tract-onnx 0.23.0-dev.3

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

use crate::model::ParsingContext;
use crate::pb::NodeProto;
use tract_hir::internal::*;

pub fn instance_normalization(
    _ctx: &ParsingContext,
    node: &NodeProto,
) -> TractResult<(Box<dyn InferenceOp>, Vec<String>)> {
    let epsilon = node.get_attr_opt("epsilon")?.unwrap_or(1e-5);
    Ok((expand(InstanceNorm::new(epsilon)), vec![]))
}

#[derive(Debug, Clone, new, Default)]
pub struct InstanceNorm {
    epsilon: f32,
}

impl Expansion for InstanceNorm {
    fn name(&self) -> StaticName {
        "InstanceNorm".into()
    }

    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        s: &mut Solver<'r>,
        inputs: &'p [TensorProxy],
        outputs: &'p [TensorProxy],
    ) -> InferenceResult {
        check_input_arity(inputs, 3)?;
        check_output_arity(outputs, 1)?;
        s.equals(&inputs[0].datum_type, &outputs[0].datum_type)?;
        s.equals(&inputs[0].datum_type, &inputs[1].datum_type)?;
        s.equals(&inputs[0].datum_type, &inputs[2].datum_type)?;
        s.equals(&inputs[1].shape, &inputs[2].shape)?;
        s.equals(&inputs[0].shape, &outputs[0].shape)?;
        s.equals(&inputs[1].shape[0], &inputs[0].shape[1])?;
        Ok(())
    }

    fn wire(
        &self,
        name: &str,
        model: &mut TypedModel,
        inputs: &[OutletId],
    ) -> TractResult<TVec<OutletId>> {
        let input_fact = model.outlet_fact(inputs[0])?.clone();
        let rank = input_fact.rank();
        let axes: Vec<_> = (0..rank as i64).filter(|&axis| axis != 1).collect();
        let mean = tract_hir::ops::nn::Reduce::new(
            Some(axes.clone()),
            true,
            tract_hir::ops::nn::Reducer::Mean,
        )
        .wire(&format!("{name}.mean"), model, &inputs[0..1])?[0];
        let diff = model.wire_node(
            format!("{name}.diff"),
            tract_hir::ops::math::sub(),
            &[inputs[0], mean],
        )?;
        let sqr_diff =
            model.wire_node(format!("{name}.sqr"), tract_hir::ops::math::square(), &diff)?;
        let vari =
            tract_hir::ops::nn::Reduce::new(Some(axes), true, tract_hir::ops::nn::Reducer::Mean)
                .wire(&format!("{name}.variance"), model, &sqr_diff)?[0];
        let epsilon = model.add_const(
            format!("{name}.epsilon.cst"),
            tensor0(self.epsilon)
                .cast_to_dt(input_fact.datum_type)?
                .into_owned()
                .broadcast_into_rank(rank)?
                .into_arc_tensor(),
        )?;
        let vari_sane = model.wire_node(
            format!("{name}.epsilon"),
            tract_hir::ops::math::add(),
            &[vari, epsilon],
        )?;
        let div =
            model.wire_node(format!("{name}.rsqrt"), tract_hir::ops::math::rsqrt(), &vari_sane)?;
        let divised = model.wire_node(
            format!("{name}.div"),
            tract_hir::ops::math::mul(),
            &[diff[0], div[0]],
        )?;
        let mut scale =
            model.wire_node(format!("{name}.add-scale-axis-n"), AxisOp::Add(0), &inputs[1..2])?;
        for i in 2..rank {
            scale =
                model.wire_node(format!("{name}.add-scale-axis-{i}"), AxisOp::Add(2), &scale)?;
        }
        let scaled = model.wire_node(
            format!("{name}.scaled"),
            tract_hir::ops::math::mul(),
            &[divised[0], scale[0]],
        )?;
        let mut bias =
            model.wire_node(format!("{name}.add-bias-axis-n"), AxisOp::Add(0), &inputs[2..3])?;
        for i in 2..rank {
            bias = model.wire_node(format!("{name}.add-bias-axis-{i}"), AxisOp::Add(2), &bias)?;
        }
        model.wire_node(name, tract_hir::ops::math::add(), &[scaled[0], bias[0]])
    }
}