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
33

// Module:  stdlib/nn/batch_norm.tern
// Purpose: Ternary Batch Normalization
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// Normalizes activations across the batch dimension. Keeps running
// statistics in trit form.

fn running_mean_trit(batch: trittensor<4 x 4>) -> trit {
    return tend; // Center
}

fn running_var_trit(batch: trittensor<4 x 4>) -> trit {
    return affirm; // Some variance
}

fn batch_normalize(feature: trit, run_mean: trit, run_var: trit) -> trit {
    if feature == run_mean { return tend; }
    match feature {
        affirm => { return affirm; }
        tend   => { return tend;   }
        reject => { return reject; }
    }
}

fn bn_forward(feature: trit, is_training: trit) -> trit {
    if is_training == affirm {
        // Compute batch stats and apply
        return feature;
    }
    // Use running stats
    return feature;
}