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

use crate::gradients::*;
use crate::prelude::*;
use rand::Rng;
use rand_distr::Uniform;

/// **Requires Nightly** Performs 2d convolutions on 3d and 4d images.
///
/// **Pytorch Equivalent**: `torch.nn.Conv2d`
///
/// Generics:
/// - `IN_CHAN`: The number of input channels in an image.
/// - `OUT_CHAN`: The number of channels in the output of the layer.
/// - `KERNEL_SIZE`: The size of the kernel applied to both width and height of the images.
/// - `STRIDE`: How far to move the kernel each step. Defaults to `1`
/// - `PADDING`: How much zero padding to add around the images. Defaults to `0`.
///
/// Examples:
/// ```ignore
/// #![cfg_attr(feature = "nightly", feature(generic_const_exprs))]
/// # use dfdx::prelude::*;
/// let m: Conv2D<16, 33, 3> = Default::default();
/// #[cfg(feature = "nightly")]
/// let _: Tensor3D<33, 30, 62> = m.forward(Tensor3D::<16, 32, 64>::zeros());
/// #[cfg(feature = "nightly")]
/// let _: Tensor4D<2, 33, 13, 12> = m.forward(Tensor4D::<2, 16, 15, 14>::zeros());
/// ```
#[derive(Default, Debug, Clone)]
pub struct Conv2D<
    const IN_CHAN: usize,
    const OUT_CHAN: usize,
    const KERNEL_SIZE: usize,
    const STRIDE: usize = 1,
    const PADDING: usize = 0,
> {
    pub weight: Tensor4D<OUT_CHAN, IN_CHAN, KERNEL_SIZE, KERNEL_SIZE>,
    pub bias: Tensor1D<OUT_CHAN>,
}

impl<const I: usize, const O: usize, const K: usize, const S: usize, const P: usize>
    CanUpdateWithGradients for Conv2D<I, O, K, S, P>
{
    fn update<G: GradientProvider>(&mut self, grads: &mut G, unused: &mut UnusedTensors) {
        self.weight.update(grads, unused);
        self.bias.update(grads, unused);
    }
}

impl<const I: usize, const O: usize, const K: usize, const S: usize, const P: usize> ResetParams
    for Conv2D<I, O, K, S, P>
{
    fn reset_params<R: Rng>(&mut self, rng: &mut R) {
        let k = (I * K * K) as f32;
        let bound = 1.0 / k.sqrt();
        let dist = Uniform::new(-bound, bound);
        self.weight.randomize(rng, &dist);
        self.bias.randomize(rng, &dist);
    }
}

#[cfg(feature = "nightly")]
impl<
        T: Tape,
        const I: usize,
        const O: usize,
        const K: usize,
        const S: usize,
        const P: usize,
        const H: usize,
        const W: usize,
    > Module<Tensor3D<I, H, W, T>> for Conv2D<I, O, K, S, P>
where
    [[[(); (W + 2 * P - K) / S + 1]; (H + 2 * P - K) / S + 1]; O]:,
{
    type Output = Tensor3D<O, { (H + 2 * P - K) / S + 1 }, { (W + 2 * P - K) / S + 1 }, T>;

    fn forward(&self, x: Tensor3D<I, H, W, T>) -> Self::Output {
        x.conv2d::<O, K, S, P>(&self.weight, &self.bias)
    }
}

#[cfg(feature = "nightly")]
impl<
        T: Tape,
        const B: usize,
        const I: usize,
        const O: usize,
        const K: usize,
        const S: usize,
        const P: usize,
        const H: usize,
        const W: usize,
    > Module<Tensor4D<B, I, H, W, T>> for Conv2D<I, O, K, S, P>
where
    [[[[(); (W + 2 * P - K) / S + 1]; (H + 2 * P - K) / S + 1]; O]; B]:,
{
    type Output = Tensor4D<B, O, { (H + 2 * P - K) / S + 1 }, { (W + 2 * P - K) / S + 1 }, T>;

    fn forward(&self, x: Tensor4D<B, I, H, W, T>) -> Self::Output {
        x.conv2d::<O, K, S, P>(&self.weight, &self.bias)
    }
}

impl<const I: usize, const O: usize, const K: usize, const S: usize, const P: usize, T> ModuleMut<T>
    for Conv2D<I, O, K, S, P>
where
    Self: Module<T>,
{
    type Output = <Self as Module<T>>::Output;
    fn forward_mut(&mut self, input: T) -> Self::Output {
        self.forward(input)
    }
}

#[cfg(feature = "nightly")]
#[cfg(test)]
mod tests {
    use super::*;
    use rand::thread_rng;

    #[test]
    fn test_forward_3d_sizes() {
        type Img = Tensor3D<3, 10, 10>;
        let _: Tensor3D<2, 8, 8> = Conv2D::<3, 2, 3>::default().forward(Img::zeros());
        let _: Tensor3D<4, 8, 8> = Conv2D::<3, 4, 3>::default().forward(Img::zeros());
        let _: Tensor3D<4, 9, 9> = Conv2D::<3, 4, 2>::default().forward(Img::zeros());
        let _: Tensor3D<4, 7, 7> = Conv2D::<3, 4, 4>::default().forward(Img::zeros());
        let _: Tensor3D<2, 4, 4> = Conv2D::<3, 2, 3, 2>::default().forward(Img::zeros());
        let _: Tensor3D<2, 3, 3> = Conv2D::<3, 2, 3, 3>::default().forward(Img::zeros());
        let _: Tensor3D<2, 10, 10> = Conv2D::<3, 2, 3, 1, 1>::default().forward(Img::zeros());
        let _: Tensor3D<2, 12, 12> = Conv2D::<3, 2, 3, 1, 2>::default().forward(Img::zeros());
        let _: Tensor3D<2, 6, 6> = Conv2D::<3, 2, 3, 2, 2>::default().forward(Img::zeros());
    }

    #[test]
    fn test_forward_4d_sizes() {
        type Img = Tensor4D<5, 3, 10, 10>;
        let _: Tensor4D<5, 2, 8, 8> = Conv2D::<3, 2, 3>::default().forward(Img::zeros());
        let _: Tensor4D<5, 4, 8, 8> = Conv2D::<3, 4, 3>::default().forward(Img::zeros());
        let _: Tensor4D<5, 4, 9, 9> = Conv2D::<3, 4, 2>::default().forward(Img::zeros());
        let _: Tensor4D<5, 4, 7, 7> = Conv2D::<3, 4, 4>::default().forward(Img::zeros());
        let _: Tensor4D<5, 2, 4, 4> = Conv2D::<3, 2, 3, 2>::default().forward(Img::zeros());
        let _: Tensor4D<5, 2, 3, 3> = Conv2D::<3, 2, 3, 3>::default().forward(Img::zeros());
        let _: Tensor4D<5, 2, 10, 10> = Conv2D::<3, 2, 3, 1, 1>::default().forward(Img::zeros());
        let _: Tensor4D<5, 2, 12, 12> = Conv2D::<3, 2, 3, 1, 2>::default().forward(Img::zeros());
        let _: Tensor4D<5, 2, 6, 6> = Conv2D::<3, 2, 3, 2, 2>::default().forward(Img::zeros());
    }

    #[test]
    fn test_2_conv_sizes() {
        type A = Conv2D<1, 2, 3>;
        type B = Conv2D<2, 4, 3>;
        let _: Tensor3D<4, 6, 6> = <(A, B)>::default().forward(Tensor3D::<1, 10, 10>::zeros());
    }

    #[test]
    fn test_3_conv_sizes() {
        type A = Conv2D<1, 2, 3>;
        type B = Conv2D<2, 4, 3>;
        type C = Conv2D<4, 1, 1, 1, 1>;

        type Img = Tensor3D<1, 10, 10>;
        let _: Tensor3D<1, 8, 8> = <(A, B, C)>::default().forward_mut(Img::zeros());
    }

    #[test]
    fn test_conv_with_optimizer() {
        let mut rng = thread_rng();

        let mut m: Conv2D<2, 4, 3> = Default::default();
        m.reset_params(&mut rng);

        let weight_init = m.weight.clone();
        let bias_init = m.bias.clone();

        let mut opt: Sgd<_> = Default::default();
        let out = m.forward(Tensor4D::<8, 2, 28, 28>::randn(&mut rng).trace());
        let gradients = backward(out.square().mean());

        assert_ne!(gradients.ref_gradient(&m.weight), &[[[[0.0; 3]; 3]; 2]; 4]);
        assert_ne!(gradients.ref_gradient(&m.bias), &[0.0; 4]);

        opt.update(&mut m, gradients).expect("unused params");

        assert_ne!(weight_init.data(), m.weight.data());
        assert_ne!(bias_init.data(), m.bias.data());
    }
}