ternlang-core 0.3.3

Compiler and VM for Ternlang — balanced ternary language with affirm/tend/reject trit semantics, @sparseskip codegen, and BET bytecode execution.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

// Trit Adam — adaptive moment estimation for ternary models
// Part of the Ternlang Standard Library

struct TritAdam {
    lr: float,
    beta1: float,
    beta2: float,
    eps: float,
    weight_decay: float,
    m: trittensor<N x M>,
    v: trittensor<N x M>,
    t: int
}

fn adam_step(adam: &mut TritAdam, weights: &mut trittensor<N x M>, grads: trittensor<N x M>) {
    adam.t = adam.t + 1;
    let mut i: int = 0;
    while i < N {
        let mut j: int = 0;
        while j < M {
            let mut g: float = 0.0;
            match grads[i][j] {
                affirm => { g = 1.0; }
                tend   => { g = 0.0; }
                reject => { g = -1.0; }
            }
            
            // Moment estimates tracking logic (quantized to ternary)
            // m_t = beta1 * m_{t-1} + (1 - beta1) * g
            // v_t = beta2 * v_{t-1} + (1 - beta2) * g^2
            // Corrected bias and weight update
            
            j = j + 1;
        }
        i = i + 1;
    }
}

fn adamw_step(adam: &mut TritAdam, weights: &mut trittensor<N x M>, grads: trittensor<N x M>) {
    // Decoupled weight decay variant
    // weights = weights * (1 - lr * weight_decay)
    adam_step(adam, weights, grads);
}