dfdx 0.9.0

Ergonomic auto differentiation in Rust, with pytorch like apis.
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

use super::utils::move_tape_and_add_backward_op;
use crate::prelude::*;

/// Broadcasts the `I`th dimension. Increases number dimensions by 1. Results in `T`. Opposite of [Reduce1].
pub trait Broadcast1<T, const I: isize> {
    /// Broadcast `self` into `T`, increasing number dimensions by 1.
    fn broadcast1(self) -> T;
}

/// Broadcasts dimensions `I1` and `I2`. Increases number dimensions by 2. Results in `T`.
pub trait Broadcast2<T, const I1: isize, const I2: isize> {
    /// Broadcast `self` into `T`, increasing number dimensions by 2.
    fn broadcast2(self) -> T;
}

/// Broadcasts dimensions `I1`, `I2`, and `I3`. Increases number dimensions by 3. Results in `T`.
pub trait Broadcast3<T, const I1: isize, const I2: isize, const I3: isize> {
    /// Broadcast `self` into `T`, increasing number dimensions by 3.
    fn broadcast3(self) -> T;
}

/// Broadcasts dimensions `I1`, `I2`, `I3`, and `I4`. Increases number dimensions by 4. Results in `T`.
pub trait Broadcast4<T, const I1: isize, const I2: isize, const I3: isize, const I4: isize> {
    /// Broadcast `self` into `T`, increasing number dimensions by 4.
    fn broadcast4(self) -> T;
}

macro_rules! impl_broadcast {
    (
        [$($Axes:expr),*],
        $SrcTy:ty, $DstTy:ty,
        $TensorTrait:tt, $fn_name:tt,
        $DeviceTrait:tt, {$($Dims:tt),*}
    ) => {
impl<$(const $Dims: usize, )* H: Tape> $TensorTrait<$DstTy, $($Axes, )*> for $SrcTy {
    fn $fn_name(self) -> $DstTy {
        let mut result = <$DstTy as Tensor>::NoTape::zeros();
        <Cpu as $DeviceTrait<_, _, $($Axes),*>>::broadcast_copy(result.mut_data(), self.data());
        move_tape_and_add_backward_op(self, result, move |t, result, grads| {
            let (t_grad, result_grad) = grads.mut_and_ref(&t, &result);
            <Cpu as $DeviceTrait<_, _, $($Axes),*>>::broadcast_add(t_grad, result_grad);
        })
    }
}
    };
}

impl<H: Tape> Broadcast1<Tensor0D<H>, -1> for Tensor0D<H> {
    fn broadcast1(self) -> Tensor0D<H> {
        self
    }
}

// 0d -> Nd
#[rustfmt::skip]
impl_broadcast!([-1], Tensor0D<H>, Tensor1D<M, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M});
impl_broadcast!([0, 1], Tensor0D<H>, Tensor2D<M, N, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N});
impl_broadcast!([0, 1, 2], Tensor0D<H>, Tensor3D<M, N, O, H>, Broadcast3, broadcast3, ForEachBroadcast3, {M, N, O});
impl_broadcast!([0, 1, 2, 3], Tensor0D<H>, Tensor4D<M, N, O, P, H>, Broadcast4, broadcast4, ForEachBroadcast4, {M, N, O, P});

// 1d -> Nd
impl_broadcast!([-1], Tensor1D<M, H>, Tensor2D<M, N, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N});
impl_broadcast!([0], Tensor1D<N, H>, Tensor2D<M, N, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N});
impl_broadcast!([1, 2], Tensor1D<M, H>, Tensor3D<M, N, O, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O});
impl_broadcast!([0, 2], Tensor1D<N, H>, Tensor3D<M, N, O, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O});
impl_broadcast!([0, 1], Tensor1D<O, H>, Tensor3D<M, N, O, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O});
impl_broadcast!([1, 2, 3], Tensor1D<M, H>, Tensor4D<M, N, O, P, H>, Broadcast3, broadcast3, ForEachBroadcast3, {M, N, O, P});
impl_broadcast!([0, 2, 3], Tensor1D<N, H>, Tensor4D<M, N, O, P, H>, Broadcast3, broadcast3, ForEachBroadcast3, {M, N, O, P});
impl_broadcast!([0, 1, 3], Tensor1D<O, H>, Tensor4D<M, N, O, P, H>, Broadcast3, broadcast3, ForEachBroadcast3, {M, N, O, P});
impl_broadcast!([0, 1, 2], Tensor1D<P, H>, Tensor4D<M, N, O, P, H>, Broadcast3, broadcast3, ForEachBroadcast3, {M, N, O, P});

// 2d -> Nd
impl_broadcast!([-1], Tensor2D<M, N, H>, Tensor3D<M, N, O, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N, O});
impl_broadcast!([1], Tensor2D<M, O, H>, Tensor3D<M, N, O, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N, O});
impl_broadcast!([0], Tensor2D<N, O, H>, Tensor3D<M, N, O, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N, O});
impl_broadcast!([2, 3], Tensor2D<M, N, H>, Tensor4D<M, N, O, P, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O, P});
impl_broadcast!([1, 3], Tensor2D<M, O, H>, Tensor4D<M, N, O, P, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O, P});
impl_broadcast!([1, 2], Tensor2D<M, P, H>, Tensor4D<M, N, O, P, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O, P});
impl_broadcast!([0, 3], Tensor2D<N, O, H>, Tensor4D<M, N, O, P, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O, P});
impl_broadcast!([0, 2], Tensor2D<N, P, H>, Tensor4D<M, N, O, P, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O, P});
impl_broadcast!([0, 1], Tensor2D<O, P, H>, Tensor4D<M, N, O, P, H>, Broadcast2, broadcast2, ForEachBroadcast2, {M, N, O, P});

// 3d -> 4d
impl_broadcast!([-1], Tensor3D<M, N, O, H>, Tensor4D<M, N, O, P, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N, O, P});
impl_broadcast!([2], Tensor3D<M, N, P, H>, Tensor4D<M, N, O, P, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N, O, P});
impl_broadcast!([1], Tensor3D<M, O, P, H>, Tensor4D<M, N, O, P, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N, O, P});
impl_broadcast!([0], Tensor3D<N, O, P, H>, Tensor4D<M, N, O, P, H>, Broadcast1, broadcast1, ForEachBroadcast1, {M, N, O, P});

#[cfg(test)]
mod tests {
    use super::*;
    use crate::tests::assert_close;
    use rand::thread_rng;

    #[test]
    fn test_valid_1d_broadcasts() {
        let _: Tensor1D<5> = Tensor0D::zeros().broadcast1();

        let _: Tensor2D<5, 3> = Tensor1D::<3>::zeros().broadcast1();
        let _: Tensor2D<5, 3> = Tensor1D::<5>::zeros().broadcast1();
        let _: Tensor2D<5, 3> = Tensor1D::<5>::zeros().broadcast1();

        let _: Tensor3D<3, 5, 7> = Tensor2D::<5, 7>::zeros().broadcast1();
        let _: Tensor3D<3, 5, 7> = Tensor2D::<3, 7>::zeros().broadcast1();
        let _: Tensor3D<3, 5, 7> = Tensor2D::<3, 5>::zeros().broadcast1();
        let _: Tensor3D<3, 5, 7> = Tensor2D::<3, 5>::zeros().broadcast1();

        let _: Tensor4D<3, 5, 7, 9> = Tensor3D::<5, 7, 9>::zeros().broadcast1();
        let _: Tensor4D<3, 5, 7, 9> = Tensor3D::<3, 7, 9>::zeros().broadcast1();
        let _: Tensor4D<3, 5, 7, 9> = Tensor3D::<3, 5, 9>::zeros().broadcast1();
        let _: Tensor4D<3, 5, 7, 9> = Tensor3D::<3, 5, 7>::zeros().broadcast1();
        let _: Tensor4D<3, 5, 7, 9> = Tensor3D::<3, 5, 7>::zeros().broadcast1();
    }

    #[test]
    fn test_valid_2d_broadcasts() {
        let _: Tensor2D<5, 3> = Tensor0D::zeros().broadcast2();

        let _: Tensor3D<3, 5, 7> = Tensor1D::<3>::zeros().broadcast2();
        let _: Tensor3D<3, 5, 7> = Tensor1D::<5>::zeros().broadcast2();
        let _: Tensor3D<3, 5, 7> = Tensor1D::<7>::zeros().broadcast2();

        let _: Tensor4D<3, 5, 7, 9> = Tensor2D::<3, 5>::zeros().broadcast2();
        let _: Tensor4D<3, 5, 7, 9> = Tensor2D::<3, 7>::zeros().broadcast2();
        let _: Tensor4D<3, 5, 7, 9> = Tensor2D::<3, 9>::zeros().broadcast2();
        let _: Tensor4D<3, 5, 7, 9> = Tensor2D::<5, 7>::zeros().broadcast2();
        let _: Tensor4D<3, 5, 7, 9> = Tensor2D::<5, 9>::zeros().broadcast2();
        let _: Tensor4D<3, 5, 7, 9> = Tensor2D::<7, 9>::zeros().broadcast2();
    }

    #[test]
    fn test_valid_3d_broadcasts() {
        let _: Tensor3D<3, 5, 7> = Tensor0D::zeros().broadcast3();

        let _: Tensor4D<3, 5, 7, 9> = Tensor1D::<3>::zeros().broadcast3();
        let _: Tensor4D<3, 5, 7, 9> = Tensor1D::<5>::zeros().broadcast3();
        let _: Tensor4D<3, 5, 7, 9> = Tensor1D::<7>::zeros().broadcast3();
        let _: Tensor4D<3, 5, 7, 9> = Tensor1D::<9>::zeros().broadcast3();
    }

    #[test]
    fn test_broadcast_backwards() {
        let mut rng = thread_rng();
        let a: Tensor1D<3> = TensorCreator::randn(&mut rng);
        let b: Tensor2D<5, 3> = TensorCreator::randn(&mut rng);
        let a_up: Tensor2D<5, 3, OwnedTape> = a.trace().broadcast1();
        assert_eq!(a_up.data(), &[*a.data(); 5]);
        let r = mul(a_up, &b);
        let g = r.exp().mean().backward();
        // a's gradient: (b * (b * a).exp()).sum(0) / 15
        // b's gradient: (a * (b * a).exp()) / 15
        let a_up: Tensor2D<5, 3> = a.clone().broadcast1();
        let a_grad = mul(mul(b.clone(), &a_up).exp(), &b).sum_axis::<0>() / 15.0;
        let b_grad = mul(mul(b.clone(), &a_up).exp(), &a_up) / 15.0;
        assert_close(g.ref_gradient(&a), a_grad.data());
        assert_close(g.ref_gradient(&b), b_grad.data());
    }
}