dev(narugo): add support for safe metrics

@lru_cache()

def _open_metric_model(model_name):

def _open_metric_model(model_name, safe: bool = False):

    return open_onnx_model(hf_hub_download(

        f'deepghs/imgutils-models',

        f'ccip/{model_name}_metrics.onnx',

        f'ccip/{model_name}_{"safe_" if safe else ""}metrics.onnx',

    ))

_VALID_MODEL_NAMES = [

    'ccip-caformer-5_fp32',

    'ccip-caformer-4_fp32',

    'ccip-caformer-2_fp32',

]

_FeatureOrImage = Union[ImageTyping, np.ndarray]

def get_ccip_similarity(x: _FeatureOrImage, y: _FeatureOrImage,

def get_ccip_similarity(x: _FeatureOrImage, y: _FeatureOrImage, safe: bool = False,

                        size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES) -> float:

    return batch_ccip_similarity([x, y], size, model_name)[0, 1].item()

    return batch_ccip_similarity([x, y], safe, size, model_name)[0, 1].item()

def batch_ccip_similarity(images: Union[np.ndarray, List[_FeatureOrImage]],

def batch_ccip_similarity(images: Union[np.ndarray, List[_FeatureOrImage]], safe: bool = False,

                          size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES):

    input_ = _preprocess_feats(images, size, model_name).astype(np.float32)

    output, = _open_metric_model(model_name).run(['output'], {'input': input_})

    output, = _open_metric_model(model_name, safe=safe).run(['output'], {'input': input_})

    return output

def ccip_clustering(images: MultiImagesTyping, threshold: float = 0.6,

def ccip_clustering(images: MultiImagesTyping, threshold: float = 0.6, safe: bool = True,

                    size: int = 384, model_name: str = _DEFAULT_MODEL_NAMES):

    images = load_images(images, mode='RGB')

    features = []

    if not features:

        return []

    feats = np.stack(features)

    differences = 1 - batch_ccip_similarity(feats, size, model_name)

    differences = 1 - batch_ccip_similarity(feats, safe, size, model_name)

    def _metric(x, y):

        return differences[int(x), int(y)]