embedded-nn 0.2.0

A pure Rust, #![no_std] neural network inference library inspired by CMSIS-NN for microcontrollers and embedded targets.
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

use embedded_nn::{
    activations::{
        activation_s16, activation_s8, leaky_relu_s8, relu6_s8, relu_s16, relu_s8, sigmoid_s16,
        sigmoid_s8, tanh_s16, tanh_s8, SIGMOID_TABLE_UINT16,
    },
    float_ops::{relu6_f32, relu_f32},
    Activation,
};

#[test]
fn test_relu_s8_comprehensive() {
    let mut data = [i8::MIN, -128, -50, -1, 0, 1, 50, 127, i8::MAX];
    relu_s8(&mut data);
    assert_eq!(data, [0, 0, 0, 0, 0, 1, 50, 127, 127]);

    let mut empty: [i8; 0] = [];
    relu_s8(&mut empty);
}

#[test]
fn test_relu6_s8_comprehensive() {
    let mut data = [-128i8, -10, 0, 1, 5, 6, 7, 50, 127];
    relu6_s8(&mut data);
    assert_eq!(data, [0, 0, 0, 1, 5, 6, 6, 6, 6]);

    let mut empty: [i8; 0] = [];
    relu6_s8(&mut empty);
}

#[test]
fn test_activation_clamping_s8_comprehensive() {
    let mut data = [-100i8, -50, -10, 0, 5, 10, 50, 100];
    let act = Activation::new(-10, 10);
    activation_s8(&mut data, act);
    assert_eq!(data, [-10, -10, -10, 0, 5, 10, 10, 10]);

    let unconstrained = Activation::int8_unconstrained();
    let mut data2 = [-128i8, 0, 127];
    activation_s8(&mut data2, unconstrained);
    assert_eq!(data2, [-128, 0, 127]);
}

#[test]
fn test_relu_s16_comprehensive() {
    let mut data = [i16::MIN, -32768, -1000, -1, 0, 1, 1000, 32767, i16::MAX];
    relu_s16(&mut data);
    assert_eq!(data, [0, 0, 0, 0, 0, 1, 1000, 32767, 32767]);

    let mut empty: [i16; 0] = [];
    relu_s16(&mut empty);
}

#[test]
fn test_activation_clamping_s16_comprehensive() {
    let mut data = [-20000i16, -1000, 0, 1000, 20000];
    let act = Activation::new(-500, 500);
    activation_s16(&mut data, act);
    assert_eq!(data, [-500, -500, 0, 500, 500]);

    let unconstrained = Activation::int16_unconstrained();
    let mut data2 = [-32768i16, 0, 32767];
    activation_s16(&mut data2, unconstrained);
    assert_eq!(data2, [-32768, 0, 32767]);
}

#[test]
fn test_leaky_relu_s8_offsets_and_alpha() {
    let input = [-20i8, -10i8, 0i8, 10i8, 20i8];
    let mut output = [0i8; 5];

    // Alpha multiplier = 0.5 (Q31 mult = 1073741824, shift = 0)
    // input_offset = 0, output_offset = 0
    leaky_relu_s8(&input, &mut output, 1073741824, 0, 0, 0);
    assert_eq!(output[0], -10);
    assert_eq!(output[1], -5);
    assert_eq!(output[2], 0);
    assert_eq!(output[3], 10);
    assert_eq!(output[4], 20);

    // Test with offsets
    let mut output_off = [0i8; 5];
    // input_offset = 10 -> shifted inputs: [-10, 0, 10, 20, 30]
    // output_offset = -5
    leaky_relu_s8(&input, &mut output_off, 1073741824, 0, 10, -5);
    // -10 * 0.5 - 5 = -10
    // 0 - 5 = -5
    // 10 - 5 = 5
    // 20 - 5 = 15
    // 30 - 5 = 25
    assert_eq!(output_off, [-10, -5, 5, 15, 25]);
}

#[test]
fn test_sigmoid_s8_full_range() {
    let input = [-128i8, -90, -40, -10, 0, 10, 40, 90, 127];
    let mut output = [0i8; 9];
    sigmoid_s8(&input, &mut output);

    // Sigmoid must be strictly monotonic
    for i in 0..output.len() - 1 {
        assert!(
            output[i] <= output[i + 1],
            "Failed monotonicity at index {}",
            i
        );
    }
    // Around 0 input, output should be around 0 in centered s8
    assert!((output[4] as i32).abs() <= 5);
    // Extreme values should saturate near -128 and 127
    assert!(output[0] < -100);
    assert!(output[8] > 100);
}

#[test]
fn test_tanh_s8_full_range() {
    let input = [-128i8, -80, -40, 0, 40, 80, 127];
    let mut output = [0i8; 7];
    tanh_s8(&input, &mut output);

    // Tanh must be symmetric: tanh(-x) approx -tanh(x)
    assert_eq!(output[3], 0); // tanh(0) == 0
    assert!((output[2] as i32 + output[4] as i32).abs() <= 2);
    assert!((output[1] as i32 + output[5] as i32).abs() <= 2);
    assert!(output[0] <= output[1]);
    assert!(output[5] <= output[6]);
}

#[test]
fn test_sigmoid_s16_shifts() {
    let input = [-1000i16, -100, 0, 100, 1000];

    for shift in [-3, -1, 0, 1, 3] {
        let mut output = [0i16; 5];
        sigmoid_s16(&input, &mut output, shift);
        // Sigmoid must be monotonic
        for i in 0..output.len() - 1 {
            assert!(output[i] <= output[i + 1]);
        }
    }
}

#[test]
fn test_tanh_s16_shifts() {
    let input = [-1000i16, -100, 0, 100, 1000];

    for shift in [-3, -1, 0, 1, 3] {
        let mut output = [0i16; 5];
        tanh_s16(&input, &mut output, shift);
        println!("shift={}: {:?}", shift, output);
        // Tanh must be monotonic
        for i in 0..output.len() - 1 {
            assert!(
                output[i] <= output[i + 1],
                "Failed at shift {} index {}: {:?}",
                shift,
                i,
                output
            );
        }
    }
}

#[test]
fn test_sigmoid_table_properties() {
    assert_eq!(SIGMOID_TABLE_UINT16.len(), 256);
    assert_eq!(SIGMOID_TABLE_UINT16[0], 32768); // Sigmoid(0) scaled Q0.16 is 0.5 * 65536 = 32768
    assert!(SIGMOID_TABLE_UINT16[255] >= 65530);
}

#[test]
fn test_f32_activations_edge_cases() {
    let mut data = [-100.0f32, -0.001, 0.0, 0.001, 5.999, 6.0, 6.001, 100.0];
    relu_f32(&mut data);
    assert_eq!(data, [0.0, 0.0, 0.0, 0.001, 5.999, 6.0, 6.001, 100.0]);

    let mut data6 = [-100.0f32, -0.001, 0.0, 0.001, 5.999, 6.0, 6.001, 100.0];
    relu6_f32(&mut data6);
    assert_eq!(data6, [0.0, 0.0, 0.0, 0.001, 5.999, 6.0, 6.0, 6.0]);
}