SpiralNet<T>

Models & Types · AiDotNet.NeuralNetworks

Implements the SpiralNet++ architecture for mesh-based deep learning.

Inheritance: NeuralNetworkBase<T> → DeclaredModelLayoutBase<T> → GraphModelLayoutBase<T> → SpiralNet<T>

For Beginners

SpiralNet++ is designed for learning from 3D mesh data.

Key concepts:

  • Mesh: A 3D surface made of vertices connected by edges/triangles
  • Spiral ordering: A consistent way to visit vertex neighbors (like a clock hand)
  • Spiral convolution: Apply weights to neighbors in spiral order

How it works:

  1. For each vertex, define a spiral ordering of its neighbors
  2. Gather neighbor features in spiral order
  3. Apply learned weights to the gathered features
  4. Pool vertices to create hierarchical representations
  5. Classify or segment the mesh

Applications:

  • 3D face reconstruction and expression recognition
  • Human body shape analysis
  • Medical surface analysis (organs, bones)
  • CAD model classification

How It Works

SpiralNet++ processes 3D meshes by applying convolutions along spiral sequences of vertex neighbors. This creates translation-equivariant operations on irregular mesh structures without requiring mesh registration.

Reference: “SpiralNet++: A Fast and Highly Efficient Mesh Convolution Operator” by Gong et al.

Example

var options = new SpiralNetOptions { InputFeatures = 3, SpiralLength = 9 };
var vertexFeatures = Tensor<float>.CreateRandom(new[] { 1, 500, 3 });
var trainX = Tensor<float>.CreateRandom(4, 8);
var trainY = Tensor<float>.CreateRandom(4, 2);
var result = new AiModelBuilder<float, Tensor<float>, Tensor<float>>()
    .ConfigureModel(new SpiralNet<float>(options))
    .Build(trainX, trainY);
var output = result.Predict(vertexFeatures);

Constructors

Constructor Summary
SpiralNet() Initializes a new instance of the SpiralNet class with default options.
SpiralNet(SpiralNetOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Initializes a new instance of the SpiralNet class with specified options.
SpiralNet(int numClasses, ILossFunction<T> lossFunction) Initializes a new instance of the SpiralNet class with simple parameters.

Properties

Property Summary
int InputFeatures { get; } Gets the number of input features per vertex.
int NumClasses { get; } Gets the number of output classes for classification.
int SpiralLength { get; } Gets the spiral sequence length.
int[] ConvChannels { get; } Gets the channel configuration for spiral convolution layers.

Methods

Method Summary
List<T> Train(List<Tensor<T>> meshFeatures, List<int[]> spiralIndices, List<int> labels, int epochs, T learningRate) Trains the network on mesh data.
ModelMetadata<T> GetModelMetadata() Gets metadata about this model.
Tensor<T> Forward(Tensor<T> input) Performs a forward pass through the network.
Vector<T> PredictProbabilities(Tensor<T> meshFeatures, int[] meshSpiralIndices) Serializes network-specific data.
int PredictClass(Tensor<T> meshFeatures, int[] meshSpiralIndices) Predicts the class for a single mesh.
void SetMultiResolutionSpiralIndices(List<int[]> spiralIndicesPerLevel) Sets spiral indices for multiple resolution levels (for hierarchical processing).
void SetSpiralIndices(int[] spiralIndices) Sets the spiral indices for the current mesh being processed.
void UpdateParameters(T learningRate) Updates network parameters using the optimizer.