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

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

use std::sync::Arc;

use num_traits::{Float, FromPrimitive};

fn make_4d<S: Shape>(strides: S::Concrete) -> [usize; 4] {
    match S::NUM_DIMS {
        3 => [0, strides[0], strides[1], strides[2]],
        4 => [strides[0], strides[1], strides[2], strides[3]],
        _ => panic!("Only implemented for 3d & 4d arrays"),
    }
}

impl super::Pool2DKind {
    fn init<E: Float>(&self) -> E {
        match self {
            super::Pool2DKind::Avg => E::zero(),
            super::Pool2DKind::Min => E::infinity(),
            super::Pool2DKind::Max => E::neg_infinity(),
        }
    }

    fn accum<E: Float>(&self, accum: &E, item: &E) -> E {
        match self {
            super::Pool2DKind::Avg => *accum + *item,
            super::Pool2DKind::Min => accum.min(*item),
            super::Pool2DKind::Max => accum.max(*item),
        }
    }

    fn normalize<E: Float + FromPrimitive>(&self, item: E, num_elements: usize) -> E {
        match self {
            super::Pool2DKind::Avg => item * E::from_f64(1.0 / num_elements as f64).unwrap(),
            super::Pool2DKind::Min => item,
            super::Pool2DKind::Max => item,
        }
    }

    fn filter<E: Float>(&self, item: E, needle: E, haystack: E) -> E {
        match self {
            super::Pool2DKind::Avg => item,
            super::Pool2DKind::Min => {
                if needle == haystack {
                    item
                } else {
                    E::zero()
                }
            }
            super::Pool2DKind::Max => {
                if needle == haystack {
                    item
                } else {
                    E::zero()
                }
            }
        }
    }
}

impl<E: Float + Dtype> super::Pool2DKernel<E> for Cpu {
    fn alloc<S: Shape>(&self, s: S) -> Result<Tensor<S, E, Self>, Self::Err> {
        self.try_zeros_like(&s)
    }
    fn forward<I: Shape, O: Shape>(
        &self,
        op: super::Pool2DOp,
        inp: &Tensor<I, E, Self>,
        out: &mut Tensor<O, E, Self>,
    ) -> Result<(), Self::Err> {
        let istr = make_4d::<I>(inp.strides);
        let ostr = make_4d::<O>(out.strides);

        let buf = inp.data.as_ref();
        let out_buf = Arc::make_mut(&mut out.data);
        for b in 0..op.batch {
            for c in 0..op.chan {
                for oh in 0..op.h_out {
                    for ow in 0..op.w_out {
                        let mut tmp = op.kind.init();
                        for k1 in 0..op.kernel {
                            let y = (oh * op.stride + op.dilation * k1).checked_sub(op.padding);
                            for k2 in 0..op.kernel {
                                let x = (ow * op.stride + op.dilation * k2).checked_sub(op.padding);
                                if let Some((y, x)) = y.zip(x) {
                                    if y < op.h_in && x < op.w_in {
                                        let inp_idx =
                                            b * istr[0] + c * istr[1] + y * istr[2] + x * istr[3];
                                        tmp = op.kind.accum(&tmp, &buf[inp_idx]);
                                    }
                                }
                            }
                        }
                        tmp = op.kind.normalize(tmp, op.kernel * op.kernel);
                        out_buf[b * ostr[0] + c * ostr[1] + oh * ostr[2] + ow * ostr[3]] = tmp;
                    }
                }
            }
        }
        Ok(())
    }
    fn backward<I: Shape, O: Shape>(
        &self,
        op: super::Pool2DOp,
        inp: &Tensor<I, E, Self>,
        grad_inp: &mut Self::Vec,
        out: &Tensor<O, E, Self>,
        grad_out: &Self::Vec,
    ) -> Result<(), Self::Err> {
        let istr = make_4d::<I>(inp.strides);
        let ostr = make_4d::<O>(out.strides);

        let inp_buf = inp.data.as_ref();
        let out_buf = out.data.as_ref();

        for b in 0..op.batch {
            for c in 0..op.chan {
                for oh in 0..op.h_out {
                    for ow in 0..op.w_out {
                        let out_idx = b * ostr[0] + c * ostr[1] + oh * ostr[2] + ow * ostr[3];
                        let go = grad_out[out_idx];
                        let go = op.kind.normalize(go, op.kernel * op.kernel);
                        let vo = out_buf[out_idx];

                        for k1 in 0..op.kernel {
                            let y = (oh * op.stride + op.dilation * k1).checked_sub(op.padding);
                            for k2 in 0..op.kernel {
                                let x = (ow * op.stride + op.dilation * k2).checked_sub(op.padding);
                                if let Some((y, x)) = y.zip(x) {
                                    if x < op.w_in && y < op.h_in {
                                        let inp_idx =
                                            b * istr[0] + c * istr[1] + y * istr[2] + x * istr[3];
                                        grad_inp[inp_idx] +=
                                            op.kind.filter(go, inp_buf[inp_idx], vo);
                                    }
                                }
                            }
                        }
                    }
                }
            }
        }
        Ok(())
    }
}