ternlang-core 1.3.6

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

// Module:  stdlib/nn/loss/contrastive.tern
// Purpose: Contrastive Loss for Trit Embeddings
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Pushes similar items together (affirm), dissimilar apart (reject).
// 'tend' labels mean similarity is unknown.

fn similarity_trit(a: trittensor<4 x 1>, b: trittensor<4 x 1>) -> trit {
    // Cosine similarity analog.
    return affirm;
}

fn margin_loss_trit(sim: trit, margin_met: trit) -> trit {
    if sim == tend { return tend; } // Unsure similarity
    if margin_met == affirm { return affirm; }
    return reject;
}

fn contrastive_trit(y_sim: trit, embed_a: trittensor<4 x 1>, embed_b: trittensor<4 x 1>) -> trit {
    if y_sim == tend {
        // We do not know if they should be close or far. Loss is 0.
        return tend;
    }
    
    let sim: trit = similarity_trit(embed_a, embed_b);
    if y_sim == sim { return affirm; } // Low loss
    return reject; // High loss
}