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

mod cpu_kernel;

#[cfg(feature = "cuda")]
mod cuda_kernel;

use crate::{
    shapes::{Dtype, HasShape, Shape},
    tensor::{HasErr, Merge, PutTape, SplitTape, Storage, Tape, Tensor},
};

pub trait ChooseKernel<E: Dtype>: Storage<E> + Storage<bool> {
    fn forward<S: Shape>(
        &self,
        cond: &Tensor<S, bool, Self>,
        lhs: &Tensor<S, E, Self>,
        rhs: &Tensor<S, E, Self>,
    ) -> Result<Tensor<S, E, Self>, Self::Err>;

    fn backward<S: Shape>(
        &self,
        cond: &Tensor<S, bool, Self>,
        lhs: &Tensor<S, E, Self>,
        grad_lhs: &mut <Self as Storage<E>>::Vec,
        rhs: &Tensor<S, E, Self>,
        grad_rhs: &mut <Self as Storage<E>>::Vec,
        grad_out: &<Self as Storage<E>>::Vec,
    ) -> Result<(), Self::Err>;
}

/// Choose values from two tensors using a boolean mask. Equivalent to `torch.where` from pytorch.
///
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev: Cpu = Default::default();
/// let cond: Tensor<Rank1<3>, bool, _> = dev.tensor([true, false, true]);
/// let a: Tensor<Rank1<3>, f32, _> = dev.tensor([1.0, 2.0, 3.0]);
/// let b: Tensor<Rank1<3>, f32, _> = dev.tensor([-1.0, -2.0, -3.0]);
/// let c = cond.choose(a, b);
/// assert_eq!(c.array(), [1.0, -2.0, 3.0]);
/// ```
pub trait ChooseFrom<Lhs, Rhs>: HasErr {
    type Output;

    /// Construct a new tensor, where the output tensor contains the elements of lhs where self is
    /// true, and rhs where self is false.
    fn choose(self, lhs: Lhs, rhs: Rhs) -> Self::Output {
        self.try_choose(lhs, rhs).unwrap()
    }

    /// Fallible version of choose
    fn try_choose(self, lhs: Lhs, rhs: Rhs) -> Result<Self::Output, Self::Err>;
}

impl<
        S: Shape,
        E: Dtype,
        D: ChooseKernel<E>,
        LhsTape: Tape<E, D> + Merge<RhsTape>,
        RhsTape: Tape<E, D>,
    > ChooseFrom<Tensor<S, E, D, LhsTape>, Tensor<S, E, D, RhsTape>> for Tensor<S, bool, D>
{
    type Output = Tensor<S, E, D, LhsTape>;

    fn try_choose(
        self,
        lhs: Tensor<S, E, D, LhsTape>,
        rhs: Tensor<S, E, D, RhsTape>,
    ) -> Result<Self::Output, Self::Err> {
        assert_eq!(self.shape(), lhs.shape());
        assert_eq!(lhs.shape(), rhs.shape());

        let (lhs, tape) = lhs.split_tape();
        let (rhs, rhs_tape) = rhs.split_tape();

        let out = lhs.device.forward(&self, &lhs, &rhs)?;

        let lhs_ghost = lhs.ghost();
        let rhs_ghost = rhs.ghost();
        let out_ghost = out.ghost();
        let mut tape = tape.merge(rhs_tape);
        tape.add_backward_op(move |grads| {
            grads.try_alloc_for(&lhs_ghost)?;
            grads.try_alloc_for(&rhs_ghost)?;
            grads.try_alloc_for(&out_ghost)?;
            let (grad_lhs, grad_rhs, grad_out) =
                grads.muts_and_ref(&lhs_ghost, &rhs_ghost, &out_ghost);
            lhs.device
                .backward(&self, &lhs, grad_lhs, &rhs, grad_rhs, grad_out)
        });

        Ok(out.put_tape(tape))
    }
}

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

    #[test]
    fn test_choose_1d_backward() {
        let dev: TestDevice = Default::default();
        let cond = dev.tensor([false, true, false, true, false]);
        let a: Tensor<Rank1<5>, f32, _> = dev.sample_normal();
        let b: Tensor<Rank1<5>, f32, _> = dev.sample_normal();
        let r = cond.choose(a.leaky_trace(), b.leaky_trace());

        let a_array = a.array();
        let b_array = b.array();
        assert_eq!(
            r.array(),
            [b_array[0], a_array[1], b_array[2], a_array[3], b_array[4]]
        );
        let g = r.exp().sum().backward();
        assert_eq!(
            g.get(&a).array(),
            [0.0, a_array[1].exp(), 0.0, a_array[3].exp(), 0.0]
        );
        assert_eq!(
            g.get(&b).array(),
            [
                b_array[0].exp(),
                0.0,
                b_array[2].exp(),
                0.0,
                b_array[4].exp()
            ]
        );
    }

    #[test]
    fn test_choose_2d_backward() {
        let dev: TestDevice = Default::default();
        let cond = dev.tensor([[false, true], [true, false]]);
        let a: Tensor<_, f32, _> = dev.sample_normal();
        let b: Tensor<_, f32, _> = dev.sample_normal();
        let r = cond.choose(a.leaky_trace(), b.leaky_trace());

        let a_array = a.array();
        let b_array = b.array();
        assert_eq!(
            r.array(),
            [
                [b_array[0][0], a_array[0][1]],
                [a_array[1][0], b_array[1][1]]
            ]
        );
        let g = r.exp().sum().backward();
        assert_eq!(
            g.get(&a).array(),
            [[0.0, a_array[0][1].exp()], [a_array[1][0].exp(), 0.0]]
        );
        assert_eq!(
            g.get(&b).array(),
            [[b_array[0][0].exp(), 0.0], [0.0, b_array[1][1].exp()]]
        );
    }
}