PTWM
neural_predictor

ptwm_core::entropy::neural_predictor::net

Tiny online-learning neural predictor: a trainable per-context embedding feeds a ReLU hidden layer and a linear output. Pure integer/fixed-point and fully deterministic, so encode and decode evolve in lockstep — losslessness does NOT depend on the net compressing well. No weights cross the wire.

ptwm_core::entropy::neural_predictor::net

Tiny online-learning neural predictor: a trainable per-context embedding feeds a ReLU hidden layer and a linear output. Pure integer/fixed-point and fully deterministic, so encode and decode evolve in lockstep — losslessness does NOT depend on the net compressing well. No weights cross the wire.

rsmod

ptwm_core::entropy::neural_predictor::net

crates/ptwm-core/src/entropy/neural_predictor/net.rs:1
mod net

Tiny online-learning neural predictor: a trainable per-context embedding feeds a ReLU hidden layer and a linear output. Pure integer/fixed-point and fully deterministic, so encode and decode evolve in lockstep — losslessness does NOT depend on the net compressing well. No weights cross the wire.

rstype

ptwm_core::entropy::neural_predictor::net::Net

crates/ptwm-core/src/entropy/neural_predictor/net.rs:45
pub struct Net

Tiny MLP predictor. `predict` records per-bit activations that `update` consumes, so the two must be called in lockstep (one `update` per `predict`).

coder

Bit-level encode/decode driver tying Net to the shared range coder, coding each byte MSB-first. Self-delimiting: a u32 LE length prefix precedes the coded bits, so decode never depends on the container's decoded_len.

outcome

Typed return for the [huffman]crate::entropy::huffman and [rans]crate::entropy::rans compressors.