Class PredictorParams

src/sharp/models/params.py:164–203

Parameters for predictors with default values.

@dataclasses.dataclass
class PredictorParams:
    """Parameters for predictors with default values."""

    # Parameters for submodules.
    initializer: InitializerParams = dataclasses.field(default_factory=InitializerParams)
    monodepth: MonodepthParams = dataclasses.field(default_factory=MonodepthParams)
    monodepth_adaptor: MonodepthAdaptorParams = dataclasses.field(
        default_factory=MonodepthAdaptorParams
    )
    gaussian_decoder: GaussianDecoderParams = dataclasses.field(
        default_factory=GaussianDecoderParams
    )
    # How to align depth map (only relevant for RGBGaussianPredictor).
    depth_alignment: AlignmentParams = dataclasses.field(default_factory=AlignmentParams)

    # Selectively reduce learning rate for different properties.
    delta_factor: DeltaFactor = dataclasses.field(default_factory=DeltaFactor)
    # The maximum scale of Gaussians relative to initial scale.
    max_scale: float = 10.0
    # The minimum scale of Gaussians relative to initial scale.
    min_scale: float = 0.0
    # Which normalization to use in prediction head.
    norm_type: NormLayerName = "group_norm"
    # How many groups to use for group normalization.
    norm_num_groups: int = 8
    # Whether to use predicted mean to sample triplane features.
    use_predicted_mean: bool = False
    # Which activation function to use for colors / opacities.
    color_activation_type: math_utils.ActivationType = "sigmoid"
    opacity_activation_type: math_utils.ActivationType = "sigmoid"
    # Colorspace of the renderer ("linearRGB" or "sRGB").
    color_space: ColorSpace = "linearRGB"
    # A small value to avoid ill-conditioned splats
    low_pass_filter_eps: float = 1e-2
    # How many layer of depth does monodepth model predict.
    num_monodepth_layers: int = 2
    # Whether to sort the monodepth output (for two layer monodepth).
    sorting_monodepth: bool = False
    # Whether to account the z offsets for estimating base scale.
    base_scale_on_predicted_mean: bool = True