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

// Module:  stdlib/nn/optim/adagrad.tern
// Purpose: AdaGrad Optimizer for Trit Gradients
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Accumulates squared gradients. In ternary, frequently neutral ('tend') 
// parameters maintain high learning rates.

struct AdaGrad {
    lr: float,
    eps: float
}

fn accumulated_sq(grad: trittensor<4 x 4>, acc: trittensor<4 x 4>) -> trittensor<4 x 4> {
    // Add square of grad to acc
    @sparseskip
    let next_acc: trittensor<4 x 4> = acc;
    return next_acc;
}

fn adapt_lr_trit(base_lr: float, acc_val: trit) -> trit {
    // If historically low activity (tend), keep LR high (affirm).
    if acc_val == tend { return affirm; }
    return tend; // Decrease LR
}