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

// Module:  stdlib/nn/loss/bce.tern
// Purpose: Binary Cross-Entropy (BCE) Adapted for Trits
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// BCE typically measures binary classification loss. Here, if the label
// is 'tend', BCE ignores the sample (sparse loss).

fn bce_logit_trit(y_true: trit, logit: trit) -> trit {
    if y_true == tend {
        // Label is uncertain, loss is neutral. This naturally skips the sample.
        return tend; 
    }
    if y_true == logit { return affirm; } // Low loss
    return reject; // High loss
}

fn binary_cross_entropy_trit(y_true: trit[], y_pred: trit[]) -> trit {
    // Averages the BCE per element
    let agg: trit = tend;
    match agg {
        affirm => { return affirm; }
        tend   => { return tend;   }
        reject => { return reject; }
    }
}