topos 0.13.0

An autodiff compiler stack in Rust.
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

use crate::{Shape, Tape, Tensor};

use super::max_pool;

#[test]
fn max_pool_takes_window_maxima() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::new(
        [1, 1, 4, 4],
        (1..=16).map(|v| v as f64).collect::<Vec<_>>(),
    ));

    let pooled = max_pool(input, 2, 2);
    assert_eq!(pooled.shape(), Shape::new([1, 1, 2, 2]));

    let pooled = pooled.symbol();
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    assert_eq!(run.of(pooled).to_vec(), &[6.0, 8.0, 14.0, 16.0]);
}

#[test]
fn max_pool_routes_the_gradient_to_the_maximum() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::new(
        [1, 1, 4, 4],
        (1..=16).map(|v| v as f64).collect::<Vec<_>>(),
    ));

    let pooled = max_pool(input, 2, 2);
    let loss = pooled.sum();

    let (loss, input) = (loss.symbol(), input.symbol());
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    let gradients = run.backward(loss);
    let mut expected = vec![0.0; 16];
    for position in [5, 7, 13, 15] {
        expected[position] = 1.0;
    }
    assert_eq!(gradients.of(input).to_vec(), expected);
}

#[test]
fn max_pool_ties_route_to_the_earliest_lane() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::filled([1, 1, 2, 2], 5.0_f64));

    let pooled = max_pool(input, 2, 2);
    let loss = pooled.sum();

    let (pooled, loss, input) = (pooled.symbol(), loss.symbol(), input.symbol());
    let network = tape.into_network();
    let run = network.forward(&network.parameters(), []);
    assert_eq!(run.of(pooled).to_vec(), &[5.0]);

    // All four window elements tie; the left-biased `maximum` fold
    // hands the whole gradient to the first lane in window order.
    let gradients = run.backward(loss);
    assert_eq!(gradients.of(input).to_vec(), &[1.0, 0.0, 0.0, 0.0]);
}

#[test]
#[should_panic(expected = "must be rank 4")]
fn pooling_rejects_non_image_input() {
    let tape: Tape<f64> = Tape::new();
    let input = tape.leaf(Tensor::filled([1, 4, 4], 0.0_f64));
    max_pool(input, 2, 2);
}