Struct VariationalAutoencoder

pub struct VariationalAutoencoder { /* private fields */ }

Variational Autoencoder for chess position manifold learning with uncertainty quantification

Implementations

impl VariationalAutoencoder

pub fn new(input_dim: usize, latent_dim: usize, beta: f32) -> Self

pub fn init_network(&mut self, hidden_dims: &[usize]) -> Result<(), String>

Initialize the VAE network with configurable architecture

pub fn forward(&self, x: &Tensor) -> CandleResult<(Tensor, Tensor, Tensor)>

Forward pass through the VAE

pub fn compute_loss( &self, x: &Tensor, reconstruction: &Tensor, mean: &Tensor, logvar: &Tensor, ) -> CandleResult<Tensor>

Compute VAE loss (reconstruction + KL divergence)

pub fn train_step(&mut self, vectors: &Array2<f32>) -> Result<f32, String>

Train the VAE on a batch of position vectors

pub fn encode( &self, vectors: &Array2<f32>, ) -> Result<(Array2<f32>, Array2<f32>), String>

Encode positions to latent space with uncertainty

pub fn sample_latent( &self, mean: &Array2<f32>, logvar: &Array2<f32>, ) -> Result<Array2<f32>, String>

Sample from the latent space

pub fn decode( &self, latent_vectors: &Array2<f32>, ) -> Result<Array2<f32>, String>

Decode from latent space

pub fn encode_with_sampling( &self, vectors: &Array2<f32>, ) -> Result<Array2<f32>, String>

Full encoding pipeline (encode then sample)

pub fn generate(&self, num_samples: usize) -> Result<Array2<f32>, String>

Generate new samples from the learned manifold

pub fn evaluate_reconstruction( &self, vectors: &Array2<f32>, ) -> Result<HashMap<String, f32>, String>

Get reconstruction quality metrics

pub fn latent_dim(&self) -> usize

Get the latent dimensionality

pub fn is_initialized(&self) -> bool

Check if the VAE is initialized