Struct NeuralOdeFunc 

pub struct NeuralOdeFunc {
    pub input_size: usize,
    pub hidden_size: usize,
    pub weights_in: Vec<f64>,
    pub bias_in: Vec<f64>,
    pub weights_hidden: Vec<f64>,
    pub bias_hidden: Vec<f64>,
    pub weights_out: Vec<f64>,
    pub bias_out: Vec<f64>,
}

The dynamics function of a Neural ODE — a small MLP that maps (t, z) to dz/dt.

Architecture: z → tanh(W_in·z + b_in) → tanh(W_h·h + b_h) → W_out·h2 + b_out.

Fields

input_size: usize

Input dimensionality (size of the state vector).

hidden_size: usize

Number of hidden units in each hidden layer.

weights_in: Vec<f64>

Weight matrix from input to first hidden layer (row-major, hidden × input).

bias_in: Vec<f64>

Bias for the first hidden layer (length hidden_size).

weights_hidden: Vec<f64>

Weight matrix from first hidden to second hidden layer (row-major, hidden × hidden).

bias_hidden: Vec<f64>

Bias for the second hidden layer (length hidden_size).

weights_out: Vec<f64>

Weight matrix from second hidden to output (row-major, input × hidden).

bias_out: Vec<f64>

Bias for the output layer (length input_size).

Implementations

impl NeuralOdeFunc

pub fn new(input_size: usize, hidden_size: usize, seed: u64) -> Self

Construct a NeuralOdeFunc with all weights initialised to small random values using a simple linear congruential generator seeded by seed.

pub fn forward(&self, t: f64, z: &[f64]) -> Vec<f64>

Evaluate the ODE right-hand side: dz/dt = f(t, z).

The time t is concatenated to z before the first layer so the network can model non-autonomous dynamics.

pub fn jvp(&self, t: f64, z: &[f64], v: &[f64], eps: f64) -> Vec<f64>

Compute the Jacobian-vector product J·v via forward-mode finite differences.

Used internally by the adjoint method to approximate (∂f/∂z) · v.

pub fn params_flat(&self) -> Vec<f64>

Return all trainable parameters as a flat vector.

Layout: weights_in | bias_in | weights_hidden | bias_hidden | weights_out | bias_out.

pub fn n_params(&self) -> usize

Total number of trainable parameters.

pub fn set_params_flat(&mut self, params: &[f64])

Restore all trainable parameters from a flat vector (same layout as params_flat).

pub fn param_grad_contrib( &self, t: f64, z: &[f64], adj: &[f64], eps: f64, ) -> Vec<f64>

Compute the parameter-gradient contribution at point (t, z) with adjoint vector adj: grad_j = Σ_i adj_i · ∂f_i(t,z)/∂θ_j

Uses central finite differences with step eps.

Trait Implementations

impl Clone for NeuralOdeFunc

fn clone(&self) -> NeuralOdeFunc

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for NeuralOdeFunc

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

Formats the value using the given formatter.