SonataOptions

Options & Configuration · AiDotNet.ComputerVision.Segmentation.PointCloud

Configuration options for Sonata, self-supervised point self-distillation of a Point Transformer V3 encoder (Wu et al., CVPR 2025).

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SegmentationModelOptions → PointTransformerV3EncoderOptions → SonataOptions

For Beginners

The pretraining settings only matter when you call Pretrain; the backbone settings decide the network.

How It Works

The defaults are the official pretrain-sonata-v1m1-0-base configuration: a PT-v3m2 encoder-only backbone (depths (3, 3, 3, 12, 3), widths (48, 96, 192, 384, 512), heads (3, 6, 12, 24, 32), patch 1024, drop path 0.3, mask token) fed coordinate, colour and normal (9 channels); OnlineCluster heads of 4096 hidden, 256 embedding and 4096 prototypes on the features up-cast two levels (1088 channels); 2 global views (scale 0.4-1.0) and 4 local views (0.1-0.4) of at most 65536 points; masking of the global views with cube size 0.1 -> 0.4 and ratio 0.3 -> 0.7; teacher temperature 0.04 -> 0.07; student temperature 0.1; loss weights 2/8 (mask), 2/8 (roll mask), 4/8 (unmask); EMA momentum 0.994 -> 1; match radius 0.32 m.

The official configuration has no KoLeo term, so KoLeoWeight is zero by default.

Constructors

Constructor Summary
SonataOptions() Initializes a new instance with the Sonata v1m1 base values.
SonataOptions(SonataOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
double GlobalViewScaleMax { get; set; } Largest global crop fraction.
double GlobalViewScaleMin { get; set; } Smallest global crop, as a fraction of the scene (capped at MaxViewSize).
double GradientClip { get; set; } Gradient-norm clip during pretraining.
double KoLeoWeight { get; set; } Weight of the feature regulariser (KoLeo by default).
double LearningRate { get; set; } Learning rate of the default pretraining optimizer (AdamW).
double LocalViewScaleMax { get; set; } Largest local crop fraction.
double LocalViewScaleMin { get; set; } Smallest local crop fraction.
double MaskJitter { get; set; } Standard deviation in metres of the jitter added to masked coordinates (clipped above at twice it).
double MaskLossWeight { get; set; } Weight of the masked-view loss against the same global view.
double MaskRatioBase { get; set; } Fraction of cubes masked after warm-up.
double MaskRatioStart { get; set; } Fraction of cubes masked at the start of warm-up.
double MaskRatioWarmupRatio { get; set; } Fraction of the pretraining steps spent warming the mask ratio up.
double MaskSizeBase { get; set; } Mask cube edge in metres after warm-up.
double MaskSizeStart { get; set; } Mask cube edge in metres at the start of warm-up.
double MaskSizeWarmupRatio { get; set; } Fraction of the pretraining steps spent warming the mask size up.
double MatchMaxRadius { get; set; } Largest distance in metres at which points of two views are matched.
double MomentumBase { get; set; } EMA momentum of the teacher at the first step (cosine schedule).
double MomentumFinal { get; set; } EMA momentum of the teacher at the last step.
double RollMaskLossWeight { get; set; } Weight of the masked-view loss against the other global view.
double StudentTemperature { get; set; } Student softmax temperature.
double TeacherTemperatureBase { get; set; } Teacher temperature after warm-up.
double TeacherTemperatureStart { get; set; } Teacher temperature at the start of warm-up.
double TeacherTemperatureWarmupRatio { get; set; } Fraction of the pretraining steps spent warming the teacher temperature up.
double UnmaskLossWeight { get; set; } Weight of the local-view loss against the principal global view.
double WeightDecay { get; set; } Weight decay of the default pretraining optimizer.
int GlobalViewCount { get; set; } Global crops per scene; masked copies of them form the masked views.
int HeadEmbedChannels { get; set; } Embedding width of each head before the prototypes.
int HeadHiddenChannels { get; set; } Hidden width of each OnlineCluster head MLP.
int LocalViewCount { get; set; } Local crops per scene.
int MaxViewSize { get; set; } Most points a crop may hold.
int PretrainingEpochs { get; set; } Passes over the scenes per Pretrain call.
int PrototypeCount { get; set; } Number of learnable prototypes per head.
int SinkhornIterations { get; set; } Sinkhorn-Knopp iterations of the default target centering.
int UpCastLevel { get; set; } How many poolings the distillation heads’ features are up-cast through (Sonata: 2, giving 512 + 384 + 192 = 1088 channels).

Methods

Method Summary
void Validate() Throws when a value on this instance cannot produce a working model or pretraining run.