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

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

/// Reduction along multiple axes using `mean`.
pub trait MeanTo: HasErr + HasShape {
    /// Mean reduction. **Pytorch equivalent**: `t.mean(Axes)`
    ///
    /// Example:
    /// ```rust
    /// # use dfdx::prelude::*;
    /// # let dev: Cpu = Default::default();
    /// let t = dev.tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]);
    /// let r = t.mean::<Rank0, _>(); // or `mean::<_, Axes2<0, 1>>()`
    /// assert_eq!(r.array(), 3.5);
    /// ```
    ///
    /// Reducing 1 axis:
    /// ```rust
    /// # use dfdx::prelude::*;
    /// # let dev: Cpu = Default::default();
    /// let t = dev.tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]);
    /// let r = t.mean::<Rank1<2>, _>(); // or `mean::<_, Axis<1>>()`
    /// assert_eq!(r.array(), [2.0, 5.0]);
    /// ```
    fn mean<Dst: Shape, Ax: Axes>(self) -> Self::WithShape<Dst>
    where
        Self::Shape: HasAxes<Ax> + ReduceShapeTo<Dst, Ax>,
    {
        self.try_mean().unwrap()
    }
    /// Fallible version of [MeanTo::mean]
    fn try_mean<Dst: Shape, Ax: Axes>(self) -> Result<Self::WithShape<Dst>, Self::Err>
    where
        Self::Shape: HasAxes<Ax> + ReduceShapeTo<Dst, Ax>;
}

impl<S: Shape, E: Dtype, D: Device<E>, T: Tape<E, D>> MeanTo for Tensor<S, E, D, T> {
    fn try_mean<Dst: Shape, Ax: Axes>(self) -> Result<Self::WithShape<Dst>, Self::Err>
    where
        Self::Shape: HasAxes<Ax> + ReduceShapeTo<Dst, Ax>,
    {
        let num_elements_reduced = <S as HasAxes<Ax>>::size(self.shape()) as f64;
        let inv_normalize = E::from_f64(1.0 / num_elements_reduced).unwrap();
        self.try_sum()?.try_mul(inv_normalize)
    }
}

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

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

    #[test]
    fn test_mean_1d() {
        let dev: TestDevice = Default::default();
        let t = dev.tensor([1.0, 2.0, 3.0]).to_dtype::<TestDtype>();
        let r = t.leaky_trace().mean();
        assert_close_to_literal!(r, 2.0);
        // NOTE: .exp() so we cover the case where .mean() has to use result grad.
        let g = r.exp().backward();
        assert_close_to_literal!(&g.get(&t), &[2.463019; 3]);
    }

    #[test]
    fn test_mean_2d() {
        let dev: TestDevice = Default::default();
        let t = dev
            .tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
            .to_dtype::<TestDtype>();
        let r = t.leaky_trace().mean();
        assert_close_to_literal!(r, 3.5);
        let g = r.backward();
        assert_close_to_literal!(g.get(&t), [[1.0 / 6.0; 3]; 2]);
    }

    #[test]
    fn test_mean_3d() {
        let dev: TestDevice = Default::default();
        let t: Tensor<Rank3<4, 2, 3>, TestDtype, _> = dev.ones();
        let r = t.leaky_trace().mean();
        assert_close_to_literal!(r, 1.0);
        let g = r.backward();
        assert_close_to_literal!(g.get(&t), [[[1.0 / 24.0; 3]; 2]; 4]);
    }

    #[test]
    fn test_mean_axis_0_2d() {
        let dev: TestDevice = Default::default();
        let t = dev
            .tensor([[1.0, 2.0, 3.0], [-2.0, 4.0, -6.0]])
            .to_dtype::<TestDtype>();
        let r = t.leaky_trace().mean::<Rank1<3>, _>();
        assert_close_to_literal!(r, [-0.5, 3.0, -1.5]);
        let g = r.exp().mean().backward();
        assert_close_to_literal!(g.get(&t), [[0.10108845, 3.3475895, 0.037188362]; 2]);
    }

    #[test]
    fn test_mean_axis_1_2d() {
        let dev: TestDevice = Default::default();
        let t = dev
            .tensor([[1.0, 2.0, 3.0], [-2.0, 4.0, -6.0]])
            .to_dtype::<TestDtype>();
        let r = t.leaky_trace().mean::<Rank1<2>, _>();
        assert_close_to_literal!(r, [2.0, -4.0 / 3.0]);
        let g = r.exp().mean().backward();
        assert_close_to_literal!(g.get(&t), [[1.2315094; 3], [0.043932855; 3]]);
    }

    #[test]
    fn test_mean_axes_3d_to_1d_02() {
        let dev: TestDevice = Default::default();
        let t: Tensor<Rank3<2, 3, 4>, TestDtype, _> = dev.sample_normal();
        let r = t.leaky_trace().mean::<Rank1<3>, _>();
        let r2 = t.leaky_trace().sum::<_, Axis<0>>().sum::<_, Axis<1>>() / 8.0;
        assert_close_to_tensor!(r, r2);
        let g = r.mean().backward();
        let g2 = r2.mean().backward();
        assert_close_to_literal!(g.get(&t), [[[1. / 24.; 4]; 3]; 2]);
        assert_close_to_tensor!(g.get(&t), g2.get(&t));
    }

    #[test]
    fn test_mean_axes_3d_to_1d_01() {
        let dev: TestDevice = Default::default();
        let t: Tensor<Rank3<2, 3, 4>, TestDtype, _> = dev.sample_normal();
        let r = t.leaky_trace().mean::<Rank1<4>, _>();
        let r2 = t.sum::<_, Axis<0>>().sum::<_, Axis<0>>() / 6.0;
        assert_close_to_tensor!(r, r2);
    }
}