MeshPoolLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

Implements mesh pooling via edge collapse for MeshCNN-style networks.

Inheritance: LayerBase<T> → MeshPoolLayer<T>

Implements: IShapeContract

For Beginners

Just like max pooling shrinks an image by combining pixels, mesh pooling shrinks a mesh by combining edges. The layer learns which edges are less important and removes them, simplifying the mesh while preserving important features.

Key concepts:

  • Edge collapse: Remove an edge by merging its two vertices into one
  • Importance score: Learned value indicating how important each edge is
  • Target edges: Number of edges to keep after pooling

The process:

  1. Compute importance scores for all edges using current features
  2. Sort edges by importance (lowest first)
  3. Collapse least important edges until target count is reached
  4. Update adjacency information for remaining edges

How It Works

MeshPoolLayer reduces the number of edges in a mesh by collapsing edges based on learned importance scores. This is analogous to pooling in image CNNs but operates on the mesh structure.

Reference: “MeshCNN: A Network with an Edge” by Hanocka et al., SIGGRAPH 2019

Constructors

Constructor Summary
MeshPoolLayer(int inputChannels, int targetEdges, int numNeighbors) Initializes a new instance of the MeshPoolLayer class.

Properties

Property Summary
bool SupportsTraining { get; } Gets a value indicating whether this layer supports training.
int InputChannels { get; set; } Gets the number of input feature channels per edge.
int TargetEdges { get; set; } Gets the target number of edges after pooling.
int[] RemainingEdgeIndices { get; set; } Gets or sets the edge indices that remain after pooling.
int[] UpdatedAdjacency { get; set; } Gets or sets the updated edge adjacency after pooling.

Methods

Method Summary
Dictionary<string, string> GetParameterRoles() Returns parameter roles for per-role optimizer configuration (e.g., weight decay exemption for biases).
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
IReadOnlyList<Tensor<T>> GetTrainableParameters() Returns all trainable parameter tensors marked with [TrainableParameter].
Tensor<T> GetBiases() Gets the bias tensor (null for this layer).
Tensor<T> GetWeights() Gets the importance weights tensor.
void ResetState() Resets the cached state from forward/backward passes.
void SetEdgeAdjacency(int[] edgeAdjacency) Sets the edge adjacency information for the current mesh.
void SetTrainableParameters(IReadOnlyList<Tensor<T>> parameters) Replaces trainable parameter tensors (e.g., with ParameterBuffer views).
void UpdateParameters(T learningRate) Updates the layer parameters using computed gradients.
void ZeroGrad() Clears all gradient fields discovered by convention ({paramName}Gradient).