Render the octree. Args: octree (Octree): octree extrinsics (torch.Tensor): (4, 4) camera extrinsics

Safely copy a Gaussian object, handling tensors with gradients.

Normalize pointmap using Scale-Shift Invariant (SSI) normalization. Args: pointmap: Pointmap tensor of shape (H, W, 3) or (3, H,

Flip X,Y coordinates from PyTorch3D to OpenCV convention IN-PLACE. Transforms both positions AND rotations consistently with other transform

Get a spatial cache.

Apply inverse remapping to recover original point coordinates.

Register a spatial cache. The spatial cache can be any thing you want to cache. The registery and retrieval of the cache is b

Args: scale: (..., 3) tensor of scale factors rotation: (..., 3, 3) tensor of rotation matrices translation: (..., 3) ten

Apply scaled dot product attention. Args: qkv (torch.Tensor): A [N, L, 3, H, C] tensor containing Qs, Ks, and Vs.

Assumes scale, rotation, translation are of shape: B, K, C --- B: batch size K: number of obj

Compute IoU for GS rendering. Similar to compute_iou but for GS.

Get efficient white colors for mask rendering with GS. Returns colors tensor compatible with colors_overwrite parameter.

Return a normalization layer.

Computes the location of zero-crossings on 'edges_x' using linear interpolation with 'edges_weight'.

Convert the torso of the model to float16.

Extract mask from GS rendering with consistent logic. Returns: mask tensor [H, W]

Mimic xFormers' memory_efficient_attention using PyTorch 2.0 scaled_dot_product_attention.

Args: points_local: (N, 3) tensor of point coordinates colors: (N, 3) tensor of colors

Zero out the parameters of a module and return it.

Drop any conditioning that are tensors Not using _pytree since we actually want to throw them instead of keeping them.

Generate shortcut step sizes d with binary-time schedule. This method ensures deterministic behavior for distributed trainin