PointTransformerV3EncoderOptions

Options & Configuration · AiDotNet.ComputerVision.Segmentation.PointCloud

The Point Transformer V3 backbone settings shared by PointTransformerV3Options and SonataOptions.

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SegmentationModelOptions → PointTransformerV3EncoderOptions

Known derived types: PointTransformerV3Options, SonataOptions

For Beginners

These settings decide how big the network is (how many blocks, how wide each stage is, how many attention heads it has) and how it groups points (the voxel size, the patch size, the space-filling curves used to order the points). The defaults of each derived options class are the values from the paper it reproduces.

How It Works

Each array has one entry per stage. The encoder arrays describe stages from finest to coarsest; the decoder arrays describe stages from finest to coarsest as well, one fewer than the encoder, exactly as Pointcept’s dec_* tuples are written.

Properties

Property Summary
IReadOnlyList<ISpaceFillingCurve> SerializationCurves { get; set; } The space-filling curves the blocks cycle through, in order.
bool EnableConditionalPositionalEncoding { get; set; } Whether every block begins with the sparse-convolution xCPE.
bool PreNorm { get; set; } Whether blocks normalise before attention and MLP (pre-norm).
bool ShuffleOrders { get; set; } Whether training shuffles the serialization orders after every pooling.
double DropPathRate { get; set; } Stochastic-depth rate of the deepest block; earlier blocks use a linear ramp from 0.
double GridSize { get; set; } Voxel edge length in metres used to grid the coordinates.
double MlpRatio { get; set; } Ratio of the MLP hidden width to the block width.
int InputChannels { get; set; } Feature channels per point after the xyz coordinates.
int[] DecoderChannels { get; set; } Channels per decoder stage, finest first.
int[] DecoderDepths { get; set; } Blocks per decoder stage, finest first.
int[] DecoderNumHeads { get; set; } Attention heads per decoder stage, finest first.
int[] DecoderPatchSizes { get; set; } Serialized patch size per decoder stage, finest first.
int[] EncoderChannels { get; set; } Channels per encoder stage.
int[] EncoderDepths { get; set; } Blocks per encoder stage.
int[] EncoderNumHeads { get; set; } Attention heads per encoder stage.
int[] EncoderPatchSizes { get; set; } Serialized patch size per encoder stage.
int[] Strides { get; set; } Grid-pooling stride between consecutive encoder stages.