nnv-rs 0.6.1

Verification and Statistics on Deep Neural Networks
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

use crate::affine::Affine2;
use crate::bounds::Bounds1;
use crate::deeppoly::deep_poly_relu;
use crate::deeppoly::deep_poly_steprelu;
use crate::dnn::layer::Layer;
use crate::star::Star2;
use crate::star_node::StarNodeType;
use crate::NNVFloat;
use ndarray::Array1;
use ndarray::Array2;
use serde::{Deserialize, Serialize};
use std::fmt::{Display, Formatter, Result};

#[derive(Clone, Debug, Deserialize, Serialize, PartialEq)]
pub struct ReLU {
    ndims: usize,
}

impl ReLU {
    pub const fn new(ndims: usize) -> Self {
        Self { ndims }
    }
}

impl Display for ReLU {
    fn fmt(&self, f: &mut Formatter) -> Result {
        write!(f, "ReLU")
    }
}

#[typetag::serde]
impl Layer for ReLU {
    fn is_activation(&self) -> bool {
        true
    }

    fn input_dims(&self) -> usize {
        self.ndims
    }

    fn output_dims(&self) -> usize {
        self.ndims
    }

    fn forward1(&self, input: &Array1<NNVFloat>) -> Array1<NNVFloat> {
        input.mapv(|x| if x.lt(&0.) { 0. } else { x })
    }

    fn forward2(&self, input: &Array2<NNVFloat>) -> Array2<NNVFloat> {
        input.mapv(|x| if x.lt(&0.) { 0. } else { x })
    }

    fn apply_bounds(
        &self,
        bounds: &Bounds1,
        lower_aff: &Affine2,
        upper_aff: &Affine2,
    ) -> (Bounds1, (Affine2, Affine2)) {
        if (self.ndims + 1) == bounds.ndim() {
            deep_poly_relu(bounds, lower_aff, upper_aff)
        } else {
            let (bounds_head, bounds_tail) = bounds.split_at(self.ndims);
            let (lower_aff_head, lower_aff_tail) = lower_aff.split_at(self.ndims);
            let (upper_aff_head, upper_aff_tail) = lower_aff.split_at(self.ndims);
            let (bounds_part, (lower_part, upper_part)) =
                deep_poly_relu(&bounds_head, &lower_aff_head, &upper_aff_head);
            (
                bounds_part.append(&bounds_tail),
                (
                    lower_part.append(&lower_aff_tail),
                    upper_part.append(&upper_aff_tail),
                ),
            )
        }
    }

    fn apply_bounds_step(
        &self,
        dim: usize,
        bounds: &Bounds1,
        lower_aff: &Affine2,
        upper_aff: &Affine2,
    ) -> (Bounds1, (Affine2, Affine2)) {
        deep_poly_steprelu(dim, bounds.clone(), lower_aff.clone(), upper_aff.clone())
    }

    fn get_activation_pattern(&self, state: &Array2<NNVFloat>) -> Option<Array2<bool>> {
        // This should only be Some in an activation layer (e.g. ReLU)
        Some(state.mapv(|x| x >= 0.0))
    }

    fn forward_star(
        &self,
        star: &Star2,
        dim: Option<usize>,
        input_bounds: Option<Bounds1>,
        parent_bounds: Option<Bounds1>,
    ) -> (Vec<Star2>, Vec<Option<Bounds1>>, bool) {
        let dim = dim.unwrap();
        let child_stars = star.step_relu2(dim, &input_bounds);
        let parent_bounds = parent_bounds.unwrap();
        let mut same_output_bounds = false;
        let mut stars = vec![];
        let mut star_input_bounds = vec![];
        let is_single_child = child_stars.0.is_some() ^ child_stars.1.is_some();

        if let Some(mut lower_star) = child_stars.0 {
            let mut bounds = parent_bounds.clone();
            bounds.index_mut(dim)[0] = 0.;
            bounds.index_mut(dim)[1] = 0.;
            if is_single_child {
                // Remove redundant constraint added by step_relu2 above
                let num_constraints = lower_star.num_constraints();
                lower_star = lower_star.remove_constraint(num_constraints - 1);
            }

            stars.push(lower_star);
            star_input_bounds.push(Some(bounds));
        }

        if let Some(mut upper_star) = child_stars.1 {
            let mut bounds = parent_bounds;
            let mut lb = bounds.index_mut(dim);
            if lb[0].is_sign_negative() {
                lb[0] = 0.;
            }
            if is_single_child {
                // Remove redundant constraint added by step_relu2 above
                let num_constraints = upper_star.num_constraints();
                upper_star = upper_star.remove_constraint(num_constraints - 1);
                same_output_bounds = true;
            }
            stars.push(upper_star);
            star_input_bounds.push(Some(bounds));
        }
        (stars, star_input_bounds, same_output_bounds)
    }

    fn construct_starnodetype(&self, child_ids: &[usize], dim: Option<usize>) -> StarNodeType {
        debug_assert_gt!(child_ids.len(), 0);
        StarNodeType::StepRelu {
            dim: dim.unwrap(),
            fst_child_idx: child_ids[0],
            snd_child_idx: child_ids.get(1).copied(),
        }
    }
}