jiro_nn 0.5.0

Neural Networks framework with model specification & data preprocessing features.
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

use serde::{Deserialize, Serialize};

use crate::{
    learning_rate::{default_learning_rate, LearningRateSchedule},
    linalg::{Matrix, MatrixTrait, Scalar},
};

pub(crate) fn default_momentum() -> Scalar {
    0.9
}

// https://arxiv.org/pdf/1207.0580.pdf
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Momentum {
    #[serde(default = "default_momentum")]
    momentum: Scalar,
    #[serde(default = "default_learning_rate")]
    learning_rate: LearningRateSchedule,
    #[serde(skip)]
    v: Option<Matrix>,
}

impl Momentum {
    pub fn new(learning_rate: LearningRateSchedule, momentum: Scalar) -> Self {
        Self {
            v: None,
            momentum,
            learning_rate,
        }
    }

    pub fn default() -> Self {
        Self {
            v: None,
            momentum: default_momentum(),
            learning_rate: default_learning_rate(),
        }
    }

    pub fn update_parameters(
        &mut self,
        epoch: usize,
        parameters: &Matrix,
        parameters_gradient: &Matrix,
    ) -> Matrix {
        let lr = self.learning_rate.get_learning_rate(epoch);

        if let None = &self.v {
            let (nrow, ncol) = parameters_gradient.dim();
            self.v = Some(Matrix::zeros(nrow, ncol));
        };

        let v = self.v.as_ref().unwrap();

        let v = v
            .scalar_mul(self.momentum)
            .component_add(&parameters_gradient.scalar_mul(lr));

        let new_params = parameters.component_sub(&v);
        self.v = Some(v);
        new_params
    }
}