anime_face_detector/__init__.py

import pathlib

import torch

from .detector import LandmarkDetector


def get_config_path(model_name: str) -> pathlib.Path:
    assert model_name in ['faster-rcnn', 'yolov3', 'hrnetv2']

    package_path = pathlib.Path(__file__).parent.resolve()
    if model_name in ['faster-rcnn', 'yolov3']:
        config_dir = package_path / 'configs' / 'mmdet'
    else:
        config_dir = package_path / 'configs' / 'mmpose'
    return config_dir / f'{model_name}.py'


def get_checkpoint_path(model_name: str) -> pathlib.Path:
    assert model_name in ['faster-rcnn', 'yolov3', 'hrnetv2']
    if model_name in ['faster-rcnn', 'yolov3']:
        file_name = f'mmdet_anime-face_{model_name}.pth'
    else:
        file_name = f'mmpose_anime-face_{model_name}.pth'

    model_dir = pathlib.Path(torch.hub.get_dir()) / 'checkpoints'
    model_dir.mkdir(exist_ok=True, parents=True)
    model_path = model_dir / file_name
    if not model_path.exists():
        url = f'https://github.com/hysts/anime-face-detector/releases/download/v0.0.1/{file_name}'
        torch.hub.download_url_to_file(url, model_path.as_posix())

    return model_path


def create_detector(face_detector_name: str = 'yolov3',
                    landmark_model_name='hrnetv2',
                    device: str = 'cuda:0',
                    flip_test: bool = True,
                    box_scale_factor: float = 1.1) -> LandmarkDetector:
    assert face_detector_name in ['yolov3', 'faster-rcnn']
    assert landmark_model_name in ['hrnetv2']
    detector_config_path = get_config_path(face_detector_name)
    landmark_config_path = get_config_path(landmark_model_name)
    detector_checkpoint_path = get_checkpoint_path(face_detector_name)
    landmark_checkpoint_path = get_checkpoint_path(landmark_model_name)
    model = LandmarkDetector(landmark_config_path,
                             landmark_checkpoint_path,
                             detector_config_path,
                             detector_checkpoint_path,
                             device=device,
                             flip_test=flip_test,
                             box_scale_factor=box_scale_factor)
    return model