Struct GruModel 

pub struct GruModel { /* private fields */ }

GRU byte-level predictor with BPTT-10 online training.

Implementations

impl GruModel

pub fn new() -> Self

Create a new GRU model with Xavier-initialized weights and zeroed buffers.

pub fn forward(&mut self, byte: u8)

Forward pass: process one byte, update hidden state, compute output probs.

Call this with the byte that was just OBSERVED. The resulting byte_probs predict the NEXT byte. After forward(), call predict_bit() to get bit probabilities for the next byte.

Also saves the step into the BPTT history ring for train() to use.

pub fn predict_bit(&self, bpos: u8, c0: u32) -> u32

Convert byte probabilities to a bit prediction for the CM MetaMixer.

bpos: bit position 0-7 (0 = MSB). c0: partial byte being built (starts at 1, accumulates bits MSB-first).

Returns: 12-bit probability [1, 4095] of next bit being 1.

pub fn train(&mut self, actual_byte: u8)

Online training with truncated BPTT.

Computes the output-layer gradient for actual_byte, then propagates d_h backwards through up to BPTT_HORIZON stored steps. Gradients are accumulated into pre-allocated buffers and applied in a single weight update at the end. This avoids the weight-corruption from immediate per-step updates that would otherwise occur in BPTT.

Call this BEFORE forward(actual_byte) (matches codec.rs flow). Both encoder and decoder see the same byte sequence so their weights and history buffers evolve identically — parity is preserved.

Trait Implementations

impl Default for GruModel

fn default() -> Self

Returns the “default value” for a type.