Struct ContinuousNormalizingFlow 

pub struct ContinuousNormalizingFlow {
    pub dynamics: CnfDynamics,
    pub base_mean: Vec<f64>,
    pub base_std: Vec<f64>,
}

Continuous Normalizing Flow model.

Transforms a base distribution p_0 = N(base_mean, diag(base_std^2)) into a target distribution via the ODE dz/dt = f(z, t; θ).

Log-likelihood computation:

log p(x) = log p_0(z_0) + ∫_0^T tr(∂f/∂z) dt

Fields

dynamics: CnfDynamics

ODE dynamics network.

base_mean: Vec<f64>

Mean of the base Gaussian distribution.

base_std: Vec<f64>

Standard deviation of the base Gaussian distribution (diagonal).

Implementations

impl ContinuousNormalizingFlow

pub fn new(z_dim: usize, hidden_dim: usize, n_layers: usize) -> Self

Create a CNF with standard-normal base distribution.

pub fn integrate_forward( &self, z0: &[f64], n_steps: usize, t_start: f64, t_end: f64, ) -> (Vec<f64>, f64)

Euler-integrate the augmented ODE (dz/dt, d_logdet/dt) from t_start to t_end.

Returns (z_T, log_det_jacobian) where log_det_jacobian = ∫ tr(∂f/∂z) dt.

pub fn integrate_backward(&self, x: &[f64], n_steps: usize) -> (Vec<f64>, f64)

Euler-integrate backwards from T=1 to T=0 (inverse direction).

Returns (z_0, log_det) where log_det = -∫_T^0 tr(∂f/∂z) dt.

pub fn log_prob(&self, x: &[f64], n_steps: usize) -> f64

Compute log p(x) = log p_0(z_0) + log_det via backward integration.

pub fn sample(&self, n_steps: usize, rng: &mut StdRng) -> Vec<f64>

Sample a point from the model by sampling from the base distribution and integrating forward.

pub fn train_step( &mut self, x_batch: &[Vec<f64>], n_steps: usize, lr: f64, ) -> f64

One training step: compute mean negative-log-likelihood over x_batch, update via FD gradient.

Returns the mean NLL loss.