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
159
160
161

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

/// Select values along a single axis `I` resulting in `T`. Equivalent
/// to `torch.select` and `torch.gather` from pytorch.
pub trait Select1<T, const I: isize> {
    type Indices: Clone;

    /// Select sub elements using [Self::Indices].
    /// The same element can be selected multiple times depending
    /// on [Self::Indices].
    fn select(self, indices: &Self::Indices) -> T;
}

macro_rules! impl_select {
    ($Axis:expr, $SrcTy:ty, $IndTy:tt, $DstTy:ty, {$($Dims:tt),*}) => {
impl<$(const $Dims: usize, )* H: Tape> Select1<$DstTy, $Axis> for $SrcTy {
    type Indices = $IndTy;
    fn select(self, indices: &Self::Indices) -> $DstTy {
        let mut result: <$DstTy as Tensor>::NoTape = TensorCreator::zeros();
        Cpu::select_axis(self.data(), indices, result.mut_data());

        #[allow(clippy::clone_on_copy)]
        let i = indices.clone();

        move_tape_and_add_backward_op(self, result, move |mut t, result, grads| {
            let (t_grad, result_grad) = grads.mut_and_ref(&t, &result);
            Cpu::fill(t.mut_data(), &mut |v| *v = 0.0);
            Cpu::select_add(t.mut_data(), &i, result_grad);
            Cpu::add(t_grad, t.data());
        })
    }
}
    };
}

// 1d
impl_select!(-1, Tensor1D<M, H>, usize, Tensor0D<H>, {M});
impl_select!(-1, Tensor1D<M, H>, [usize; Z], Tensor1D<Z, H>, {M, Z});

// 2d
impl_select!(0, Tensor2D<M, N, H>, usize, Tensor1D<N, H>, {M, N});
impl_select!(0, Tensor2D<M, N, H>, [usize; Z], Tensor2D<Z, N, H>, {M, N, Z});
impl_select!(-1, Tensor2D<M, N, H>, [usize; M], Tensor1D<M, H>, {M, N});
impl_select!(-1, Tensor2D<M, N, H>, [[usize; Z]; M], Tensor2D<M, Z, H>, {M, N, Z});

// 3d
impl_select!(0, Tensor3D<M, N, O, H>, usize, Tensor2D<N, O, H>, {M, N, O});
impl_select!(0, Tensor3D<M, N, O, H>, [usize; Z], Tensor3D<Z, N, O, H>, {M, N, O, Z});
impl_select!(1, Tensor3D<M, N, O, H>, [usize; M], Tensor2D<M, O, H>, {M, N, O});
impl_select!(1, Tensor3D<M, N, O, H>, [[usize; Z]; M], Tensor3D<M, Z, O, H>, {M, N, O, Z});
impl_select!(-1, Tensor3D<M, N, O, H>, [[usize; N]; M], Tensor2D<M, N, H>, {M, N, O});
impl_select!(-1, Tensor3D<M, N, O, H>, [[[usize; Z]; N]; M], Tensor3D<M, N, Z, H>, {M, N, O, Z});

// 4d
impl_select!(0, Tensor4D<M, N, O, P, H>, usize, Tensor3D<N, O, P, H>, {M, N, O, P});
impl_select!(0, Tensor4D<M, N, O, P, H>, [usize; Z], Tensor4D<Z, N, O, P, H>, {M, N, O, P, Z});
impl_select!(1, Tensor4D<M, N, O, P, H>, [usize; M], Tensor3D<M, O, P, H>, {M, N, O, P});
impl_select!(1, Tensor4D<M, N, O, P, H>, [[usize; Z]; M], Tensor4D<M, Z, O, P, H>, {M, N, O, P, Z});
impl_select!(2, Tensor4D<M, N, O, P, H>, [[usize; N]; M], Tensor3D<M, N, P, H>, {M, N, O, P});
impl_select!(2, Tensor4D<M, N, O, P, H>, [[[usize; Z]; N]; M], Tensor4D<M, N, Z, P, H>, {M, N, O, P, Z});
impl_select!(-1, Tensor4D<M, N, O, P, H>, [[[usize; O]; N]; M], Tensor3D<M, N, O, H>, {M, N, O, P});
impl_select!(-1, Tensor4D<M, N, O, P, H>, [[[[usize; Z]; O]; N]; M], Tensor4D<M, N, O, Z, H>, {M, N, O, P, Z});

#[cfg(test)]
mod tests {
    use rand::thread_rng;

    use super::*;

    #[test]
    fn test_valid_selects_1d() {
        let _: Tensor0D = Tensor1D::<5>::zeros().select(&0);
        let _: Tensor1D<3> = Tensor1D::<5>::zeros().select(&[1, 2, 3]);
        let _: Tensor1D<10> = Tensor1D::<5>::zeros().select(&[0, 1, 2, 3, 4, 0, 1, 2, 3, 4]);
    }

    #[test]
    fn test_select_1d_backward() {
        let mut rng = thread_rng();
        let t: Tensor1D<5> = TensorCreator::randn(&mut rng);
        let r: Tensor0D<OwnedTape> = t.trace().select(&0);
        assert_eq!(r.data(), &t.data()[0]);
        let g = r.exp().mean().backward();
        assert_eq!(g.ref_gradient(&t), &[t.data()[0].exp(), 0.0, 0.0, 0.0, 0.0]);
    }

    #[test]
    fn test_select_1d_less_backward() {
        let mut rng = thread_rng();
        let t: Tensor1D<5> = TensorCreator::randn(&mut rng);
        let r: Tensor1D<2, OwnedTape> = t.trace().select(&[0, 3]);
        assert_eq!(r.data(), &[t.data()[0], t.data()[3]]);
        let g = r.mean().backward();
        assert_eq!(g.ref_gradient(&t), &[0.5, 0.0, 0.0, 0.5, 0.0]);
    }

    #[test]
    fn test_select_1d_more_backward() {
        let mut rng = thread_rng();
        let t: Tensor1D<5> = TensorCreator::randn(&mut rng);
        let _t = *t.data();
        let r: Tensor1D<8, OwnedTape> = t.trace().select(&[0, 1, 2, 3, 4, 2, 4, 4]);
        assert_eq!(
            r.data(),
            &[_t[0], _t[1], _t[2], _t[3], _t[4], _t[2], _t[4], _t[4]]
        );
        let g = r.mean().backward();
        assert_eq!(
            g.ref_gradient(&t),
            &[1.0 / 8.0, 1.0 / 8.0, 2.0 / 8.0, 1.0 / 8.0, 3.0 / 8.0]
        );
    }

    #[test]
    fn test_select_last_1d() {
        let t: Tensor1D<3> = Tensor1D::new([1.0, 2.0, 3.0]);
        let r: Tensor0D<OwnedTape> = t.trace().select(&2);
        assert_eq!(r.data(), &3.0);
        // NOTE: .exp() so we make sure its using result grad properly
        let gradients = r.exp().backward();
        assert_eq!(gradients.ref_gradient(&t), &[0.0, 0.0, 20.085537]);
    }

    #[test]
    fn test_select_last_2d() {
        let t: Tensor2D<2, 3> = Tensor2D::new([[1.0, 2.0, 3.0], [-1.0, -2.0, -3.0]]);
        let r: Tensor1D<2, OwnedTape> = t.trace().select(&[1, 2]);
        assert_eq!(r.data(), &[2.0, -3.0]);
        let gradients = r.mean().backward();
        assert_eq!(
            gradients.ref_gradient(&t),
            &[[0.0, 0.5, 0.0], [0.0, 0.0, 0.5]]
        );
    }

    #[test]
    fn test_select_last_3d() {
        let t: Tensor3D<4, 2, 3> = Tensor3D::new([
            [[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]],
            [[-1.0, -2.0, -3.0], [-4.0, -5.0, -6.0]],
            [[-3.0, 2.0, -1.0], [-6.0, 5.0, -4.0]],
            [[1.0, -2.0, 3.0], [4.0, -5.0, 6.0]],
        ]);
        let r: Tensor2D<4, 2, OwnedTape> = t.trace().select(&[[0, 1], [2, 2], [1, 1], [0, 0]]);
        assert_eq!(
            r.data(),
            &[[1.0, 5.0], [-3.0, -6.0], [2.0, 5.0], [1.0, 4.0]]
        );
        let gradients = r.mean().backward();
        assert_eq!(
            gradients.ref_gradient(&t),
            &[
                [[0.125, 0.0, 0.0], [0.0, 0.125, 0.0]],
                [[0.0, 0.0, 0.125], [0.0, 0.0, 0.125]],
                [[0.0, 0.125, 0.0], [0.0, 0.125, 0.0]],
                [[0.125, 0.0, 0.0], [0.125, 0.0, 0.0]]
            ]
        );
    }
}