SinkhornKnoppCentering<T>

Models & Types · AiDotNet.ComputerVision.Segmentation.PointCloud

Sinkhorn-Knopp centering (Caron et al. 2020; Sonata sinkhorn_knopp): alternately normalises the prototypes and the samples of exp(scores / temperature) so every prototype receives the same total mass and every sample’s row is a distribution.

Implements: ISonataTargetCentering<T>

How It Works

Runs in single precision regardless of T, as Sonata does (feat = feat.float()): exp(1 / 0.04) is beyond half precision. The iteration runs in log space: each normalisation is a log-softmax along its axis minus the log of the axis’ target mass, and the result is exponentiated once at the end. That is algebraically Sonata’s linear update, but a prototype whose every exp(score / temperature) underflows to zero no longer divides 0 by 0 into NaN. The result carries no gradient: it is the teacher’s target.

Constructors

Constructor Summary
SinkhornKnoppCentering(int iterations) Creates the centering with the given number of iterations.

Properties

Property Summary
int Iterations { get; } The number of row/column normalisation rounds.

Methods

Method Summary
Tensor<T> Center(Tensor<T> teacherScores, double temperature)