Struct DiagonalSSM 

pub struct DiagonalSSM { /* private fields */ }

Available on crate feature alloc only.

Non-selective diagonal SSM with fixed A, B, C, and step size.

This is the simplest SSM variant: all parameters are determined at construction time and do not adapt to input content. The hidden state is an N-dimensional vector that evolves through element-wise recurrence.

When to Use

Use DiagonalSSM when:

• You want a simple temporal smoothing/memory layer
• Input-dependent selectivity is not needed
• You need a fast baseline to compare against SelectiveSSM

Example

use irithyll_core::ssm::diagonal::DiagonalSSM;
use irithyll_core::ssm::SSMLayer;

let mut ssm = DiagonalSSM::new(8, 0.1);
let output = ssm.forward(&[1.0]);
assert_eq!(output.len(), 1);

Implementations

impl DiagonalSSM

pub fn new(n_state: usize, delta: f64) -> Self

Create a new non-selective diagonal SSM.

Initializes A via the Mamba strategy (A_n = -(n+1)), sets B and C to ones, D (skip connection) to zero, and pre-computes the discretized matrices using ZOH.

Arguments

• n_state – number of hidden state dimensions (N)
• delta – fixed discretization step size (positive)

Example

use irithyll_core::ssm::diagonal::DiagonalSSM;

let ssm = DiagonalSSM::new(16, 0.01);

pub fn with_params( n_state: usize, delta: f64, b: Vec<f64>, c: Vec<f64>, d_skip: f64, ) -> Self

Create a diagonal SSM with custom B, C vectors and skip connection.

Arguments

• n_state – number of hidden state dimensions
• delta – fixed discretization step size
• b – input projection vector (length n_state)
• c – output projection vector (length n_state)
• d_skip – skip connection weight

Panics

Debug-asserts that b.len() == n_state and c.len() == n_state.

pub fn forward_scalar(&mut self, x: f64) -> f64

Process a single scalar input and return the scalar output.

Updates the hidden state via:

h_n = a_bar_n * h_n + b_bar_factor_n * b_n * x
y = C^T * h + D * x

pub fn n_state(&self) -> usize

Get the number of state dimensions.

pub fn log_a(&self) -> &[f64]

Get the log-A parameters (for inspection/serialization).

pub fn delta(&self) -> f64

Get the fixed step size.

Trait Implementations

impl SSMLayer for DiagonalSSM

fn forward(&mut self, input: &[f64]) -> Vec<f64>

Process one input timestep and return the output vector.

fn state(&self) -> &[f64]

Get a reference to the current hidden state.

fn output_dim(&self) -> usize

Output dimension of this SSM layer.

fn reset(&mut self)

Reset hidden state to zeros, as if no data has been seen.