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

//! Convolution variants checked against their definitions.

use fastnn::prelude::*;
use fastnn::tensor::Window;

const TOLERANCE: f32 = 1e-4;

#[track_caller]
fn assert_close(got: f32, want: f32, what: &str) {
    assert!(
        (got - want).abs() < TOLERANCE,
        "{what}: got {got}, want {want}"
    );
}

/// With a shared weight and no bias, <conv(x), y> = <x, convT(y)>.
#[test]
fn conv_transpose_is_the_exact_adjoint_of_conv() {
    manual_seed(12);
    let window = Window::square(3, 2, 1);
    let conv = Conv2d::with_window(2, 3, window, false);
    let transpose = ConvTranspose2d::with_window(3, 2, window, false);

    // Both layers flatten the weight to the same [out, in*kh*kw] buffer.
    transpose.weight.set_value(conv.weight.value());

    // 7 satisfies (size + 2p - k) % s = 0, a size like 6 would need output_padding.
    let x = Tensor::randn(&[2, 2, 7, 7]);
    let forward = no_grad(|| conv.forward(&x)); // [2, 3, 4, 4]
    let y = Tensor::randn(&[2, 3, 4, 4]);
    let backward = no_grad(|| transpose.forward(&y)); // [2, 2, 7, 7]

    let lhs = forward.mul(&y).sum().item();
    let rhs = x.mul(&backward).sum().item();
    assert!(
        (lhs - rhs).abs() / lhs.abs().max(1.0) < TOLERANCE,
        "adjoint identity violated: ⟨conv x, y⟩ = {lhs} but ⟨x, convᵀ y⟩ = {rhs}"
    );
}

/// A 1x1 input through a 2x2 all-ones kernel at stride 2 fills a 2x2 output.
#[test]
fn conv_transpose_upsamples_a_pixel_by_hand() {
    let layer = ConvTranspose2d::with_window(1, 1, Window::square(2, 2, 0), false);
    layer
        .weight
        .set_value(Tensor::from_vec(vec![1.0; 4], &[1, 1, 2, 2]));

    let out = no_grad(|| layer.forward(&Tensor::from_vec(vec![5.0], &[1, 1, 1, 1])));
    assert_eq!(out.shape(), &[1, 1, 2, 2]);
    assert_eq!(out.to_vec(), vec![5.0; 4]);
}

/// With 1x1 kernels this is a per-channel scale and the channels must not mix.
#[test]
fn depthwise_keeps_channels_separate() {
    let layer = Conv2d::depthwise(2, 1, 1, 0);
    layer
        .weight
        .set_value(Tensor::from_vec(vec![2.0, 3.0], &[2, 1, 1, 1]));
    if let Some(bias) = &layer.bias {
        bias.set_value(Tensor::zeros(&[2]));
    }

    let input = Tensor::from_vec(vec![1.0, 10.0, 100.0, 1000.0], &[1, 2, 1, 2]);
    let out = no_grad(|| layer.forward(&input)).to_vec();
    assert_eq!(
        out,
        vec![2.0, 20.0, 300.0, 3000.0],
        "channel 0 ×2, channel 1 ×3, no mixing"
    );
}

/// Kernel `[a, b]` at dilation 2 gives `out[t] = a*x[t] + b*x[t+2]`.
#[test]
fn dilation_stretches_the_kernel_footprint() {
    let window = Window {
        kernel: (1, 2),
        stride: (1, 1),
        padding: (0, 0),
        dilation: (1, 2),
    };
    let layer = Conv2d::with_window(1, 1, window, false);
    layer
        .weight
        .set_value(Tensor::from_vec(vec![10.0, 1.0], &[1, 1, 1, 2]));

    let x = Tensor::from_vec(vec![1.0, 2.0, 3.0, 4.0, 5.0], &[1, 1, 1, 5]);
    let out = no_grad(|| layer.forward(&x)).to_vec();
    assert_eq!(out, vec![13.0, 24.0, 35.0]);
}

/// Poking a later input must not change earlier outputs.
#[test]
fn causal_conv1d_cannot_see_the_future() {
    manual_seed(13);
    let layer = Conv1d::causal(1, 1, 3, 2);
    let x = Tensor::randn(&[1, 1, 8]);

    let mut poked = x.to_vec();
    poked[5] += 10.0;
    let poked = Tensor::from_vec(poked, &[1, 1, 8]);

    let out_a = no_grad(|| layer.forward(&x));
    let out_b = no_grad(|| layer.forward(&poked));
    assert_eq!(out_a.shape(), &[1, 1, 8], "causal padding preserves length");

    let (a, b) = (out_a.to_vec(), out_b.to_vec());
    for t in 0..5 {
        assert_eq!(a[t], b[t], "output {t} changed when input 5 was poked");
    }
    assert!(a[5] != b[5], "output 5 should see input 5");
}

/// Each group must match its own small convolution built from the weight slices.
#[test]
fn grouped_conv_equals_independent_group_convs() {
    manual_seed(14);
    let layer = Conv2d::grouped(4, 6, 3, 1, 1, 2);
    let x = Tensor::randn(&[2, 4, 5, 5]);
    let out = no_grad(|| layer.forward(&x));

    let window = Window::square(3, 1, 1);
    for group in 0..2 {
        let slice = x.narrow(1, group * 2, 2);
        let kernels = layer
            .weight
            .value()
            .narrow(0, group * 3, 3)
            .reshape(&[3, -1]);
        let reference = no_grad(|| {
            kernels.matmul(&slice.im2col(window)).add(
                &layer
                    .bias
                    .as_ref()
                    .unwrap()
                    .value()
                    .narrow(0, group * 3, 3)
                    .reshape(&[1, 3, 1]),
            )
        });

        let got = out.narrow(1, group * 3, 3).to_vec();
        for (i, (g, r)) in got.iter().zip(reference.to_vec()).enumerate() {
            assert_close(*g, r, &format!("group {group} element {i}"));
        }
    }
}

#[test]
fn gradients_reach_all_conv_variant_parameters() {
    let layers: Vec<(&str, Box<dyn Module>, Tensor)> = vec![
        (
            "conv1d",
            Box::new(Conv1d::new(2, 3, 3, 1, 1)),
            Tensor::randn(&[1, 2, 6]),
        ),
        (
            "causal conv1d",
            Box::new(Conv1d::causal(2, 3, 3, 2)),
            Tensor::randn(&[1, 2, 6]),
        ),
        (
            "grouped",
            Box::new(Conv2d::grouped(4, 4, 3, 1, 1, 2)),
            Tensor::randn(&[1, 4, 5, 5]),
        ),
        (
            "transpose2d",
            Box::new(ConvTranspose2d::new(2, 3, 3, 2, 1)),
            Tensor::randn(&[1, 2, 4, 4]),
        ),
        (
            "transpose1d",
            Box::new(ConvTranspose1d::new(2, 3, 4, 2, 1)),
            Tensor::randn(&[1, 2, 5]),
        ),
    ];

    for (name, layer, input) in layers {
        layer.forward(&input).sum().backward();
        for (param_name, param) in layer.named_parameters() {
            assert!(
                param.grad().is_some(),
                "{name}: no gradient reached {param_name}"
            );
        }
    }
}

#[test]
fn output_padding_hits_the_requested_size() {
    let plain = ConvTranspose2d::new(1, 1, 3, 2, 1);
    let padded = ConvTranspose2d::new(1, 1, 3, 2, 1).output_padding(1);
    let x = Tensor::randn(&[1, 1, 4, 4]);

    assert_eq!(no_grad(|| plain.forward(&x)).shape(), &[1, 1, 7, 7]);
    assert_eq!(no_grad(|| padded.forward(&x)).shape(), &[1, 1, 8, 8]);
}