"""Camera Depth Model configuration: the constructor arguments of the official ``RGBDDepth``
(ByteDance-Seed/manip-as-in-sim-suite @ 757b2ea2, cdm/rgbddepth/dpt.py) and the preprocessing defaults of cdm/infer.py."""

from transformers import PreTrainedConfig


class CameraDepthModelConfig(PreTrainedConfig):
    r"""
    encoder: "vits" | "vitb" | "vitl" | "vitg" (infer.py model_configs; the released models are vitl).
    features / out_channels / use_bn / use_clstoken: the DPT head arguments (infer.py model_configs).
    max_depth: the ``RGBDDepth`` argument (stored; the forward does not use it).
    camera: the depth camera the checkpoint is trained for (None for the base model).
    """

    model_type = "camera_depth_model"

    encoder: str = "vitl"
    features: int = 256
    out_channels: list | tuple = (256, 512, 1024, 1024)
    use_bn: bool = False
    use_clstoken: bool = False
    max_depth: float = 20.0
    camera: str | None = None


__all__ = ["CameraDepthModelConfig"]