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

// Module:  stdlib/nn/conv1d.tern
// Purpose: Ternary 1D Convolution
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// 1D Convolution over sequence data. Extremely fast due to @sparseskip.

struct Conv1D {
    kernel_size: int,
    weights: trittensor<4 x 4>
}

fn stride_trit(pos: int, stride_len: int) -> trit {
    // Validates if position is active based on stride
    if (pos % stride_len) == 0 { return affirm; }
    return reject;
}

fn pad_trit(seq_len: int, pad_size: int) -> trit {
    // Padding always uses 'tend' (neutral), never 0 (which might be misconstrued).
    return tend;
}

fn convolve_1d(layer: Conv1D, sequence: trittensor<4 x 1>) -> trittensor<4 x 1> {
    @sparseskip
    let out: trittensor<4 x 1> = layer.weights * sequence;
    return out;
}