dfdx 0.13.0

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
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174

use crate::{shapes::*, tensor::*};

/// Broadcast self into a new shape.
///
/// **pytorch equivalent** `torch.broadcast_to`.
///
/// Use shape generic or output type to dictate what shape you want:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let a: Tensor<Rank2<3, 7>, f32, _> = dev.zeros();
/// // broadcast axis 1
/// let _: Tensor<Rank3<3, 5, 7>, _, _> = a.clone().broadcast();
/// // broadcast axis 0 and axis 2
/// let _ = a.clone().broadcast::<Rank4<1, 3, 5, 7>, _>();
/// ```
///
/// Use axes generic to dis-ambiguate:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let a: Tensor<Rank1<1>, f32, _> = dev.zeros();
/// // It's ambiguous what axes to broadcast here - explicitly say axes 0 and 2
/// let _: Tensor<Rank3<1, 1, 1>, _, _> = a.clone().broadcast::<_, Axes2<0, 2>>();
/// ```
pub trait BroadcastTo: HasErr + HasShape {
    /// Broadcast into shape `Dst` along axes `Ax`.
    fn broadcast<Dst: ConstShape, Ax: Axes>(self) -> Self::WithShape<Dst>
    where
        Self::Shape: BroadcastShapeTo<Dst, Ax>,
    {
        self.try_broadcast_like::<Dst, _>(&Default::default())
            .unwrap()
    }
    /// Fallible version of [BroadcastTo::broadcast]
    fn try_broadcast<Dst: ConstShape, Ax: Axes>(self) -> Result<Self::WithShape<Dst>, Self::Err>
    where
        Self::Shape: BroadcastShapeTo<Dst, Ax>,
    {
        self.try_broadcast_like::<Dst, _>(&Default::default())
    }
    /// Same as [BroadcastTo::broadcast], but the target shape is given
    fn broadcast_like<Dst: HasShape, Ax: Axes>(self, dst: &Dst) -> Self::WithShape<Dst::Shape>
    where
        Self::Shape: BroadcastShapeTo<Dst::Shape, Ax>,
    {
        self.try_broadcast_like(dst).unwrap()
    }
    /// fallible version of [BroadcastTo::broadcast_like]
    fn try_broadcast_like<Dst: HasShape, Ax: Axes>(
        self,
        dst: &Dst,
    ) -> Result<Self::WithShape<Dst::Shape>, Self::Err>
    where
        Self::Shape: BroadcastShapeTo<Dst::Shape, Ax>;
}

impl<S: Shape, E, D: Storage<E>, T: Tape<E, D>> BroadcastTo for Tensor<S, E, D, T> {
    fn try_broadcast_like<Dst: HasShape, Ax: Axes>(
        self,
        dst: &Dst,
    ) -> Result<Self::WithShape<Dst::Shape>, Self::Err>
    where
        Self::Shape: BroadcastShapeTo<Dst::Shape, Ax>,
    {
        self.shape().check(dst.shape());

        Ok(Tensor {
            id: self.id,
            data: self.data,
            shape: *dst.shape(),
            strides: self.shape.broadcast_strides(self.strides),
            device: self.device,
            tape: self.tape,
        })
    }
}

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

    #[test]
    #[should_panic]
    fn test_broadcast_incorrect_dims() {
        let dev: TestDevice = Default::default();
        let a: Tensor<(usize,), TestDtype, _> = dev.zeros_like(&(5,));
        let _: Tensor<(Const<3>, usize), TestDtype, _> = a.broadcast_like(&(Const, 7));
    }

    #[test]
    fn test_broadcast_with_tensor() {
        let dev: TestDevice = Default::default();
        let a1: Tensor<_, TestDtype, _> = dev.zeros_like(&(5,));
        let b: Tensor<_, TestDtype, _> = dev.zeros_like(&(2, 5, 3));
        let a2 = a1.broadcast_like::<_, Axes2<0, 2>>(&b);
        assert_eq!(a2.shape(), &(2, 5, 3));
    }

    #[test]
    fn test_valid_1d_broadcasts() {
        let dev: TestDevice = Default::default();
        let _: Tensor<Rank1<5>, TestDtype, _> = dev.zeros::<Rank0>().broadcast();
        let _: Tensor<Rank2<5, 3>, TestDtype, _> = dev.zeros::<Rank1<3>>().broadcast();
        let _: Tensor<Rank2<5, 3>, TestDtype, _> = dev.zeros::<Rank1<5>>().broadcast();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank2<5, 7>>().broadcast();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank2<3, 7>>().broadcast();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank2<3, 5>>().broadcast();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank2<3, 5>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank3<5, 7, 9>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank3<3, 7, 9>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank3<3, 5, 9>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank3<3, 5, 7>>().broadcast();
    }

    #[test]
    fn test_valid_2d_broadcasts() {
        let dev: TestDevice = Default::default();
        let _: Tensor<Rank2<5, 3>, TestDtype, _> = dev.zeros::<Rank0>().broadcast();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank1<3>>().broadcast();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank1<5>>().broadcast();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank1<7>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank2<3, 5>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank2<3, 7>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank2<3, 9>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank2<5, 7>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank2<5, 9>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank2<7, 9>>().broadcast();
    }

    #[test]
    fn test_valid_3d_broadcasts() {
        let dev: TestDevice = Default::default();
        let _: Tensor<Rank3<3, 5, 7>, TestDtype, _> = dev.zeros::<Rank0>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank1<3>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank1<5>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank1<7>>().broadcast();
        let _: Tensor<Rank4<3, 5, 7, 9>, TestDtype, _> = dev.zeros::<Rank1<9>>().broadcast();
    }

    #[test]
    fn test_broadcast_backwards() {
        let dev: TestDevice = Default::default();
        let a: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
        let b: Tensor<Rank2<5, 3>, TestDtype, _> = dev.sample_normal();
        let a_up = a.leaky_trace().broadcast::<Rank2<5, 3>, _>();
        assert_close!(a_up.array(), [a.array(); 5], 1e-4);
        let r = a_up * b.clone();
        let g = r.exp().mean().backward();

        let a_up = a.clone().broadcast::<Rank2<5, 3>, _>();
        // a's gradient: (b * (b * a).exp()).sum(0) / 15
        let a_grad = (b.clone() * (b.clone() * a_up.clone()).exp()).sum::<Rank1<3>, _>() / 15.0;
        // b's gradient: (a * (b * a).exp()) / 15
        let b_grad = (a_up.clone() * (b.clone() * a_up).exp()) / 15.0;
        assert_close_to_tensor!(g.get(&a), a_grad, 1e-4);
        assert_close_to_tensor!(g.get(&b), b_grad, 1e-4);
    }

    #[test]
    fn test_broadcast_summed() {
        let dev: TestDevice = Default::default();
        let a: Tensor<Rank1<3>, TestDtype, _> = dev.sample_normal();
        let g = a
            .leaky_trace()
            .broadcast::<Rank2<4, 3>, _>()
            .exp()
            .mean()
            .backward();
        assert_close_to_tensor!(g.get(&a), a.exp() / 3.0);
    }
}