Zinyn - ML and AI made possible for fun

Goal

Create a framework for training DNNs in Zig (using hardware accelerators) and compare results

Why?

To quote a wise man: "For the joy of programming!"

Features

Generic tensor using comptime
Binops such as add, sub, mul, div
General tensor ops (clamp, reshape, sum, min, max...)
SGD optimizer
Common loss functions such as mse and mae + 3 more
Activation functions
Simple layers
Autograd
Computational graph
Hardware acceleration support

Install (Per 0.14 beta)

Run this command in the parent directory of your project

zig fetch --save git+https://github.com/Jafagervik/zybel.git

Then add these lines to build.zig before b.installArtifact(exe)

const zybel = b.dependency("zybel", .{});

exe.root_module.addImport("zybel", zybel.module("zybel"));

Example

const std = @import("std");

const zb = @import("zybel");
const Tensor = zb.Tensor;
const TF32 = Tensor(f32);

pub fn main() !void {
    var gpa = std.heap.GeneralPurposeAllocator(.{}){};
    const allocator = gpa.allocator();
    defer _ = gpa.deinit();

    // Creates a F32 Tensor of shape (1, 3, 3)
    var t: TF32 = try TF32.ones(allocator, &[_]u32{ 1, 3, 3 });
    defer t.deinit();

    std.debug.print("First value is {d:.2}\n", .{t.getFirst()});

    bon.setVal(0, 2.0);

    std.debug.print("First value is now {d:.2}\n", .{t.getFirst()});

    // Prints input about the tensor
    t.print();

    std.debug.print("Sum is {d:.2}\n", .{t.sum()});
}

Contribute

Hell yea! Feel free to add suggestions/corrections. I am a mere mortal, not a god