briny_ai 0.5.1

A tiny & efficient AI inference engine
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

use super::exp;
use crate::nn::{
    tensors::{Tensor, WithGrad},
    TensorFloat,
};
use tensor_optim::TensorOps;

#[cfg(feature = "dyntensor")]
use alloc::vec;

#[cfg(feature = "alloc")]
use alloc::boxed::Box;
#[cfg(not(feature = "alloc"))]
use box_closure::{Align8, OpaqueFn};

/// Applies a tanh activation function element-wise.
#[must_use]
#[cfg(feature = "dyntensor")]
pub fn tanh(
    input: &WithGrad<Tensor<TensorFloat>>,
) -> (
    Tensor<TensorFloat>,
    Box<dyn Fn(Tensor<TensorFloat>) -> Tensor<TensorFloat> + '_>,
) {
    let shape = input.get_value().shape();
    let input_data = input.get_value().data();

    let mut out_data = vec![TensorFloat::default(); input_data.len()];

    // Forward
    out_data
        .iter_mut()
        .zip(input_data.iter())
        .for_each(|(y, &x)| {
            // numerically stable tanh
            let e2x = exp(x + x);
            *y = (e2x - TensorFloat::from(1.0)) / (e2x + TensorFloat::from(1.0));
        });

    let out = Tensor::new(shape, &out_data);

    // Backward
    let back = {
        let y_saved = out_data;
        move |grad_output: Tensor<TensorFloat>| {
            let dy = grad_output.data();
            let mut grad = vec![TensorFloat::default(); dy.len()];

            grad.iter_mut()
                .zip(y_saved.iter())
                .zip(dy.iter())
                .for_each(|((g, &y), &dyi)| {
                    *g = dyi * (TensorFloat::from(1.0) - y * y);
                });

            Tensor::new(shape, &grad)
        }
    };

    (out, Box::new(back))
}

/// Applies a tanh activation function element-wise.
#[must_use]
#[cfg(all(feature = "alloc", not(feature = "dyntensor")))]
pub fn tanh<const N: usize, const D: usize>(
    input: &WithGrad<Tensor<TensorFloat, N, D>>,
) -> (
    Tensor<TensorFloat, N, D>,
    Box<dyn Fn(Tensor<TensorFloat, N, D>) -> Tensor<TensorFloat, N, D> + '_>,
) {
    use tensor_optim::ConstTensorOps;

    let shape: &[usize; D] = input.get_value().shape_array();
    let input_data = input.get_value().data();

    let mut out_data = [TensorFloat::default(); N];

    // Forward
    out_data
        .iter_mut()
        .zip(input_data.iter())
        .for_each(|(y, &x)| {
            let e2x = exp(x + x);
            *y = (e2x - TensorFloat::from(1.0)) / (e2x + TensorFloat::from(1.0));
        });

    let out = Tensor::new(shape, &out_data);

    // Backward
    let back = {
        let y_saved = out_data;
        move |grad_output: Tensor<TensorFloat, N, D>| {
            let dy = grad_output.data();
            let mut grad = [TensorFloat::default(); N];

            grad.iter_mut()
                .zip(y_saved.iter())
                .zip(dy.iter())
                .for_each(|((g, &y), &dyi)| {
                    *g = dyi * (TensorFloat::from(1.0) - y * y);
                });

            Tensor::new(shape, &grad)
        }
    };

    (out, Box::new(back))
}

/// Applies a tanh activation function element-wise.
#[must_use]
#[cfg(not(feature = "alloc"))]
pub fn tanh<const N: usize, const D: usize>(
    input: &WithGrad<Tensor<TensorFloat, N, D>>,
) -> (
    Tensor<TensorFloat, N, D>,
    OpaqueFn<'_, Tensor<TensorFloat, N, D>, Tensor<TensorFloat, N, D>, Align8<128>>,
) {
    use tensor_optim::ConstTensorOps;

    let shape: &[usize; D] = input.get_value().shape_array();
    let input_data = input.get_value().data();

    let mut out_data = [TensorFloat::default(); N];

    // Forward
    out_data
        .iter_mut()
        .zip(input_data.iter())
        .for_each(|(y, &x)| {
            let e2x = exp(x + x);
            *y = (e2x - TensorFloat::from(1.0)) / (e2x + TensorFloat::from(1.0));
        });

    let out = Tensor::new(shape, &out_data);

    let back = move |grad_output: Tensor<TensorFloat, N, D>| {
        let dy = grad_output.data();
        let mut grad = [TensorFloat::default(); N];

        grad.iter_mut()
            .zip(out_data.iter())
            .zip(dy.iter())
            .for_each(|((g, &y), &dyi)| {
                *g = dyi * (TensorFloat::from(1.0) - y * y);
            });

        Tensor::new(shape, &grad)
    };

    (out, OpaqueFn::new(back))
}