nuvec-nn 0.3.0

Neural-network modules, layers and optimizers 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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243

//! `nuvec-nn` against values recorded from MLX 0.32.4 (`mlx.nn`, `mlx.optimizers`).

use nuvec::{Array, Result, random};
use nuvec_nn::losses::{Reduction, cross_entropy};
use nuvec_nn::optim::{Adam, Optimizer, Sgd};
use nuvec_nn::{Linear, Module, Parameters, QuantizedLinear, relu, value_and_grad};

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

fn close(got: &[f32], expected: &[f64], what: &str) {
    assert_eq!(got.len(), expected.len(), "{what}");
    for (g, e) in got.iter().zip(expected) {
        assert!(
            (f64::from(*g) - e).abs() <= 1e-6 * e.abs().max(1.0),
            "{what}: {g} vs MLX {e}"
        );
    }
}

#[test]
fn linear_initialisation_equals_mlx() {
    random::seed(0);
    let l = Linear::new(3, 2, true).unwrap();
    let w: Vec<f32> = [
        -0.4937573969364166,
        -0.5538788437843323,
        -0.4882282614707947,
        0.4372609853744507,
        -0.387317955493927,
        -0.4599017798900604,
    ]
    .map(|v: f64| v as f32)
    .to_vec();
    assert_eq!(values(&l.weight), w, "bit for bit");
    let b: Vec<f32> = [0.004011034965515137f64, -0.022510409355163574]
        .map(|v| v as f32)
        .to_vec();
    assert_eq!(values(l.bias.as_ref().unwrap()), b);
    assert_eq!(l.weight.shape(), &[2, 3]);
}

#[test]
fn cross_entropy_equals_mlx() {
    let logits = Array::from_slice(&[2.0f32, -1.0, -1.0, 2.0], &[2, 2]).unwrap();
    let targets = Array::from_slice(&[0i32, 1], &[2]).unwrap();
    let none = cross_entropy(&logits, &targets, Reduction::None).unwrap();
    close(
        &values(&none),
        &[0.04858732968568802, 0.04858732968568802],
        "none",
    );
    let mean = cross_entropy(&logits, &targets, Reduction::Mean).unwrap();
    close(&values(&mean), &[0.04858732968568802], "mean");
    let logits = Array::from_slice(&[0.5f32, 1.5, -0.3, 2.0, 0.1, 0.4], &[2, 3]).unwrap();
    let targets = Array::from_slice(&[2u32, 0], &[2]).unwrap();
    let loss = cross_entropy(&logits, &targets, Reduction::None).unwrap();
    close(
        &values(&loss),
        &[2.2273428440093994, 0.30118927359580994],
        "three classes",
    );
    let floats = Array::from_slice(&[0.0f32, 1.0], &[2]).unwrap();
    assert!(cross_entropy(&logits, &floats, Reduction::None).is_err());
    let wrong = Array::from_slice(&[0u32, 1, 2], &[3]).unwrap();
    assert!(cross_entropy(&logits, &wrong, Reduction::None).is_err());
}

/// Two steps from `w = [1, 2, 3]` with the constant gradient `[0.1, -0.2, 0.3]`.
fn steps(opt: &mut impl Optimizer) -> Vec<Vec<f32>> {
    let mut w = vec![Array::from_slice(&[1.0f32, 2.0, 3.0], &[3]).unwrap()];
    let g = vec![Array::from_slice(&[0.1f32, -0.2, 0.3], &[3]).unwrap()];
    (0..2)
        .map(|_| {
            w = opt.apply(&w, &g).unwrap();
            values(&w[0])
        })
        .collect()
}

/// An optimizer's name, its two steps, and MLX's values for them.
type Case = (&'static str, Vec<Vec<f32>>, [[f64; 3]; 2]);

#[test]
fn optimizers_equal_mlx() {
    let cases: [Case; 5] = [
        (
            "sgd",
            steps(&mut Sgd::new(0.1)),
            [[0.99, 2.02, 2.97], [0.98, 2.04, 2.94]],
        ),
        (
            "sgd momentum + weight decay",
            steps(&mut Sgd::new(0.1).momentum(0.9).weight_decay(0.01)),
            [
                [0.9890000224113464, 2.0179998874664307, 2.9670000076293945],
                [0.9681110382080078, 2.0521819591522217, 2.9043331146240234],
            ],
        ),
        (
            "sgd nesterov",
            steps(&mut Sgd::new(0.1).momentum(0.9).nesterov(true)),
            [
                [0.9810000061988831, 2.0380001068115234, 2.943000078201294],
                [0.9538999795913696, 2.0922000408172607, 2.8617000579833984],
            ],
        ),
        (
            "adam",
            steps(&mut Adam::new(0.01)),
            [
                [0.9683773517608643, 2.0316226482391357, 2.9683773517608643],
                [0.9258815050125122, 2.0741186141967773, 2.9258813858032227],
            ],
        ),
        (
            "adam bias correction",
            steps(&mut Adam::new(0.01).bias_correction(true)),
            [
                [0.9900000691413879, 2.009999990463257, 2.990000009536743],
                [0.9800001978874207, 2.0199999809265137, 2.9800000190734863],
            ],
        ),
    ];
    for (name, got, expected) in cases {
        for (step, (g, e)) in got.iter().zip(&expected).enumerate() {
            close(g, e, &format!("{name} step {}", step + 1));
        }
    }
}

#[derive(Clone)]
struct Mlp(Vec<Linear>);

impl Module for Mlp {
    fn visit(&mut self, p: &mut Parameters<'_>) {
        p.module("layers", &mut self.0);
    }
}

/// The trainable parameters, without names.
fn arrays(m: &impl Module) -> Vec<Array> {
    m.trainable_parameters()
        .into_iter()
        .map(|(_, a)| a)
        .collect()
}

impl Mlp {
    fn forward(&self, x: &Array) -> Result<Array> {
        let mut h = x.clone();
        for (i, l) in self.0.iter().enumerate() {
            h = l.forward(&h)?;
            if i + 1 < self.0.len() {
                h = relu(&h)?;
            }
        }
        Ok(h)
    }
}

#[test]
fn value_and_grad_matches_the_core_transform_and_training_learns() {
    random::seed(3);
    let mut model = Mlp(vec![
        Linear::new(4, 8, true).unwrap(),
        Linear::new(8, 3, true).unwrap(),
    ]);
    let x = random::normal(&[16, 4], nuvec::DType::F32, None, None, None).unwrap();
    let labels: Vec<u32> = (0..16).map(|i| i % 3).collect();
    let y = Array::from_slice(&labels, &[16]).unwrap();
    // Class-dependent shift so the problem is learnable.
    let shift = Array::from_slice(
        &labels.iter().map(|&l| l as f32).collect::<Vec<_>>(),
        &[16, 1],
    )
    .unwrap();
    let x = nuvec::add(&x, &shift).unwrap();
    let loss = |m: &Mlp| cross_entropy(&m.forward(&x)?, &y, Reduction::Mean);

    let (value, grads) = value_and_grad(&model, loss).unwrap();
    assert_eq!(grads.len(), 4);
    let params = arrays(&model);
    let f = |ps: &[Array]| {
        let mut m = model.clone();
        m.update_trainable(ps)?;
        loss(&m)
    };
    let (core_value, core_grads) = nuvec::value_and_grad(f, &[0, 1, 2, 3])(&params).unwrap();
    assert_eq!(values(&value), values(&core_value));
    for (a, b) in grads.iter().zip(&core_grads) {
        assert_eq!(values(a), values(b));
    }

    let mut opt = Adam::new(0.05);
    let start = values(&value)[0];
    for _ in 0..60 {
        let (_, grads) = value_and_grad(&model, loss).unwrap();
        opt.update(&mut model, &grads).unwrap();
        nuvec::eval(&arrays(&model)).unwrap();
    }
    let end = values(&value_and_grad(&model, loss).unwrap().0)[0];
    assert!(end < 0.3 * start, "loss {start} -> {end}");
    assert!(model.update_trainable(&params[..2]).is_err());
}

#[test]
fn quantized_linear_equals_mlx() {
    random::seed(0);
    let q = QuantizedLinear::from_linear(&Linear::new(64, 3, true).unwrap(), 32, 4).unwrap();
    assert_eq!(q.weight.shape(), &[3, 8]);
    assert_eq!(q.scales.shape(), &[3, 2]);
    let x = nuvec::arange(0.0, 128.0, 1.0, nuvec::DType::F32).unwrap();
    let x = (x.reshape(&[2, 64]).unwrap() / 64.0) - 0.5;
    // mlx.nn.QuantizedLinear.from_linear(nn.Linear(64, 3), group_size=32, bits=4) after
    // mx.random.seed(0), on the CPU.
    let expected = [
        0.3646765351295471,
        -0.15546883642673492,
        0.23967400193214417,
        -0.7248094081878662,
        -0.07878664135932922,
        -0.29476773738861084,
    ];
    close(
        &values(&q.forward(&x).unwrap()),
        &expected,
        "QuantizedLinear",
    );
    // Only the bias trains (as MLX's from_linear); gradients reach it through the layer.
    let names = |ps: Vec<(String, Array)>| ps.into_iter().map(|(n, _)| n).collect::<Vec<_>>();
    assert_eq!(
        names(q.parameters()),
        ["weight", "scales", "biases", "bias"]
    );
    assert_eq!(names(q.trainable_parameters()), ["bias"]);
    let (_, grads) = value_and_grad(&q, |m: &QuantizedLinear| {
        nuvec::sum(&m.forward(&x)?, &[], false)
    })
    .unwrap();
    assert_eq!(values(&grads[0]), vec![2.0; 3]);
}