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

use embedded_nn::{
    recurrent::{lstm_step_s16, lstm_step_s8_s16, svdf_s8, svdf_state_s16_s8, LstmGateParams},
    Activation, PerTensorQuantParams,
};

#[test]
fn test_svdf_s8_rank1_execution() {
    let input = [10i8, 20i8]; // 2 input channels
    let weights_feature = [1i8, 2i8, 3i8, 4i8]; // 2 filters, rank 1, 2 in_c
    let weights_time = [1i8, 1i8]; // 2 filters, rank 1, memory_size 1
    let bias = [0i32, 0i32];

    let mut state = [0i8; 2]; // 2 filters * rank 1 * memory_size 1 = 2
    let mut output = [0i8; 2];

    let input_quant = PerTensorQuantParams::new(1073741824, 0); // 0.5
    let output_quant = PerTensorQuantParams::new(1073741824, 0); // 0.5
    let act = Activation::int8_unconstrained();

    // Step 1
    svdf_s8(
        0, // input_offset
        0, // output_offset
        1, // rank
        &input,
        &mut state,
        &weights_feature,
        &weights_time,
        Some(&bias),
        &input_quant,
        &output_quant,
        &act,
        &mut output,
    )
    .unwrap();

    // Feature dot:
    // Filter 0: (10*1 + 20*2) * 0.5 = 25
    // Filter 1: (10*3 + 20*4) * 0.5 = 55
    // Time dot with state:
    // Filter 0: 25 * 0.5 = 12.5 -> rounds to 13
    // Filter 1: 55 * 0.5 = 27.5 -> rounds to 28
    assert_eq!(output[0], 13);
    assert_eq!(output[1], 28);
}

#[test]
fn test_svdf_state_s16_s8_execution() {
    let input = [5i8, 10i8];
    let weights_feature = [2i8, 2i8];
    let weights_time = [1i16];
    let bias = [0i32];

    let mut state = [0i16; 1];
    let mut output = [0i8; 1];

    let input_quant = PerTensorQuantParams::new(2147483647, 0); // 1.0
    let output_quant = PerTensorQuantParams::new(2147483647, 0); // 1.0
    let act = Activation::int8_unconstrained();

    svdf_state_s16_s8(
        0, // input_offset
        0, // output_offset
        1, // rank
        &input,
        &mut state,
        &weights_feature,
        &weights_time,
        Some(&bias),
        &input_quant,
        &output_quant,
        &act,
        &mut output,
    )
    .unwrap();

    // Feature dot: 5*2 + 10*2 = 30 -> written to state[0]
    assert_eq!(state[0], 30);
    // Time dot: 30 * 1 = 30 -> >> 15 = 0
    assert_eq!(output[0], 0);
}

#[test]
fn test_lstm_step_s8_s16_execution() {
    let input = [10i8, 20i8]; // input_dim = 2
    let mut hidden_state = [0i8; 1]; // hidden_dim = 1
    let mut cell_state = [0i16; 1];

    // Weights: 4 * hidden_dim x input_dim = 4 * 1 * 2 = 8
    let weight_input = [1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8];
    // Weights hidden: 4 * hidden_dim x hidden_dim = 4 * 1 * 1 = 4
    let weight_hidden = [0i8, 0i8, 0i8, 0i8];
    // Bias: 4 * hidden_dim = 4
    let bias = [0i32, 0i32, 0i32, 0i32];

    let gate_params = LstmGateParams {
        input_offset: 0,
        hidden_offset: 0,
        multiplier: 1073741824,
        shift: 0,
    };
    let output_quant = PerTensorQuantParams::new(1073741824, 0);

    lstm_step_s8_s16(
        &input,
        &mut hidden_state,
        &mut cell_state,
        &weight_input,
        &weight_hidden,
        &bias,
        &gate_params,
        32767,
        &output_quant,
        0,
        &Activation::int8_unconstrained(),
    )
    .unwrap();
}

#[test]
fn test_lstm_step_s16_execution() {
    let input = [100i16, 200i16];
    let mut hidden_state = [0i16; 1];
    let mut cell_state = [0i16; 1];

    let weight_input = [1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8];
    let weight_hidden = [0i8, 0i8, 0i8, 0i8];
    let bias = [0i64, 0i64, 0i64, 0i64];

    let gate_params = LstmGateParams {
        input_offset: 0,
        hidden_offset: 0,
        multiplier: 1073741824,
        shift: 0,
    };
    let output_quant = PerTensorQuantParams::new(1073741824, 0);

    lstm_step_s16(
        &input,
        &mut hidden_state,
        &mut cell_state,
        &weight_input,
        &weight_hidden,
        &bias,
        &gate_params,
        32767,
        &output_quant,
        &Activation::int16_unconstrained(),
    )
    .unwrap();
}