Naive implementation of scaled dot product attention.

Normalizes the given weights to be non-negative. If input weights are None, it creates and returns a set of weights of ones.

Prepare time sequence and step size for inference

Backward compatibility wrapper - only applies dual transforms

Apply joint transforms to rgb_image and rgb_image_mask.

Ran this script: out = model(input) loss = out.sum() loss.backward() for name, p in dino_model.named_paramet

Create sinusoidal position embeddings. Args: x: a 1-D Tensor of N indices Returns: an (N, D) Tensor

Connects four neighboring dual vertices to form a quadrilateral. The quadrilaterals are then split into triangles based on the gamma

Align mesh vertices to target point cloud using scale and translation.

Legacy function for compatibility - creates a copy. Use apply_gs_transform_inplace for better performance.

Apply transformation to Gaussian Splatting object IN-PLACE for efficiency. Uses provided backup_data without creating new backups (more effic

Apply ICP transformation matrix to Gaussian scaling and rotation parameters. Returns: Tuple of (R, scale_icp, t) where:

Apply dropout to pointmap embeddings. Drops entire pointmap for selected samples during training or when forced. Whe

Bake texture to a mesh from multiple observations. Args: vertices (np.array): Vertices of the mesh. Shape (V, 3). faces (np.

Return the attention configuration of the model.

Create an additive attention mask for block-diagonal attention. The result is shape [sum_q, sum_kv], with 0.0 in the valid region(s) and

Calculate serialization and partitioning for a set of coordinates. Args: tensor (SparseTensor): The input tensor. window_siz

Calculate serialization and partitioning for a set of coordinates. Args: tensor (SparseTensor): The input tensor. window_siz

Loss function for GS optimization using both RGB and mask supervisions. Args: rendered: GS rendered output with 'color' key and 'alp

Compute a bounding box around a binary mask with optional size adjustment. Args: mask: A 2D binary tensor where non-zero values repr

Compute self-consistency target for shortcut model's self-consistency objective. This method uses a mixed approach where:

Create a 4-channel RGBA image from a 3-channel RGB image and a mask.

construct a dense grid based on resolution

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Create a copy of Gaussian with updated positions.

Crop mesh vertices to only those visible in the mask.

Args: cubes [Vx8] verts index for each cube value [Vx8xM] value to be scattered Operation: reduced[cubes[i][j]][k] +=

Convert normalized intrinsics to absolute pixel coordinates.

Downsample sparse structure coordinates when there are more than max_coords. Downsamples by rescaling coordinates, effectively shrinking the

Process pointmap into window embeddings without positional encoding

Calculate the area covered by the mask's 2D bounding box as a fraction of total image area.

Evaluate spherical harmonics at unit directions using hardcoded SH polynomials. Works with torch/np/jnp. ... Can be 0 or more batch d

e.g. Takes tensor of (a, b, c) and returns a tensor of (a, b, c, 1, 1, ...)

Extract RGB from GS rendering with consistent logic. Args: rendered: GS rendered output dictionary Returns: RGB tensor

Extract target points from MoGe pointmap using visible mask.

For converting latent tokens from generator to cube Args: flat_latent: (B, T, C) Returns: cube: (B,

Gather the data from the leaf children. Args: data (torch.Tensor): the data to gather. The first dimension should be the

Gather the data from the non-leaf children. Args: data (torch.Tensor): the data to gather. The first dimension should be

Apply gaussian blur to the input image. Args: image (torch.Tensor): Image tensor of shape (C, H, W), float32, with values in [0, 1].

Setup GS renderer for mask rendering. Forces square rendering (H=W) to work with unmodified gaussian_render.py.

Apply initial transformation to Gaussian Splatting object. Similar to get_mesh but for GS. Args: Gaussian: Original Gaussian obj

Generate MoGe point cloud from image tensor.

Check if the mask has internal holes or has been split into fragments. This may indicate internal occlusion.

Infer camera intrinsics from a point map. Exact implementation matching moge library's inference logic. Args: points: P