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

#![allow(clippy::needless_range_loop)]

use crate::shapes::{Axes, Dtype, RemoveDimTo, ReplaceDimTo, Shape};
use crate::tensor::{
    cpu::{index_to_i, LendingIterator, NdIndex},
    Cpu, Storage, Tensor, ZerosTensor,
};

impl<E: Dtype> super::ReplaceDimKernel<E> for Cpu {
    fn forward<Src: Shape, Dst: Shape, Idx: Shape>(
        &self,
        inp: &Tensor<Src, E, Self>,
        idx: &Tensor<Idx, usize, Self>,
    ) -> Result<Tensor<Dst, E, Self>, Self::Err>
    where
        Src: ReplaceDimTo<Dst, Idx>,
    {
        let ax = Src::Ax::as_array()[0] as usize;
        assert!(<Idx as Shape>::NUM_DIMS >= ax);

        let offset = <Idx as Shape>::NUM_DIMS - ax;

        let mut out = self.try_zeros_like(&inp.shape.replace(idx.shape))?;
        let mut out_iter = out.iter_mut_with_index();
        while let Some((x, i_replaced)) = out_iter.next() {
            let mut i_idx: <Idx as Shape>::Concrete = Default::default();
            let mut i_inp: Src::Concrete = Default::default();

            // # Construct the `idx` indices
            // the indices into `idx` will always share the same "head"
            // as the indices for `out`.
            for j in 0..<Idx as Shape>::NUM_DIMS {
                i_idx[j] = i_replaced[j];
            }

            // # Construct the `inp` indices
            for j in 0..Src::NUM_DIMS {
                i_inp[j] = match j.cmp(&ax) {
                    std::cmp::Ordering::Less => i_replaced[j],
                    std::cmp::Ordering::Equal => idx[i_idx],
                    std::cmp::Ordering::Greater => i_replaced[j - 1 + offset],
                };
            }
            *x = inp[i_inp];
        }
        Ok(out)
    }

    fn backward<Src: Shape, Dst: Shape, Idx: Shape>(
        &self,
        inp: &Tensor<Src, E, Self>,
        grad_inp: &mut <Self as Storage<E>>::Vec,
        idx: &Tensor<Idx, usize, Self>,
        out: &Tensor<Dst, E, Self>,
        grad_out: &<Self as Storage<E>>::Vec,
    ) -> Result<(), Self::Err>
    where
        Src: ReplaceDimTo<Dst, Idx>,
    {
        let ax = Src::Ax::as_array()[0] as usize;

        // NOTE: this is the same exact indexing logic as forward
        let offset = <Idx as Shape>::NUM_DIMS - ax;

        let mut out_idx = NdIndex::new(out.shape, out.strides);
        while let Some((i_out, i_replaced)) = out_idx.next_with_idx() {
            let mut i_idx: <Idx as Shape>::Concrete = Default::default();
            let mut i_inp: Src::Concrete = Default::default();
            for j in 0..<Idx as Shape>::NUM_DIMS {
                i_idx[j] = i_replaced[j];
            }
            for j in 0..Src::NUM_DIMS {
                i_inp[j] = match j.cmp(&ax) {
                    std::cmp::Ordering::Less => i_replaced[j],
                    std::cmp::Ordering::Equal => idx[i_idx],
                    std::cmp::Ordering::Greater => i_replaced[j - 1 + offset],
                };
            }
            grad_inp[index_to_i(&inp.shape, &inp.strides, i_inp)] += grad_out[i_out];
        }
        Ok(())
    }
}

impl<E: Dtype> super::RemoveDimKernel<E> for Cpu {
    fn forward<Src: Shape, Dst: Shape, Idx: Shape>(
        &self,
        inp: &Tensor<Src, E, Self>,
        idx: &Tensor<Idx, usize, Self>,
    ) -> Result<Tensor<Dst, E, Self>, Self::Err>
    where
        Src: RemoveDimTo<Dst, Idx>,
    {
        let ax = Src::Ax::as_array()[0] as usize;

        let mut out = self.try_zeros_like(&inp.shape.remove(idx.shape))?;
        let mut out_iter = out.iter_mut_with_index();
        while let Some((x, i_replaced)) = out_iter.next() {
            let mut i_idx: <Idx as Shape>::Concrete = Default::default();
            let mut i_inp: Src::Concrete = Default::default();

            // # Construct the `idx` indices
            // the indices into `idx` will always share the same "head"
            // as the indices for `out`.
            for j in 0..<Idx as Shape>::NUM_DIMS {
                i_idx[j] = i_replaced[j];
            }

            // # Construct the `inp` indices
            for j in 0..Src::NUM_DIMS {
                i_inp[j] = match j.cmp(&ax) {
                    std::cmp::Ordering::Less => i_replaced[j],
                    std::cmp::Ordering::Equal => idx[i_idx],
                    std::cmp::Ordering::Greater => i_replaced[j - 1],
                };
            }
            *x = inp[i_inp];
        }
        Ok(out)
    }

    fn backward<Src: Shape, Dst: Shape, Idx: Shape>(
        &self,
        inp: &Tensor<Src, E, Self>,
        grad_inp: &mut <Self as Storage<E>>::Vec,
        idx: &Tensor<Idx, usize, Self>,
        out: &Tensor<Dst, E, Self>,
        grad_out: &<Self as Storage<E>>::Vec,
    ) -> Result<(), Self::Err>
    where
        Src: RemoveDimTo<Dst, Idx>,
    {
        let ax = Src::Ax::as_array()[0] as usize;

        let mut out_idx = NdIndex::new(out.shape, out.strides);
        while let Some((i_out, i_replaced)) = out_idx.next_with_idx() {
            let mut i_idx: <Idx as Shape>::Concrete = Default::default();
            let mut i_inp: Src::Concrete = Default::default();
            for j in 0..<Idx as Shape>::NUM_DIMS {
                i_idx[j] = i_replaced[j];
            }
            for j in 0..Src::NUM_DIMS {
                i_inp[j] = match j.cmp(&ax) {
                    std::cmp::Ordering::Less => i_replaced[j],
                    std::cmp::Ordering::Equal => idx[i_idx],
                    std::cmp::Ordering::Greater => i_replaced[j - 1],
                };
            }
            grad_inp[index_to_i(&inp.shape, &inp.strides, i_inp)] += grad_out[i_out];
        }
        Ok(())
    }
}