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

// Module:  stdlib/nn/loss/focal.tern
// Purpose: Ternary Focal Loss
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Focal loss emphasizes hard examples. In ternary, 'tend' predictions
// are considered hard examples because the network is uncertain.

fn modulating_factor(y_true: trit, y_pred: trit) -> trit {
    // If prediction is tend, we highly modulate (focus) on it.
    if y_pred == tend {
        return affirm; // High focus
    }
    if y_true == y_pred {
        return reject; // Low focus
    }
    return tend; // Medium focus
}

fn hard_example_weight(y_pred: trit) -> trit {
    if y_pred == tend { return affirm; } // Max weight
    return tend;
}

fn focal_trit(y_true: trit[], y_pred: trit[]) -> trit {
    let state: trit = tend;
    match state {
        affirm => { return affirm; }
        tend   => { return tend;   }
        reject => { return reject; }
    }
}