autograd 0.8.0

Tensors and differentiable operations in Rust
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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181

extern crate ndarray;
extern crate rayon;

use self::rayon::iter::*;
use ndarray_ext;
use ndarray_ext::NdArray;
use op;
use tensor::Tensor;

// `Tensordot` is implemented in `ops/mod.rs`.

pub struct MatMul {
    pub transpose_a: bool,
    pub transpose_b: bool,
}

pub struct BatchMatMul {
    pub transpose_a: bool,
    pub transpose_b: bool,
}

impl op::Op for MatMul {
    fn name(&self) -> &str {
        "MatMul"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        let xs = ctx.grab_inputs();
        let x0 = xs[0];
        let x1 = xs[1];
        let x0_shape = x0.shape();
        let x1_shape = x1.shape();
        assert_eq!(
            x0_shape.len(),
            2,
            "First input to the matmul should be Matrix"
        );
        assert_eq!(
            x1_shape.len(),
            2,
            "Second input to the matmul should be Matrix"
        );
        let x0_view = x0.view();
        let x1_view = x1.view();
        // unwrap is always safe
        let mut a = x0_view.into_shape((x0_shape[0], x0_shape[1])).unwrap();
        let mut b = x1_view.into_shape((x1_shape[0], x1_shape[1])).unwrap();
        if self.transpose_a {
            // almost zero cost
            a.swap_axes(0, 1);
        }
        if self.transpose_b {
            // almost zero cost
            b.swap_axes(0, 1);
        }
        vec![Ok(a.dot(&b).into_dyn())]
    }

    fn grad(&self, gy: &Tensor, inputs: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>> {
        let opa = Tensor::builder()
            .set_inputs(vec![gy, inputs[1]])
            .build(MatMul {
                transpose_a: false,
                transpose_b: true,
            });

        let opb = Tensor::builder()
            .set_inputs(vec![inputs[0], gy])
            .build(MatMul {
                transpose_a: true,
                transpose_b: false,
            });

        vec![Some(opa), Some(opb)]
    }
}

impl op::Op for BatchMatMul {
    fn name(&self) -> &str {
        "BatchMatMul"
    }

    fn compute(&self, ctx: ::runtime::OpComputeContext) -> op::ComputeResult {
        let xs = ctx.grab_inputs();
        let x0: &NdArray = xs[0];
        let x1: &NdArray = xs[1];
        let shape0 = x0.shape();
        let shape1 = x1.shape();
        let rank0 = x0.ndim();
        let rank1 = x1.ndim();

        if rank0 != rank1 || shape0[..rank0 - 2] != shape1[..rank0 - 2] {
            panic!("Input shape mismatch: {:?} vs {:?}", shape0, shape1);
        }

        let row0 = shape0[rank0 - 2];
        let row1 = shape1[rank0 - 2];

        let col0 = shape0[rank0 - 1];
        let col1 = shape1[rank0 - 1];

        // squashes dims (remains last two dims)
        // unwrap is always safe
        let x0_flattened = {
            let mut a = x0.view()
                .into_shape((x0.len() / row0 / col0, row0, col0))
                .unwrap();
            if self.transpose_a {
                a.swap_axes(1, 2);
            }
            a
        };

        let x1_flattened = {
            let mut b = x1.view()
                .into_shape((x1.len() / row1 / col1, row1, col1))
                .unwrap();
            if self.transpose_b {
                b.swap_axes(1, 2);
            }
            b
        };

        // parallel dot
        let dot = (0..x0_flattened.shape()[0] as isize)
            .into_par_iter()
            .map(|i| {
                let x0_mat = x0_flattened
                    .slice(s![i..i + 1, .., ..])
                    .remove_axis(ndarray::Axis(0))
                    .to_owned();
                let x1_mat = x1_flattened
                    .slice(s![i..i + 1, .., ..])
                    .remove_axis(ndarray::Axis(0))
                    .to_owned();
                x0_mat.dot(&x1_mat).into_dyn()
            })
            .collect::<Vec<_>>();

        // owned to ref
        let mut dot_view = Vec::with_capacity(dot.len());
        for i in 0..dot.len() {
            dot_view.push(ndarray_ext::expand_dims_view(dot[i].view(), 0));
        }

        // stack dot result
        let stacked = ndarray::stack(ndarray::Axis(0), dot_view.as_slice()).unwrap();

        let dst_shape = {
            let stacked_shape = stacked.shape();
            shape0[..rank0 - 2]
                .into_iter()
                .chain(&[stacked_shape[1], stacked_shape[2]])
                .cloned()
                .collect::<Vec<usize>>()
        };

        // reshape to dst shape with safe unwrapping
        vec![Ok(stacked
            .into_shape(ndarray::IxDyn(dst_shape.as_slice()))
            .unwrap())]
    }

    fn grad(&self, gy: &Tensor, inputs: &[&Tensor], _: &Tensor) -> Vec<Option<Tensor>> {
        let opa = Tensor::builder()
            .set_inputs(vec![gy, inputs[1]])
            .build(BatchMatMul {
                transpose_a: false,
                transpose_b: true,
            });

        let opb = Tensor::builder()
            .set_inputs(vec![inputs[0], gy])
            .build(BatchMatMul {
                transpose_a: true,
                transpose_b: false,
            });

        vec![Some(opa), Some(opb)]
    }
}