Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing: 1) Uses the first position (index=0) for the first frame onl

Convert tensor to numpy array and validate shape.

Convert ImageList back to PIL Image(s). Returns: PIL.Image.Image or tuple: Single PIL Image if list contains one image,

Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy). batch: list, tuple, dict of tensors or other things device: pytor

Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy). batch: list, tuple, dict of tensors or other things device: pyto

Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy). batch: list, tuple, dict of tensors or other things device: pyto

Output a (H,W,2) array of int32 with output[j,i,0] = i + origin[0] output[j,i,1] = j + origin[1]

Output a (H,W,2) array of int32 with output[j,i,0] = i + origin[0] output[j,i,1] = j + origin[1]

modified from https://github.com/MoonshotAI/Moonlight/blob/master/examples/toy_train.py#L49 Major change: G is [b, d, d] rather than [d, d]

Newton-Schulz orthogonalisation. Args: G: [b, d, d] input matrices to orthogonalise. steps: number of Newton-Schulz iterations.

Initializes the position generator with an empty cache.

Initializes the position generator with an empty cache.

Initializes the position generator with an empty cache.

Make R3's vendored packages win over benchmark/models/depth_anything_3.

Build the compatibility scalar confidence from split logits.

Align depth map to input extrinsics

Temporarily manage sys.path and sys.modules, isolate amb3r dependencies such as vggt/dust3r/croco. amb3r dependency chain: amb3r_root/amb3r

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Apply inverse logarithm transform: sign(y) * (exp(|y|) - 1) Args: input_tensor: Input tensor

Rescale accumulated online pose state after multi-frame metric bootstrap.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Apply positional embedding to tensor x.

Rebuild point cloud geometry from current state.

Helper function to get camera parameters based on camera type. Args: fidx: Frame index intrinsics: Camera intrinsic paramete

Helper function to get camera parameters based on camera type. Args: fidx: Frame index intrinsics: Camera intrinsic paramete

Build the result dict from combined outputs. Args: c2w: (N, 4, 4) cam2world float64 points: (N, H, W, 3) world points

(N, 3, 4) c2w -> list of (4, 4) SE3 numpy matrices.

Convert OpenCV c2w (4x4 or 3x4) to a Viser camera pose.

Generate a unique color for the i-th camera (uint8 RGB).

Return camera-local movement axes in world coordinates.

Clean up old completed/failed tasks to prevent memory buildup.

Compute mean optical flow magnitude induced by camera motion. Projects current frame pixels into the last keyframe camera using the current d

Compute Q, K, V and learning rates for input x. Returns: Tuple of (q, k, v, lr0, lr1, lr2) all appropriately shaped

numpy fallback (used when evo is not installed).

Convert raw model output to Prediction object.