nuvec-core 0.3.0

Lazy array graph, primitives, dtypes and evaluation for nuvec
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

//! Custom transform rules and gradient checkpointing.

use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};

use nuvec_core::*;

fn sum_all(a: &Array) -> Result<Array> {
    sum(a, &[], false)
}

fn values(a: &Array) -> Vec<f32> {
    a.to_vec::<f32>().unwrap()
}

/// Equal up to a few ulps (Miri perturbs transcendental functions slightly).
fn assert_close(got: &Array, expected: &Array) {
    for (g, e) in values(got).iter().zip(values(expected)) {
        assert!((g - e).abs() <= 1e-5 * e.abs().max(1.0), "{g} vs {e}");
    }
}

#[test]
fn custom_rules_are_used() {
    let jvp_calls = Arc::new(AtomicUsize::new(0));
    let vmap_calls = Arc::new(AtomicUsize::new(0));
    let (j, v) = (jvp_calls.clone(), vmap_calls.clone());
    let square = custom_function(|xs| Ok(vec![multiply(&xs[0], &xs[0])?]))
        .with_vjp(|_, _, cot| Ok(vec![multiply(&cot[0], 10)?]))
        .with_jvp(move |_, t| {
            j.fetch_add(1, Ordering::SeqCst);
            Ok(vec![multiply(&t[0], 100)?])
        })
        .with_vmap(move |xs, axes| {
            v.fetch_add(1, Ordering::SeqCst);
            Ok((vec![multiply(&xs[0], &xs[0])?], axes[0].unwrap_or(0)))
        });
    let x = Array::from_slice(&[1.0f32, 2.0], &[2]).unwrap();
    let out = square.call(std::slice::from_ref(&x)).unwrap();
    assert_eq!(values(&out[0]), [1.0, 4.0]);

    let s = square.clone();
    let g = grad(move |x: &Array| sum_all(&s.call(std::slice::from_ref(x))?[0]));
    assert_eq!(values(&g(&x).unwrap()), [10.0, 10.0]);

    let s = square.clone();
    let (_, t) = jvp(
        |xs| s.call(xs),
        std::slice::from_ref(&x),
        std::slice::from_ref(&x),
    )
    .unwrap();
    assert_eq!(values(&t[0]), [100.0, 200.0]);
    assert_eq!(jvp_calls.load(Ordering::SeqCst), 1);

    let s = square.clone();
    let batched = vmap(move |xs| s.call(xs), &[], &[])(std::slice::from_ref(&x)).unwrap();
    assert_eq!(values(&batched[0]), [1.0, 4.0]);
    assert_eq!(vmap_calls.load(Ordering::SeqCst), 1);
}

#[test]
fn default_rules_transform_the_function() {
    let body = |xs: &[Array]| Ok(vec![multiply(&sin(&xs[0])?, &xs[1])?, exp(&xs[0])?]);
    let f = custom_function(body);
    let x = Array::from_slice(&[0.5f32, 1.0, 1.5], &[3]).unwrap();
    let y = Array::from_slice(&[2.0f32, 3.0, 4.0], &[3]).unwrap();
    let cots = [
        Array::from_slice(&[1.0f32, 1.0, 1.0], &[3]).unwrap(),
        Array::from_slice(&[0.5f32, 0.5, 0.5], &[3]).unwrap(),
    ];
    let primals = [x.clone(), y.clone()];
    let (_, custom) = vjp(|xs| f.call(xs), &primals, &cots).unwrap();
    let (_, plain) = vjp(body, &primals, &cots).unwrap();
    for (c, p) in custom.iter().zip(&plain) {
        assert_close(c, p);
    }
    let (_, custom) = jvp(
        |xs| f.call(xs),
        &primals,
        &[cots[0].clone(), cots[0].clone()],
    )
    .unwrap();
    let (_, plain) = jvp(body, &primals, &[cots[0].clone(), cots[0].clone()]).unwrap();
    for (c, p) in custom.iter().zip(&plain) {
        assert_close(c, p);
    }
    let batched = vmap(|xs| f.call(xs), &[Some(0), None], &[])(&[x.clone(), Array::scalar(2.0f32)]);
    let plain = vmap(body, &[Some(0), None], &[])(&[x, Array::scalar(2.0f32)]);
    for (c, p) in batched.unwrap().iter().zip(&plain.unwrap()) {
        assert_close(c, p);
    }
}

#[test]
fn wrong_rule_results_are_errors() {
    let x = Array::from_slice(&[1.0f32], &[1]).unwrap();
    let f = custom_function(|xs| Ok(vec![xs[0].clone()])).with_vjp(|_, _, _| Ok(vec![]));
    let g = grad(move |x: &Array| sum_all(&f.call(std::slice::from_ref(x))?[0]));
    assert!(matches!(g(&x), Err(Error::InvalidArgument(_))));
    let f =
        custom_function(|xs| Ok(vec![xs[0].clone()])).with_jvp(|_, t| Ok(vec![t[0].clone(); 2]));
    let r = jvp(
        |xs| f.call(xs),
        std::slice::from_ref(&x),
        std::slice::from_ref(&x),
    );
    assert!(matches!(r, Err(Error::InvalidArgument(_))));
    let empty = custom_function(|_| Ok(vec![]));
    assert!(matches!(empty.call(&[x]), Err(Error::InvalidArgument(_))));
}

#[test]
fn checkpoint_recomputes_with_equal_gradients() {
    let calls = Arc::new(AtomicUsize::new(0));
    let c = calls.clone();
    let block = move |xs: &[Array]| {
        c.fetch_add(1, Ordering::SeqCst);
        Ok(vec![tanh(&matmul(&xs[0], &xs[1])?)?])
    };
    let checkpointed = checkpoint(block.clone());
    let loss = move |xs: &[Array]| sum_all(&multiply(&checkpointed(xs)?[0], 2.0)?);
    let plain = move |xs: &[Array]| sum_all(&multiply(&block(xs)?[0], 2.0)?);
    let a = arange(0.0, 6.0, 1.0, DType::F32)
        .unwrap()
        .reshape(&[2, 3])
        .unwrap();
    let b = arange(-1.0, 0.5, 0.25, DType::F32)
        .unwrap()
        .reshape(&[3, 2])
        .unwrap();
    let args = [&a * 0.1, b];
    calls.store(0, Ordering::SeqCst);
    let (value, grads) = value_and_grad(loss, &[0, 1])(&args).unwrap();
    assert_eq!(
        calls.load(Ordering::SeqCst),
        2,
        "forward plus recomputation"
    );
    let (expected_value, expected) = value_and_grad(plain, &[0, 1])(&args).unwrap();
    assert_close(&value, &expected_value);
    for (g, e) in grads.iter().zip(&expected) {
        assert_close(g, e);
    }
}