(model_name):
    return _open_onnx_model(hf_hub_download(
        'deepghs/imgutils-models',
        f'face_detect/{model_name}'
    ))


_LABELS = ['head']


def detect_faces(image: ImageTyping, level: str = 's', max_infer_size=1216,
                 conf_threshold: float = 0.25, iou_threshold: float = 0.7) \
        -> 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_face_detect_model(level).run(['output0'], {'images': data})
    return _data_postprocess(output[0], conf_threshold, iou_threshold, old_size, new_size, _LABELS)


def _gr_detect_faces(image: ImageTyping, model_name: str, max_infer_size=1216,
                     conf_threshold: float = 0.25, iou_threshold: float = 0.7):
    ret = detect_faces(image, model_name, max_infer_size, conf_threshold, iou_threshold)
    return detection_visualize(image, ret, _LABELS)