Struct TernaryMLP 

pub struct TernaryMLP {
    pub w1: TritMatrix,
    pub w2: TritMatrix,
    pub in_features: usize,
    pub hidden_size: usize,
    pub out_features: usize,
}

A 2-layer ternary multi-layer perceptron.

Architecture: input (in_features) → hidden (hidden_size) → output (out_features)

All weights are ternary {-1, 0, +1}. Forward pass uses sparse_matmul. No bias terms (ternary bias adds nothing that weight magnitude can’t cover).

Fields

w1: TritMatrix

w2: TritMatrix

in_features: usize

hidden_size: usize

out_features: usize

Implementations

impl TernaryMLP

pub fn new(w1: TritMatrix, w2: TritMatrix) -> Self

Construct from pre-quantized weight matrices.

pub fn from_f32( in_features: usize, hidden_size: usize, out_features: usize, w1_f32: &[f32], w2_f32: &[f32], ) -> Self

Initialise from f32 weight slices using BitNet threshold quantization.

pub fn forward(&self, input: &TritMatrix) -> (TritMatrix, usize, usize)

Forward pass: input [1 × in_features] → output [1 × out_features].

Returns (output_row, layer1_skips, layer2_skips).

pub fn predict(&self, input: &TritMatrix) -> usize

Classify a single input row: returns the column index of the max activated output (most +1, breaking ties by column index).

pub fn layer1_sparsity(&self) -> f64

pub fn layer2_sparsity(&self) -> f64

pub fn forward_logits(&self, input: &[f32]) -> Vec<f32>

F32 forward pass: returns raw f32 logits (no final ternary clipping).

Uses quantized {-1,0,+1} weights but accumulates in f32, which makes the output suitable for softmax / cross-entropy in perplexity evaluation.

input — flat f32 slice of length in_features