tract-core 0.2.0

Tiny, no-nonsense, self contained, TensorFlow and ONNX inference
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

use crate::ops::prelude::*;

#[derive(Debug, Clone)]
pub enum PaddingSpec {
    Explicit(TVec<usize>, TVec<usize>),
    Valid,
    SameUpper,
    SameLower,
}

impl Default for PaddingSpec {
    fn default() -> PaddingSpec {
        PaddingSpec::Valid
    }
}

#[derive(Debug, Clone)]
pub struct ComputedPaddedDim<D: DimLike> {
    pub pad_before: TVec<D>,
    pub pad_after: TVec<D>,
    pub output: TVec<D>,
}

impl PaddingSpec {
    pub fn valid_dim(&self, d: usize) -> bool {
        match self {
            PaddingSpec::Valid => true,
            PaddingSpec::Explicit(a, b) => a[d] == 0 && b[d] == 0,
            _ => false,
        }
    }

    pub fn rm_axis(&self, d: usize) -> PaddingSpec {
        match self {
            PaddingSpec::Explicit(a, b) => {
                let mut a = a.clone();
                let mut b = b.clone();
                a.remove(d);
                b.remove(d);
                PaddingSpec::Explicit(a, b)
            }
            _ => self.clone(),
        }
    }

    pub fn compute<D: DimLike, KD: Into<D> + Copy>(
        &self,
        input_spatial_shape: &[D],
        kernel_spatial_shape: &[KD],
        dilations: &[usize],
        strides: &[usize],
    ) -> ComputedPaddedDim<D> {
        assert_eq!(dilations.len(), strides.len());
        assert_eq!(dilations.len(), input_spatial_shape.len());
        assert_eq!(dilations.len(), kernel_spatial_shape.len());
        match self {
            PaddingSpec::Valid => Self::explicit(
                input_spatial_shape,
                kernel_spatial_shape,
                dilations,
                strides,
                &*vec![0; kernel_spatial_shape.len()],
                &*vec![0; kernel_spatial_shape.len()],
            ),
            PaddingSpec::Explicit(ref bef, ref aft) => Self::explicit(
                input_spatial_shape,
                kernel_spatial_shape,
                dilations,
                strides,
                bef,
                aft,
            ),
            PaddingSpec::SameUpper => self.same(
                input_spatial_shape,
                kernel_spatial_shape,
                dilations,
                strides,
                true,
            ),
            PaddingSpec::SameLower => self.same(
                input_spatial_shape,
                kernel_spatial_shape,
                dilations,
                strides,
                false,
            ),
        }
    }

    fn explicit<D: DimLike, KD: Into<D> + Copy>(
        data_spatial_shape: &[D],
        kernel_spatial_shape: &[KD],
        dilations: &[usize],
        strides: &[usize],
        bef: &[usize],
        aft: &[usize],
    ) -> ComputedPaddedDim<D> {
        let spatial_rank = data_spatial_shape.len();
        assert_eq!(spatial_rank, kernel_spatial_shape.len());
        assert_eq!(spatial_rank, dilations.len());
        assert_eq!(spatial_rank, strides.len());
        assert_eq!(spatial_rank, aft.len());
        assert_eq!(spatial_rank, bef.len());
        let output_spatial_shape = (0..spatial_rank)
            .map(|ax| {
                let kernel_field = (kernel_spatial_shape[ax].into() - 1) * dilations[ax] + 1;
                let dim = (data_spatial_shape[ax] + bef[ax] + aft[ax] - kernel_field + 1)
                    .div_ceil(strides[ax]);
                dim
            })
            .collect();
        ComputedPaddedDim {
            output: output_spatial_shape,
            pad_before: bef.iter().map(|&x| D::from(x)).collect(),
            pad_after: aft.iter().map(|&x| D::from(x)).collect(),
        }
    }

    fn same<D: DimLike, KD: Into<D> + Copy>(
        &self,
        data_spatial_shape: &[D],
        kernel_spatial_shape: &[KD],
        dilations: &[usize],
        strides: &[usize],
        upper: bool,
    ) -> ComputedPaddedDim<D> {
        let spatial_rank = data_spatial_shape.len();
        let mut dims = tvec![];
        let mut pad_before = tvec![];
        let mut pad_after = tvec![];
        for ax in 0..spatial_rank {
            let dim = data_spatial_shape[ax].div_ceil(strides[ax]);
            let kernel_field = (kernel_spatial_shape[ax].into() - 1) * dilations[ax] + 1;
            dims.push(dim);
            let pad = (dim - 1) * strides[ax] + kernel_field - data_spatial_shape[ax];
            let lower_pad = pad / 2;
            let higher_pad = pad - pad / 2;
            if upper {
                pad_before.push(lower_pad);
                pad_after.push(higher_pad);
            } else {
                pad_after.push(lower_pad);
                pad_before.push(higher_pad);
            }
        }
        ComputedPaddedDim {
            pad_before,
            pad_after,
            output: dims,
        }
    }
}