xnn 0.2.0

A lightweight ML framework with GPU-first architecture
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
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153

//! Linear regression example using gradient descent.
//!
//! Learns y = 2x + 1 from synthetic data.

use xnn::{Context, Error, Tensor};

/// Training hyperparameters.
struct Config {
    samples: u16,
    learning_rate: f32,
    epochs: usize,
}

/// Linear model: y = w*x + b.
struct Model {
    w: Tensor<f32>,
    b: Tensor<f32>,
}

impl Model {
    fn new(ctx: &Context) -> Result<Self, Error> {
        let w = Tensor::constant(ctx, &[1], &[0.0])?;
        let b = Tensor::constant(ctx, &[1], &[0.0])?;
        Ok(Self { w, b })
    }

    fn weights(&self) -> Result<(f32, f32), Error> {
        let w = self.w.to_vec()?[0];
        let b = self.b.to_vec()?[0];
        Ok((w, b))
    }

    /// Forward pass: pred = w * x + b.
    fn forward(&self, x: &Tensor<f32>) -> Result<Tensor<f32>, Error> {
        let wx = x.mul(&self.w)?;
        wx.add(&self.b)
    }

    /// Gradient descent step.
    fn step(
        &mut self,
        x: &Tensor<f32>,
        y: &Tensor<f32>,
        lr: &Tensor<f32>,
    ) -> Result<Tensor<f32>, Error> {
        // Forward: pred = w * x + b
        let pred = self.forward(x)?;

        // diff = pred - y
        let diff = pred.sub(y)?;

        // grad_w = sum(diff * x)
        let diff_x = diff.mul(x)?;
        let grad_w = diff_x.sum_reduce(&[0], false)?;

        // grad_b = sum(diff)
        let grad_b = diff.sum_reduce(&[0], false)?;

        // w = w - lr * grad_w
        let w_update = grad_w.mul(lr)?;
        self.w = self.w.sub(&w_update)?;

        // b = b - lr * grad_b
        let b_update = grad_b.mul(lr)?;
        self.b = self.b.sub(&b_update)?;

        Ok(diff)
    }
}

/// Compute MSE loss from diff tensor.
fn compute_loss(diff: &Tensor<f32>) -> Result<f32, Error> {
    let mse = diff.sqr()?.mean_reduce(&[0])?;
    Ok(mse.to_vec()?[0])
}

/// Generate synthetic training data: y = 2x + 1.
fn generate_data(ctx: &Context, n: u16) -> Result<(Tensor<f32>, Tensor<f32>), Error> {
    let n_f32 = f32::from(n);
    let x_data: Vec<f32> = (0..n).map(|i| f32::from(i) / n_f32).collect();
    let y_data: Vec<f32> = x_data.iter().map(|&x| 2.0 * x + 1.0).collect();
    let x = Tensor::from_slice(ctx, &x_data)?;
    let y = Tensor::from_slice(ctx, &y_data)?;
    Ok((x, y))
}

/// Print inference results.
fn print_inference(w: f32, b: f32) {
    println!("\nInference:");
    println!(
        "{:>6} | {:>10} | {:>10} | {:>8}",
        "x", "predicted", "expected", "error"
    );
    println!("{:-<6}-+-{:-<10}-+-{:-<10}-+-{:-<8}", "", "", "", "");

    for test_x in [0.0f32, 0.25, 0.5, 0.75, 1.0, 2.0] {
        let predicted = w * test_x + b;
        let expected = 2.0 * test_x + 1.0;
        let error = (predicted - expected).abs();
        let error_str = format_error(error);
        println!("{test_x:6.2} | {predicted:10.4} | {expected:10.4} | {error_str:>8}");
    }
}

fn format_error(e: f32) -> String {
    match e {
        e if e < 1e-6 => "<1e-6".into(),
        e if e < 1e-5 => "<1e-5".into(),
        e if e < 1e-4 => "<1e-4".into(),
        e if e < 1e-3 => "<1e-3".into(),
        e if e < 1e-2 => "<1e-2".into(),
        _ => format!("{e:.1e}"),
    }
}

fn main() -> Result<(), Box<dyn core::error::Error>> {
    let ctx = Context::try_default()?;

    let cfg = Config {
        samples: 2048,
        learning_rate: 0.1,
        epochs: 500,
    };

    let (x, y) = generate_data(&ctx, cfg.samples)?;

    let mut model = Model::new(&ctx)?;

    // Learning rate scaled by 2/n for gradient descent
    let n = f32::from(cfg.samples);
    let lr = Tensor::constant(&ctx, &[1], &[2.0 * cfg.learning_rate / n])?;

    println!("Training linear regression: y = wx + b");
    println!("Target: w = 2.0, b = 1.0\n");

    for epoch in 0..cfg.epochs {
        let diff = model.step(&x, &y, &lr)?;

        if epoch % 100 == 0 || epoch == cfg.epochs - 1 {
            let loss = compute_loss(&diff)?;
            let (w, b) = model.weights()?;
            println!("Epoch {epoch:3}: loss = {loss:.6}, w = {w:.4}, b = {b:.4}");
        }
    }

    let (w, b) = model.weights()?;
    println!("\nTarget: w = 2.0000, b = 1.0000");
    println!("Final:  w = {w:.4}, b = {b:.4}");

    print_inference(w, b);

    Ok(())
}