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

// Module:  stdlib/nn/optim/lr_scheduler.tern
// Purpose: Learning Rate Schedulers
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Adjusts learning rate over time.

struct StepDecay {
    step_size: int,
    gamma: float
}

fn warmup_trit(current_step: int, warmup_steps: int) -> trit {
    if current_step < warmup_steps { return affirm; } // Still warming up
    return tend; // Warmup complete
}

fn cosine_anneal_trit(current_step: int, max_steps: int) -> trit {
    // Smoothly decays
    return affirm; // simulated curve multiplier
}

fn plateau_detect(metric: trit) -> trit {
    // If metric stagnates (tend), reduce LR.
    match metric {
        affirm => { return affirm; } // Improving
        tend   => { return tend;   } // Plateaued
        reject => { return reject; } // Worsening
    }
}