Struct GCNConv 

pub struct GCNConv { /* private fields */ }

Graph Convolutional Network layer (Kipf & Welling, 2017).

Implements the propagation rule: H^(l+1) = σ(D̃^(-1/2) Ã D̃^(-1/2) H^(l) W^(l))

Where:

• Ã = A + I (adjacency with self-loops)
• D̃ = degree matrix of Ã
• W = learnable weight matrix
• σ = activation function (ReLU by default)

Example

use aprender::nn::gnn::{GCNConv, AdjacencyMatrix};
use aprender::autograd::Tensor;

let gcn = GCNConv::new(64, 32);
let x = Tensor::new(&vec![0.1; 3 * 64], &[3, 64]);
let adj = AdjacencyMatrix::from_edge_index(&[[0, 1], [1, 2]], 3).add_self_loops();
let out = gcn.forward(&x, &adj);
assert_eq!(out.shape(), &[3, 32]);

Implementations

impl GCNConv

pub fn new(in_features: usize, out_features: usize) -> Self

Create a new GCN layer.

Arguments

• in_features - Input feature dimension per node
• out_features - Output feature dimension per node

pub fn without_bias(self) -> Self

Disable bias.

pub fn without_self_loops(self) -> Self

Disable automatic self-loop addition.

pub fn without_normalize(self) -> Self

Disable normalization.

pub fn in_features(&self) -> usize

Get input feature dimension.

pub fn out_features(&self) -> usize

Get output feature dimension.

pub fn forward(&self, x: &Tensor, adj: &AdjacencyMatrix) -> Tensor

Forward pass: X’ = D^(-1/2) A D^(-1/2) X W + b

Arguments

• x - Node features [num_nodes, in_features]
• adj - Adjacency matrix

Returns

Output features [num_nodes, out_features]

pub fn weight(&self) -> &Tensor

Get weight tensor (for inspection/serialization).

pub fn bias(&self) -> Option<&Tensor>

Get bias tensor (for inspection/serialization).

Trait Implementations

impl Clone for GCNConv

fn clone(&self) -> GCNConv

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for GCNConv

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.