fastnn 0.3.3

A robust deep learning library in Rust with CUDA kernels
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

//! Tiny models of each layer family, trained until they fit.

use fastnn::prelude::*;

/// Steps until the loss is under `target` or `steps` run out, returns the last loss.
fn train_until(
    opt: &mut dyn Optimizer,
    steps: usize,
    target: f32,
    mut compute_loss: impl FnMut() -> Tensor,
) -> f32 {
    let mut value = f32::MAX;
    for _ in 0..steps {
        let loss = compute_loss();
        opt.zero_grad();
        loss.backward();
        opt.step();
        value = loss.item();
        if value < target {
            break;
        }
    }
    value
}

#[test]
fn mlp_converges_on_xor() {
    manual_seed(3);
    let inputs = Tensor::from_vec(vec![0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0], &[4, 2]);
    let targets = [0usize, 1, 1, 0];

    let model = Sequential::new()
        .add(Linear::new(2, 8))
        .add(Tanh)
        .add(Linear::new(8, 2));
    let mut opt = Adam::new(model.parameters(), 5e-2);

    let loss_value = train_until(&mut opt, 1000, 0.01, || {
        cross_entropy(&model.forward(&inputs), &targets)
    });
    assert!(
        loss_value < 0.01,
        "xor did not converge: final loss {loss_value}"
    );
}

#[test]
fn linear_regression_recovers_slope_and_intercept() {
    manual_seed(4);
    // y = 2x + 1 with no noise.
    let xs: Vec<f32> = (0..32).map(|i| i as f32 / 16.0 - 1.0).collect();
    let ys: Vec<f32> = xs.iter().map(|x| 2.0 * x + 1.0).collect();
    let inputs = Tensor::from_vec(xs, &[32, 1]);
    let targets = Tensor::from_vec(ys, &[32, 1]);

    let model = Linear::new(1, 1);
    let mut opt = SGD::new(model.parameters(), 0.5);

    for _ in 0..500 {
        let loss = mse(&model.forward(&inputs), &targets);
        opt.zero_grad();
        loss.backward();
        opt.step();
    }

    let params = model.parameters();
    let (weight, bias) = (params[0].value().to_vec()[0], params[1].value().to_vec()[0]);
    assert!((weight - 2.0).abs() < 0.01, "slope {weight}, wanted 2");
    assert!((bias - 1.0).abs() < 0.01, "intercept {bias}, wanted 1");
}

#[test]
fn cnn_overfits_a_small_batch() {
    manual_seed(5);
    let inputs = Tensor::randn(&[10, 1, 8, 8]);
    let targets: Vec<usize> = (0..10).collect();

    let model = Sequential::new()
        .add(Conv2d::new(1, 4, 3, 1, 1))
        .add(ReLU)
        .add(MaxPool2d::new(2))
        .add(Flatten::new())
        .add(Linear::new(4 * 4 * 4, 10));
    let mut opt = Adam::new(model.parameters(), 1e-2);

    let loss_value = train_until(&mut opt, 200, 0.05, || {
        cross_entropy(&model.forward(&inputs), &targets)
    });
    assert!(
        loss_value < 0.05,
        "cnn did not overfit ten samples: loss {loss_value}"
    );
}

#[test]
fn lstm_memorizes_a_short_sequence() {
    manual_seed(6);
    let inputs = Tensor::randn(&[1, 6, 4]);
    let targets = [2usize, 0, 3, 1, 2, 0];

    let lstm = LSTM::new(4, 24);
    let head = Linear::new(24, 4);
    let mut params = lstm.parameters();
    params.extend(head.parameters());
    let mut opt = Adam::new(params, 1e-2);

    let loss_value = train_until(&mut opt, 300, 0.05, || {
        let logits = head.forward(&lstm.forward(&inputs)).reshape(&[6, -1]);
        cross_entropy(&logits, &targets)
    });
    assert!(
        loss_value < 0.05,
        "lstm did not memorize: loss {loss_value}"
    );
}

#[test]
fn transformer_overfits_a_small_batch() {
    manual_seed(7);
    let source = Tensor::randn(&[2, 3, 16]);
    let target_in = Tensor::randn(&[2, 4, 16]);
    let wanted = Tensor::randn(&[2, 4, 16]);

    let model = Transformer::new(16, 2, 32, 1, 1, 0.0);
    let mut opt = Adam::new(model.parameters(), 3e-3);

    let loss_value = train_until(&mut opt, 400, 0.05, || {
        mse(&model.run(&source, &target_in), &wanted)
    });
    assert!(
        loss_value < 0.05,
        "transformer did not overfit: loss {loss_value}"
    );
}

#[test]
fn saved_model_predicts_identically_after_reload() {
    manual_seed(8);
    let inputs = Tensor::randn(&[4, 6]);
    let model = Sequential::new()
        .add(Linear::new(6, 12))
        .add(ReLU)
        .add(Linear::new(12, 3));

    // Train a little so the weights are not just the init.
    let targets = Tensor::zeros(&[4, 3]);
    let mut opt = SGD::new(model.parameters(), 0.01);
    for _ in 0..5 {
        let loss = mse(&model.forward(&inputs), &targets);
        opt.zero_grad();
        loss.backward();
        opt.step();
    }
    let before = no_grad(|| model.forward(&inputs)).to_vec();
    assert!(
        before.iter().all(|v| v.is_finite()),
        "training diverged before the roundtrip"
    );

    let path = std::env::temp_dir().join("fastnn_roundtrip_test.fdl");
    save(&model, &path).unwrap();
    let reloaded = Sequential::new()
        .add(Linear::new(6, 12))
        .add(ReLU)
        .add(Linear::new(12, 3));
    load(&reloaded, &path).unwrap();
    std::fs::remove_file(&path).ok();

    let after = no_grad(|| reloaded.forward(&inputs)).to_vec();
    assert_eq!(before, after, "reloaded model predicts differently");
}

#[test]
fn clip_grad_value_bounds_every_gradient_element() {
    let param = fastnn::nn::Param::new(Tensor::from_vec(vec![1.0, -2.0], &[2]));
    param.tensor().mul_scalar(100.0).sum().backward();

    clip_grad_value(std::slice::from_ref(&param), 1.5);
    let grad = param.grad().unwrap().to_vec();
    assert!(
        grad.iter().all(|g| g.abs() <= 1.5),
        "gradient escaped the clamp: {grad:?}"
    );
}

/// Learning rates are per optimizer, a shared one would test tuning.
#[test]
fn every_optimizer_minimizes_least_squares() {
    use fastnn::nn::Param;
    type Recipe = fn(Vec<Param>) -> Box<dyn Optimizer>;

    let recipes: Vec<(&str, Recipe)> = vec![
        ("sgd", |p| Box::new(SGD::new(p, 0.1))),
        ("adam", |p| Box::new(Adam::new(p, 0.05))),
        ("adamw", |p| Box::new(AdamW::new(p, 0.05))),
        ("rmsprop", |p| Box::new(RMSprop::new(p, 0.01))),
        ("rmsprop-centered", |p| {
            Box::new(RMSprop::new(p, 0.01).momentum(0.9).centered())
        }),
        ("adagrad", |p| Box::new(Adagrad::new(p, 0.5))),
        ("adadelta", |p| Box::new(Adadelta::new(p))),
        ("radam", |p| Box::new(RAdam::new(p, 0.05))),
        ("lion", |p| Box::new(Lion::new(p, 0.005))),
        ("lookahead-adam", |p| {
            Box::new(Lookahead::new(Adam::new(p, 0.05), 5, 0.5))
        }),
    ];

    for (name, make) in recipes {
        manual_seed(9);
        let inputs = Tensor::randn(&[64, 3]);
        let true_map = Tensor::from_vec(vec![1.0, -2.0, 0.5], &[3, 1]);
        let targets = inputs.matmul(&true_map);

        let model = Linear::new(3, 1);
        let mut opt = make(model.parameters());

        let loss_value = train_until(opt.as_mut(), 1500, 1e-3, || {
            mse(&model.forward(&inputs), &targets)
        });
        assert!(
            loss_value < 1e-3,
            "{name} failed to converge: loss {loss_value}"
        );
    }
}