Operations¶

Low-level tensor operations for sparse and segmented data.

Defined in warpconvnet/ops/.

Reductions¶

from warpconvnet.ops.reductions import REDUCTIONS, row_reduction

REDUCTIONS¶

Enum of supported reduction types:

Member	Value	Description
REDUCTIONS.MIN	"min"	Element-wise minimum per segment
REDUCTIONS.MAX	"max"	Element-wise maximum per segment
REDUCTIONS.MEAN	"mean"	Mean per segment
REDUCTIONS.SUM	"sum"	Sum per segment
REDUCTIONS.MUL	"mul"	Product per segment
REDUCTIONS.VAR	"var"	Variance per segment
REDUCTIONS.STD	"std"	Standard deviation per segment
REDUCTIONS.RANDOM	"random"	Random sample per segment

row_reduction¶

row_reduction(
    features: Tensor,    # (N, F) — concatenated features
    row_offsets: Tensor,  # (M+1,) — segment boundaries
    reduction: REDUCTIONS,
    eps: float = 1e-6,
) -> Tensor  # (M, F) — one row per segment

Apply a reduction over contiguous segments of a concatenated feature tensor. row_offsets uses CSR format: segment i spans rows row_offsets[i] to row_offsets[i+1].

Sampling¶

from warpconvnet.ops.sampling import farthest_point_sampling

farthest_point_sampling¶

farthest_point_sampling(
    points: Tensor,   # (N, 3) — packed point cloud
    offsets: Tensor,   # (B+1,) — batch offsets
    K: int,            # number of points to sample per batch item
) -> Tensor  # (B*K,) int32 — global indices of selected points

Iteratively selects the point farthest from all previously selected points, producing a well-spread subset. Runs on GPU via a CUDA kernel.