Struct GraphGAN 

pub struct GraphGAN { /* private fields */ }

Graph Generative Adversarial Network (GraphGAN) Learns to generate realistic graphs through adversarial training

Implementations

impl GraphGAN

pub fn new( latent_dim: usize, hidden_dim: usize, output_features: usize, use_bias: bool, ) -> Result<Self, TorshError>

Create a new Graph GAN

pub fn generate(&self, num_nodes: usize) -> Result<GraphData, TorshError>

Generate fake graph from random noise

pub fn discriminate(&self, graph: &GraphData) -> Result<f32, TorshError>

Discriminator forward pass (returns real/fake probability in (0, 1))

Errors

Propagates discriminator tensor-operation failures.

pub fn discriminate_logit(&self, graph: &GraphData) -> Result<f32, TorshError>

Discriminator forward pass returning the raw pre-sigmoid logit.

The losses are computed from this value rather than from the sigmoid output: in f32 the sigmoid saturates to exactly 0.0 or 1.0 for logits beyond roughly +-17, and ln(0) would make the loss infinite.

Errors

Propagates discriminator tensor-operation failures.

pub fn generator_loss(&self, num_nodes: usize) -> Result<f32, TorshError>

Train generator (maximize discriminator error)

Computes -log D(G(z)) as softplus(-logit), which is finite for every finite logit.

Errors

Propagates generator/discriminator tensor-operation failures.

pub fn discriminator_loss( &self, real_graph: &GraphData, num_nodes: usize, ) -> Result<f32, TorshError>

Train discriminator (distinguish real from fake)

Computes -log D(real) - log(1 - D(fake)) in the numerically stable binary-cross-entropy-with-logits form softplus(-logit_real) + softplus(logit_fake).

Errors

Propagates generator/discriminator tensor-operation failures.

pub fn generator_parameters(&self) -> Vec<Tensor>

Get generator parameters

pub fn discriminator_parameters(&self) -> Vec<Tensor>

Get discriminator parameters

Trait Implementations

impl Debug for GraphGAN

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

Formats the value using the given formatter.