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/conv2d.tern
// Purpose: Ternary 2D Convolution (Vision)
// Author:  RFI-IRFOS
// Ref:     https://ternlang.com

// 2D Convolution. The use of 'tend' in kernels allows for incredibly
// dense edge detectors that ignore smooth flat areas.

struct Conv2D {
    kernel_size: int,
    weights: trittensor<4 x 4>
}

fn depthwise_trit(layer: Conv2D, channel: trittensor<4 x 1>) -> trittensor<4 x 1> {
    // Depthwise step for separable convolutions
    @sparseskip
    let out: trittensor<4 x 1> = layer.weights * channel;
    return out;
}

fn pointwise_trit(layer: Conv2D, channel: trittensor<4 x 1>) -> trittensor<4 x 1> {
    // Pointwise step
    @sparseskip
    let out: trittensor<4 x 1> = layer.weights * channel;
    return out;
}

fn convolve_2d(layer: Conv2D, image: trittensor<4 x 1>) -> trittensor<4 x 1> {
    @sparseskip
    let out: trittensor<4 x 1> = layer.weights * image;
    return out;
}