Merge pull request #77 from deepghs/dev/cls

    def _raw_predict(self, image: ImageTyping, model_name: str):

        image = load_image(image, force_background='white', mode='RGB')

        model = self._open_model(model_name)

        batch, channels, height, width = model.get_inputs()[0].shape

        if channels != 3:

            raise RuntimeError(f'Model {model_name!r} required {[batch, channels, height, width]!r}, '

                               f'channels not 3.')  # pragma: no cover

        if isinstance(height, int) and isinstance(width, int):

            input_ = _img_encode(image, size=(width, height))[None, ...]

        else:

            input_ = _img_encode(image)[None, ...]

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

        return output